Beruflich Dokumente
Kultur Dokumente
By Mateusz P. Dziubinski
A dissertation submitted to Business and Social Sciences, Aarhus University, in partial fulfilment of the requirements of the PhD degree in Economics and Management
Contents
Preface Summary Chapter 1. Option Valuation with the Simplied Component GARCH Model Chapter 2. Conditionally-Uniform Feasible Grid Search Algorithm Chapter 3. Commodity Derivatives Pricing with Inventory Eects 57 25 1 v vii
iii
Preface
This PhD dissertation was written in the period from February 2008 to February 2012 during my studies at the Department of Economics and Business at Aarhus University. I am grateful to the department and to the Center for Research in Econometric Analysis of Time Series (CREATES), funded by the Danish National Research Foundation, for generous nancial support in connection with courses and conferences. I am further grateful to CREATES for providing excellent research facilities. A number of people have contributed to the making of this thesis. I would like to express my sincere gratitude to my main advisor Timo Tersvirta for providing guidance and numerous insightful comments and suggestions throughout my PhD studies. I am grateful for having had the opportunity to work with my co-author Christian Bach on one of the chapters. I hope we can continue our collaboration in the years to come. At Aarhus University I would like to thank the faculty and fellow students. I would also like to thank CREATES, Department of Economics and Business, Department of Mathematics, and Computer Science Department for providing outstanding courses, seminars, conferences, and support. Special thanks also to Steen Thorbjrnsen for showing me that stochastic processes can be easy and Svend Erik Graversen for revealing the mysteries of stochastic calculus to me and even spending his free time doing so. At the Department of Economics and Business I would like to thank my fellow PhD students for providing me with a welcoming, friendly environment and opportunities to discuss economics, nance and (un)related topics. Christian deserves my gratitude for sharing an oce with me and providing me with endless opportunities for conversation, debate and learning relevant to our common interests. Finally, I would like to thank Anders for accompanying me on our individual yet common journey in the probability theory wonderland. Last but not least, I would like to thank my family for their continuing love and support.
vi
PREFACE
The predefense was held on March 21, 2012. I would like to thank
the members of the assessment committee, Peter Christoersen, Olaf Posch (chair), and Lars Stentoft for their comments and suggestions for improvements. Most of those have been incorporated into the dissertation.
Summary
The unifying themes of this dissertation are nancial econometrics and derivatives pricing. An underlying topic is the price behavior of a nancial asset, be it a stock market index or a commodity derivative, with an additional focus on the factors aecting this behavior, like the volatility and inventory levels. Practical implementation issues arising in applying the models, together with an empirical motivation for the
why
how.
The thesis contains three independent chapters, of which the rst two can be viewed as contributions to nancial econometrics, whereas the third chapter has to do with derivatives pricing. In the rst chapter, entitled "Option Valuation with the Simplied Component GARCH Model", I introduce the Simplied Component GARCH (SCGARCH) option pricing model, show and discuss sucient conditions for nonnegativity of the conditional variance, apply the model to both lowfrequency and high-frequency nancial data, and consider the option valuation, comparing the model performance with similar models from the literature. Two volatility components in my model allow me to satisfactorily model time structure of volatility. The SCGARCH model builds on Engle and Lee (1999), Heston and Nandi (2000), and Christoersen et al. (2008) (hereafter referred to as CJOW) models. Engle and Lee (1999) introduced the volatility component model in the GARCH context, while Heston and Nandi (2000) introduced a model with a closed-form solution for the European call option-pricing formulas. The CJOW model is a generalization of the Heston and Nandi model allowing for a time-varying long-run component. sured. In the second chapter, entitled "Conditionally-Uniform Feasible Grid Search Algorithm", I present and evaluate a numerical optimization method (together with an algorithm for choosing the starting values) pertinent to the constrained optimization problem where the variables have to satisfy a sequentially dependent set of constraints. In practice, these arise (for instance) in the estimation of the models with inequality constraints, in particular GARCH models such as the Engle and Lee (1999) GARCH model and the Simplied Component The SCGARCH model is a simplied variant of the CJOW model, in which the non-negativity of the conditional variance is en-
vii
viii
Summary
GARCH (SCGARCH) model. The numerical optimization method, the Conditionally-Uniform Feasible Grid Search (CUFGS), is essentially a particular kind of a random grid search coupled with a constrained feasible Sequential Quadratic Programming (SQP) algorithm. One of the reasons for developing it are the problems encountered when using non-specialized gradient-based algorithms due to constrained feasible space requirement and scaling. For example, in relation to the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm, popular among econometricians, Nocedal and Wright (2006) write: (1)
(2)
BFGS updating is generally less eective for constrained problems than in the unconstrained case because of the requirement of maintaining a positive denite approximation to an underlying matrix that often does not have this property. SQP methods are most ecient if the number of active constraints is nearly as large as the number of variables, that is, if the number of free variables is relatively small. They require few evaluations of the functions, in comparison with augmented Lagrangian methods, and can be more robust on badly scaled problems.
I also provide the objective function and analytical gradient computation algorithms for the SCGARCH model, which are useful for the 1 practical implementation purposes. In the third chapter, "Commodity Derivatives Pricing with Inventory Eects" (written jointly with Christian Bach), we introduce tractable models for commodity derivatives pricing with inventory and volatility eects, introduce a new, maturity-wise calibration method compatible with these models and apply it to modeling the commodity derivatives associated with the crude oil market. The role of inventories in explaining price and volatility of commodities has been studied in several papers. Brennan (1958) and Telser (1958) are early studies on the eect of the level of inventory on agricultural commodities, but the inventory eect has also been documented for metals (Ng and Pirrong (1994)) and oil and natural gas markets (Geman and Ohana (2009)). Instead of relying on a proxy for inventory data, we use weekly data on oil inventories. Geman and Nguyen (2005) construct a database of soybean inventory over a 10-year period and show that volatility can be written as an exact inverse function of inventory. In this chapter we do not nd such a clear relationship,
Summary
ix
although we see strong signs of the relationship between inventory and volatility being negative. We contribute to the existing literature in several respects. First, whereas the previous literature uses futures data for investigating the relationship between inventory and volatility, we use the information available in options traded on futures. Second, performance assessment in the previous literature has primarily evolved around explaining moments of data or forecasting prices of futures. Instead, we asses the performance of our model by considering both its ability of explaining prices in-sample and out-of-sample assessing both the pricingperformance and the hedging-performance of the models. Third, we model the futures surface rather than the spot price process, and limit the number of parameters to calibrate (using the observed inventory process instead of a latent one). We introduce a new, maturity-wise calibration method compatible with this modeling methodology. Fourth, we use actual data on inventories rather than a proxy. Fifth, our model is very exible and allows for analyzing several dierent types of relationships between inventory and volatility.
Bibliography
Brennan, M. J. (1958). The supply of storage.
Brooks, C., S. P. Burke, and G. Persand (2001). the accuracy of GARCH model estimation.
Christoersen, P., K. Jacobs, C. Ornthanalai, and Y. Wang (2008). Option valuation with long-run and short-run volatility components.
Cointegration, Causality, and Forecasting: A Festschrift in Honuor of Clive W.J. Granger, pp. 475497. Oxford Management Science 51 (7), 10761091.
Soybean inventory and forward
Geman, H. and S. Ohana (2009). Forward curves, scarcity and price volatility in oil and natural gas markets. 576585.
The Journal of
Zivot, E. (2009). Practical issues in the analysis of univariate GARCH models. In T. G. Andersen, R. A. Davis, J.-P. Krei, and T. Mikosch (Eds.),
xi
CHAPTER 1
GARCH) option pricing model, show and discuss sucient conditions for non-negativity of the conditional variance, apply it to low-frequency and high-frequency nancial data, and consider the option valuation, comparing the model performance with similar models from the literature. Two volatility components in our model, short-term and long-term, allow us to model time-structure of volatility.
JEL Classication.
G12, C32.
1.
Introduction
In this paper we introduce a discrete-time volatility model in which the conditional variance of the underlying asset follows a particular GARCH process. Our model can be used for option pricing, while two volatility components allow us to model time structure thereof. The model builds on Engle and Lee (1999), Heston and Nandi (2000) and Christoersen et al. (2008) (hereafter referred to as CJOW) models. The model by Engle and Lee (1999) introduced the volatility component model in the GARCH context, while Heston and Nandi (2000) introduced a model with a closed-form solution for the European call option-pricing formulas. The CJOW model is a generalization of the Heston and Nandi model allowing for a time-varying long-run component. Our model is a simplied specication of the CJOW model, which solves the problem of ensuring the non-negativity of the conditional variance.
option pricing. We wish to thank Timo Tersvirta for a discussion regarding non-negativity conditions. We acknowledge nancial support by the Center for Research in Econometric Analysis of Time Series, CREATES, funded by the Danish National Research Foundation. All errors, omissions and mistakes are author's own responsibility.
2
The paper proceeds as follows. In Section 2 we provide basic denitions and notation. We introduce the model in Section 3 (in which we also discuss the non-negativity conditions), and discuss the estimation thereof in Section 4. our conclusions. In Section 5 we present the estimation results. Section 6 is devoted to option pricing, and, nally, Section 7 contains
2.
(, F , P ) and a ltration F = (Ft )tT , where, depending on the context, we shall assume T = Z+ or T = Z [0, T ], T > 0 or T = Z [1, T ], T 0. We refer to P as the physical probability measure and we call (, F , F, P ) a ltered physical probability space. We shall also use probability measure Q on (, F ) and refer to it as the risk-neutral probability measure.
We assume as given a probability space
stochastic process X on (, F , P ) is a collection of R-valued random variables (Xt )tT , and we denote it by X = (Xt )tT .
A The process
X X
is said to be
t T).
adapted
if
Xt Ft t T (that is,
if
it is
Ft
is said to be
predictable
X is said to be (F, P )-white noise with mean X and P 2 2 variance X , written X W N (X , X ) if and only if, under probability measure P , X has mean X R and covariance function (s, t) = 2 2 R++ . |ts| , where h := {0} (h) is the Kronecker delta and X X
X P.
Xt Ft1 t T,
and we
The process
(F, P )-Gaussian white noise with mean P P 2 2 X and variance X , written X GW N (X , X ) if and only if X P 2 2 ) and Xt N (X , X ) t T. W N (X , X X
is said to be First-order partial dierential operator with respect to
x is denoted x .
For further details regarding stochastic processes and time series we refer the reader to Protter (2005) and Brockwell and Davis (1991).
3.
The Model
We begin by presenting the SCGARCH and the CJOW models. The advantages of the CJOW model are the existence of a (quasi-)closedform solution for the option pricing formulas, improved ability to model the smirk and the path of spot volatility and, distinctively, the ability to model the volatility term structure for details, see Christoersen et al. (2008). A problem with this model is that the volatility components
MATT P. DZIUBINSKI
may admit negative values. This leads to a contradiction in the context of conditional variance modeling, as the conditional variance cannot be negative. We propose the SCGARCH model as a more parsimonious model which solves this problem and discuss its relation to the CJOW model. Furthermore, we shall consider the properties of the model and discuss the estimation of its parameters.
Assumption 1. The spot asset price, S (including accumulated interest or dividends) follows (over time steps of length 1) the following process under the physical probability measure P ,
rt+1 log
with
t+1 = rf + vt+1
2 ux,t = (wt 1) P
w GW N (0, 1)
where rf is the continuously compounded interest rate for the time interval of length , vt is the conditional variance of the log return between t 1 and t, with v P .
We use a notation that is closely linked to the interpretation of our model. First, the with cess
being its physical conditional mean, while is the long-run volatility component.
v .
s = v x. Under weak stationarity (discussed in the sequel) we have E[vt+1 ] = E[xt+1 ] = mx nx /(1 px ). Thus, mx is the unconditional mean of x and v , with nx being the numerator of nx /(1 px ), directly proportional to the unconditional mean level. Third, the ux and uv processes serve as mean-zero innovations for x and v , respectively, with the coecients ix and iv measuring the strength of the impact of those innovations. The coecient px measures the persistence of x. Analo2 gously, the persistence of v is measured by bv = pv iv gv , with gv being the asymmetry coecient. Finally, the source of the randomness w is the (F, P )-Gaussian white noise with mean 0 and variance 1, hereafter also referred to as the (F, P )-standard Gaussian white noise.
3.1.
First, we shall
look at the CJOW model and consider the issues regarding the nonnegativity of the conditional variance arising in its application.
3.1.1.
CJOW Model.
(3.8)
arrive at a possible contradiction. In order to examine the seriousness of the problem, we perform a simulation study and analyze the behavior of the model.
3.1.2.
to examine the behavior of this model performing a grid search with respect to
px
searching from
0.0
to
1.0
0.001.
We
do this both for the original CJOW model and a deterministic version thereof (i.e. the one where the driving noise process is assumed to be identically equal to zero instead of a standard GWN), xing all the other parameter values to those in Table 1 in Christoersen et al. (2008) (for convenience, we reproduce it in Table 1) in addition setting rf 1 to 1.000 10 . We choose this particular parameter value, since it is
the one used by Christoersen et al. (2008) to dierentiate between the Component and the Persistent Component (px is that for
= 1)
models. Further-
more, the reason we consider the unit interval as the parameter range
px < 0
px > 1
invalid
px
invalid
and
valid
values,
v.
We
nd that the low parameter values are invalid, while the higher ones are valid the boundary being at approximately all
0.9.
our simulation study there exists a Note, that in practice this leads
MATT P. DZIUBINSKI
T = rf nx iv ix pv px gv gx
1,000
Simulation
Table 1.
1.000 101 2.092 10+0 8.208 107 1.580 106 2.480 106 6.437 101 9.896 101 4.151 10+2 6.324 10+1
NaN
conforming architecture.
Since commonly
applied optimization routines will reject arguments leading to NaNs (or terminate with an error, leading to restarting the optimization with dierent starting values), this potentially explains the estimate of
px =
et al. (2008), which is very close to property of the model, one cannot
necessarily infer high persistence to hold in this case. This is because the high estimate might well be a numerical artifact, as opposed to being an empirical property of the data described by the model. In addition, as we change the sample size
T,
creases as the sample size increases. A possible interpretation of this nding is that as the model runs for a longer time (i.e., as we have more draws in the generated sample) the chance of drawing at least one negative value increases. However, this is not solely due to Gaussianity of
w,
model (i.e. even for a biased forecast) in fact, the boundary is higher for the deterministic case than the stochastic one.
3.1.3.
assuming the CJOW model, we rewrite (3.2) and (3.3), substituting (3.5) and (3.8), respectively:
(3.9) (3.10)
square root function to argument outside its domain, due to attempt to take the square root of a negative number.
2IEEE
oating-point representation of real numbers today on computers for further reference, see IEEE Task P754 (2008).
where
nx = mx (1 px ).
Rearranging terms, we obtain
(3.11)
px > 0 and v > 0. Consider two cases with respect to ix . 2 If we assume ix < 0, we have, in (3.13), that ix gx vt > 0 and ix > 0 2 this, however, results in ix (wt gx vt ) < 0. On the other hand, if we assume 0 < ix (and we may also want ix < nx < nx , so that 2 nx ix > 0), then ix gx vt < 0. Hence, we conclude that P (t T : xt < 0) > 0. However, since x is the long-run volatility component, it
Now, assume should remain non-negative over time. Furthermore, even if we assume
(3.12) (3.13)
vt+1
where (3.18) In fact, we can obtain P and the fact that w (3.19)
iv = 0, P (t T : vt < 0) > 0. the result for an arbitrary t T, using (3.17) GW N (0, 1):
Now, for all
2 = P (iv + bv vt + iv (wt gv vt ) < 0|vt > 0) (3.20) 2 = P ((wt gv vt ) < (iv bv vt )/iv |vt > 0) (3.21) = P ((iv bv vt )/iv < wt gv vt < (iv bv vt )/iv |vt > 0) (3.22) (iv bv vt ) (iv bv vt ) = P ( + gv vt < wt < + gv vt | vt > 0) > 0, iv iv
(3.23)
3Note
that this is the weakest assumption possible to ensure non-negativity of weaker than assuming
x,
i.e.,
x0
x > 0.
MATT P. DZIUBINSKI
gx
leads to
a model that can result in negative values for the volatility components. Note, that this is an inherent problem of the CJOW model per se in particular, this is not merely a peripheral problem limited to a given particular numerical treatment of the model (as in, for instance, particular discretization schemes applied to the Heston model). This also implies that the numerical optimization problem arising in the estimation of the CJOW model will, in general, be an ill-posed problem, due to inherent numerical instability associated with the presence of the NaN results on the IEEE 754 conforming architecture.
3.1.4.
Section 4.2) provide the mapping between CJOW and Heston-Nandi GARCH(2,2). We can write the conditional variance in the component model as a Heston-Nandi GARCH(2,2) process.
rt+1 = rf + ht+1 +
(3.24)
ht1 )2
(3.25) (3.26)
(3.30)
(3.31)
(3.32) (3.33)
Note that we can easily ensure non-negativity of the conditional variance in the HN-GARCH(2,2) model (HN) imposing (sucient) conditions given by the following inequality constraints:
(3.34)
As such, the original, unrestricted HN model does not suer from the lack of simple non-negativity conditions. In contrast, nonlinear parameter restrictions (3.27)(3.33) are precisely what makes it quite dicult
to come up with sensible constraints in particular, note that restrictions (3.47)-(3.49) (derived in the following section, operating under the
gx = 0, also desirable for the interpretation of the model) would imply that a2 < 0 (with the sign of a1 inversely related to the sign of a2 ) and make the signs of b1 and b2 dependent on
additional assumption of a relation analogous to (3.49). This illustrates the trade-o between the CJOW and the HN models although Christoersen et al. (2004, Section 4.2) argue that coming up with sensible parameter starting values (in the estimation context) and the stationarity requirements is simpler in the CJOW model compared to the HN model, it is easier to obtain simple non-negativity constraints in the HN model, as in (3.34). Here, we oer an alternative approach, allowing us to proceed with a model with a structure similar to CJOW (thus preserving the ease of starting values interpretation and simple stationarity conditions) with an additional benet of also having relatively simple non-negativity conditions.
3.1.5.
A Solution.
specication which allows us to derive sucient conditions for the volatility components to stay non-negative, given Assume that component
x0 > 0
and
v0 > 0.
gx = 0.
x.
presented in Engle and Lee (1999, Section 6) where the "leverage" term is signicant only in the transitory component (corresponding to non-zero
gv )
for all the data sets (including the S&P 500 index stud-
ied here). Engle and Lee (1999) also cite the nding in Gallant et al. (1993) (using a non-parametric approach) strongly supporting this hypothesis. The theoretical explanation oered by Engle and Lee (1999) states that while the debt-equity ratio may be hard to adjust in the short run, there is no reason that rms will not be able to adjust their capital structure over time toward a long-term "target value" (and thus Engle and Lee (1999) anticipate no asymmetric response of the volatility expectation to shocks in the long run). Hence, this assumption is consistent both with this theoretical hypothesis and with the empirical ndings of Engle and Lee (1999) and Gallant et al. (1993). We have
(3.35)
(3.36)
10
MATT P. DZIUBINSKI
2 xt+1 = mx (1 px ) + px xt + ix (wt 1) 2 = nx ix + px xt + ix wt
(3.37) (3.38)
where
nx = mx (1 px ).
(1995) .
(3.39)
Note, that (3.38) follows the GMACH(1, 1) model by Yang and Bewley
x,
we need
nx > ix > 0
which we also
nx . 1 px
(3.40)
mx
x.
(3.41)
(3.42)
2 = nx ix + px xt + ix wt
(3.43)
2 pv = bv + iv gv
we have
2 vt+1 = nx ix + px xt + ix wt
(3.44)
2 i x wt
(3.45)
(3.46)
bv > 0, iv > 0 and ix > 0, we have a sucient condition for non-negativity of v , which is (nx ix iv ) + (px pv )x > 0. Since we have already established conditions for non-negativity of x, we need to ensure that in addition to them, (nx ix iv ) > 0, (px pv ) > 0 and bv > 0. Thus, the joint sucient conditions for non-negativity of
Now, assuming
4This
is similar to assuming
c=0
11
and
Restrictions in (3.47) are analogous to those in Engle and Lee (1999). Note, that similarly to Engle and Lee (1999) we also assume that the weak stationarity restriction
px < 1
holds.
tation of (3.48) is that the mean long-term volatility level has to be suciently high relative to the strength of the innovation impact (recall from (3.40) that i.e. stated as
nx
mx nx /(1 px )). The interpretation of (3.49), which can also be px > bv > 0, is that the persistence of the long-run component
has to be higher than the one of the short-run component and that the impact of the innovation(s) to the short-run component cannot be as strong as to outweigh the persistence. Hereafter we shall denote our parameter vector by
:= (rf , , nx , iv , ix , pv , px , gv )T
and the restricted parameter space
:= { :
where
(3.47)
(3.49)},
Rp , p = 8.
4.
t = 0.
P First, note that by the assumptions (3.1)(3.7) we have that w P GW N (0, 1) and v P . Using this and (3.1) yields rt |Ft1 N (t , vt ).
Hence, the conditional probability density function (PDF) of
rt |Ft1
is
(4.1)
12
MATT P. DZIUBINSKI
( vt wt )2 1 f (rt |Ft1 ) = exp 2vt 2vt 2 1 vt w t = exp 2vt 2vt 2 1 wt = exp . 2 2vt
PDFs are parametrized by a parameter vector as:
(4.2)
(4.3)
(4.4)
then the M-estimator using the log-likelihood of the sample over t = 0, 1, ..., N can be written = arg max Qn ()
N
(4.5)
1 Qn () = N
t ( )
t ( ) t=0
Note, that using proportional and monotonic transformations, we can state our problem for the purposes of minimization as follows:
= arg min Q n ( )
N
(4.9)
n () = Q
t=0
lt ()
(4.10) (4.11)
2 lt () = log(vt ) + wt .
For the numerical details, including objective function computation algorithm and analytical gradient formulas, we refer the reader to Dziubinski (2010).
5.
Estimation Results
Due to the results in Dziubinski (2010) we choose Conditionally-Uniform Feasible Grid Search (CUFGS) with Feasible Sequential Quadratic Programming (FSQP) to estimate the models. FSQP allows us to solve the constrained optimization problem (4.9), while coupling it with CUFGS enables us to widen the search space thus increasing the chance of convergence.
13
We use the S&P 500 index data to calculate the (log) returns.
We
t our model to both daily (source: Yahoo Finance, period 1/3/1950 7/22/2009) and high-frequency (5-minute) data (source: Price-Data.com S&P 500, period 4/21/198212/6/2007 from Price-Data.com). For the purposes of research reproducibility, we use the same starting values as the ones in the column Estimation Starting Values in Table 1 in Dziubinski (2010).
We have considered three methods of obtaining the standard errors the OPG method, the numerical Hessian, and the sandwich estimator. Since there are numerical issues present when inverting the numerical Hessian (even if it is obtained using analytical rst derivatives), we choose to report the OPG standard errors.
Note that looking at the estimates for two data sets sampled at dierent frequencies is a way to empirically investigate temporal aggregation properties of our model. See Zivot (2009, Section 3.4) for a discussion of temporal aggregation in a context of GARCH models. The estimates obtained using the FSQP-AL CUFGS optimization algo8 rithm appear in Table 2. The problems with the large standard errors (causing insignicance) were practically not encountered in case of the FSQP optimization (where we used sandwich estimation and only used OPG or Hessian errors in case of numerical problems; the only problem was with a NaN deal.
This conrms
our belief that the choice of the optimization method matters a great Unsurprisingly, as in the similar models in the term-structure components GARCH literature, there are still some issues with estimating
and
gv .
pv
and
px )
pv
and
px
px
to be optimal at
pv
px
generated by CUFGS.
The results for the FSQP-NL CUFGS optimization algorithm in Table 3 are mostly similar to those discussed above. It can be seen (looking
6Note,
that as Zivot (2009) reports, a poor choice of starting values can lead to
an ill-behaved log-likelihood and cause convergence problems, which is why we use the starting values that satisfy the non-negativity conditions.
7As
In fact, Fiorentini
et al. (1996) and Hafner and Herwartz (2008) report, in the context of GARCH estimation, that the analytical Hessian signicantly outperforms the approximation. However, in our model, this comes at a cost of calculating ensuring that the estimates
remain in and using 8(8+1) = 36 derivatives). Also, bootstrapping the errors due to Young's Theorem, 2 is a possibility.
gorithms see Dziubinski (2010).
8For a discussion of the FSQP-AL and the FSQP-NL CUFGS optimization al-
14
MATT P. DZIUBINSKI
at the estimates and associated standard errors) that estimated more accurately in this case.
gv
seems to be
Daily Data
5-minute Data
T r f n x i v i x p v p x g v ) n ( Q
Table 2.
14, 984 4 5.604 10 (1.048 104 ) 7.024 101 (1.392 10+0 ) 4.744 106 (5.998 107 ) 2.320 106 (4.376 107 ) 2.396 106 (9.433 107 ) 9.375 101 (3.896 103 ) 9.375 101 (7.828 103 ) 2.183 10+2 (N aN ) 1.29423 10+5
523, 068 2.685 10 (1.548 106 ) 4.392 10+0 (1.475 10+0 ) 1.047 107 (8.741 109 ) 2.742 108 (1.785 109 ) 7.714 108 (6.401 109 ) 9.157 101 (6.563 103 ) 9.157 101 (6.652 103 ) 1.595 10+1 (1.247 10+2 ) 6.7172 10+6
10
ses) obtained for the S&P 500 data using FSQP-AL CUFGS optimization algorithm.
Daily Data
5-minute Data
T r f n x iv i x p v p x g v ) n ( Q
Table 3.
14, 984 7.028 104 (1.174 104 ) 4.007 101 (1.488 10+0 ) 4.040 106 (8.989 107 ) 3.474 106 (1.910 108 ) 5.336 107 (1.097 107 ) 9.460 101 (4.777 103 ) 9.460 101 (4.312 102 ) 1.022 10+2 (4.339 10+0 ) 1.29362 10+5
523, 068 2.906 1029 (1.540 106 ) 3.913 10+0 (1.465 10+0 ) 1.049 107 (2.714 1010 ) 2.018 1010 (8.008 105 ) 1.047 107 (8.008 105 ) 9.155 101 (1.350 101 ) 9.155 101 (1.784 104 ) 5.031 10+2 (9.509 10+1 ) 6.71722 10+6
ses) obtained for the S&P 500 data using FSQP-NL CUFGS optimization algorithm.
6.
Option Pricing
We shall consider option pricing under our model. In general, there are several approaches to look at:
15
(1) CJOW risk-neutralization using the conditional moment generating function (MGF), based on Christoersen et al. (2008) (2) Monte-Carlo, Empirical Martingale Simulation (EMS), based on Duan and Simonato (1998) (3) Monte-Carlo, Empirical Martingale Correction (EMC), based on Chorro et al. (2010) (4) alternative risk-neutralization method, (5) dierent model specication and derivation of the pertinent non-negativity conditions: (a) change the source of randomness distribution with positive support, (b) change the
so that it follows a
and
tions in log terms in the spirit of an EGARCH model. The pricing formulas using the analytical methods might be harder to derive (and would often be infeasible in the case of exotic options). The disadvantage of the Monte-Carlo-based pricing methods might be slower performance and, besides, they might need further adjustment to ensure the martingale property, see Duan and Simonato (1998) and Chorro et al. (2010).
6.1.
CJOW-MGF Approach.
Christoersen et al. (2008) model we may, in principle, consider using the option-pricing formulas presented there. diculty arises in attempts to apply them.
option valuation one needs to derive the moment generating function (MGF) for the component GARCH process (provided in Appendix A of Christoersen et al. (2008)), specify the dynamics under the riskneutral measure
(2008)) and proceed with the option-valuation formula (given in section 4.4 of Christoersen et al. (2008)). The problem arises in the second step, the risk-neutralization. Following Christoersen et al. (2008) we need which requires that
EQ [exp(rt+1 )] = exp(rf ),
rt+1 log
with
(6.1)
1 Q t+1 = rf vt+1 . 2
This in turn implies that
(6.2)
1 Q wt +1 = wt+1 + ( + ) vt+1 . 2
(6.3)
16
MATT P. DZIUBINSKI
We also want to ensure the equality of the conditional variances under the two measures:
(6.4)
Q Q (wt gi vt )2 = (wt gi vt )2 , 1 Q gi = gi + + , 2 gx = 0.
i = v, x.
(6.5)
i = v, x.
(6.6)
Thus
1 Q gx =+ . 2
This means
(6.8)
1 Q gx = 0 = . 2
Hence, without restricting the market price of risk
(6.9)
to a value which
Q-dynamic
is
going to remain such that we stay within our class of models (where we can apply the sucient conditions for non-negativity of the conditional variance). 6.2.
EMS is a variance-
reduction method ensuring the martingale property to be used with Monte Carlo pricing. The problem from our point of view is, however, that the EMS relies on the formulation of the model under the Q measure that is, a prior risk-neutralization. However, our model (similarly to CJOW) is stated under the P measure, so analytical riskneutralization would be required. But then, the one available method (CJOW, discussed above) is not applicable if we want to stay within our class of models. This excludes the EMS from any further considerations. 6.3.
tingale Correction method is, in fact, inspired by the EMS, see Chorro et al. (2010). The fundamental dierence is that it is applicable to the models stated under the P measure, such as ours. In this method, we make no assumption on the risk-neutralization (i.e. the shape of dQ ), and we the pricing kernel, involving Radon-Nikodym derivative dP compute prices for options with time to maturity (T t) by simulating sampled paths of the stochastic model under the historical measure P.
17
To rule out arbitrage opportunities, we directly impose risk neutrality th constraints. The i sampled historical nal price for the underlying is denoted by
ST,i .
The Empirical Martingale Correction works such that the previously sampled prices are replaced by:
T,i = S
The sampled average of
ST,i
1 N N i=1
ST,i
St er(T t) . St er(T t) ,
(6.10)
T,i S
is exactly equal to
risk neutral conditional expectation. With this approach, we only shift the historical distribution in a way that prevents arbitrage opportunities by implicitly changing the drift of this distribution. Chorro et al. (2010) compare this approach with the ane Stochastic Discount Factor (SDF) methodology in Cochrane (2002) and nd that the prices obtained by these two methods are close to each other.
6.4.
results in our model with those in the Black-Scholes-Merton model and Heston-Nandi GARCH(1, 1) model (HN). In this section we use daily data only. For comparison, we consider option pricing under the SCGARCH model using the estimates obtained using FSQP-AL and FSQP-NL (applying CUFGS in both cases). We estimate the HN model using fOptions R package for details, see Wuertz (2007). The estimation results are shown in Table 4.
= 14,984
Table 4.
3.451 10 1.139 10281 3.671 106 9.005 101 1.196 10+2 88559.65 0.953 7.806764 105
Daily Data +0
We present the pricing results across moneyness and maturity in Tables 58, with implied volatilities (IVs) reported in Tables 911. Note that Hull and White (1987) nd that the Black and Scholes (1973) model tends to overprice at-the-money (ATM) options and underprice
18
MATT P. DZIUBINSKI
deep in-the-money (ITM) and deep out-of-the-money (OTM) options in the presence of stochastic volatility. We can observe that ITM underpricing is even more pronounced in the SCGARCH (FSQP-NL estimates) model (this is also reected by the IVs for the SCGARCH model being consistently lower than for the HN model, especially for the ITM options recall that this is OptionPrice consistent with vega = being always positive for calls and puts), Table 8 similarly to the ane Heston (1993) stochastic volatility model (AF-SV) as reported in Christoersen et al. (2006). The prices for the SCGARCH (FSQP-AL estimates) model, Table 7 are practically quite close (this is reected in the IVs that are within 1% of each other, see Tables 1011) and reported mostly for completeness note, however, that the degree of the ITM underpricing is lesser than that of the SCGARCH (FSQP-NL estimates) model. However, as for the ATM options, HN (Table 6) overprices even more than BSM (Table 5) in contrast, note that lower prices in Table 8 suggest that this mispricing does not aect our model (or at least not as much). For the OTM options, both the HN and the SCGARCH (FSQP-NL estimates) models underprice more than the BSM model. We can also note that the SCGARCH (FSQP-NL estimates) model underprices (relatively to the other models) options with low strike prices and the long-maturity options. The long-maturity mispricing can be related to the one reported by Christoersen et al. (2010). Christoersen et al. (2010) introduce the generalized realized volatility (GRV) model (which nests the daily Heston and Nandi (2000) GARCH model as a special case, and also nests a variance dynamic with realized volatility only as a special case, referred to as the RV model) and develop another model where expected realized volatility is used in the variance dynamic in conjunction with squared returns (referred to as the GERV model, and the corresponding special case that only models expected realized volatility is referred to as the ERV model). Studying ve models in total: GRV, RV, GERV, ERV, and the benchmark Heston-Nandi GARCH model, Christoersen et al. (2010) show that all ve models tend to underprice long-maturity options (except for GRV) and overprice short-maturity options. Note, that for the very-short-maturity options (
= 1/24)
the HN
model exhibits a volatility smile, while the SCGARCH model exhibits a volatility smirk. However, for the longer maturities all models show a volatility smirk, presence of which being consistent with the empirical observations, see Christoersen et al. (2009).
19
K K K K K
BSM
Table 5.
year with
K,
underlying
(fraction of a
K K K K K
HN
Table 6.
underlying year with
K,
(fraction of a
K K K K K
AL
Table 7.
days).
K,
underlying
20
MATT P. DZIUBINSKI
K K K K K
NL
Table 8.
days).
K,
underlying
K K K K K
HN
Table 9.
underlying year with
K,
(fraction of a
K K K K K
HN
Table 10.
days).
K,
underlying
K K K K K
HN
Table 11.
days).
K,
underlying
21
One could add that another empirically interesting exercise would be to use the option prices in estimation, similarly to what Christoersen et al. (2008) suggested. Note, however, that in our model we do not use analytical option pricing formulas, but instead apply a Monte-Carlo method. On a single-core CPU (central processing unit) this method is too slow to be used in this application.
proach, however, would be to use parallelized many-core GPU (graphics processing unit) computation since MC is a so-called
embarrass-
improvements. In particular, in an application of MC pricing involving path-dependent options, Joshi (2010) demonstrates that it is possible 4 to get accuracy of 2 10 in less than a ftieth of a second, con-
cluding that GPU technology has rendered the Monte Carlo pricing of Asian options suciently fast that there is no longer any need for analytic approximations. This approach would also make possible to investigate forecasting properties of the model.
7.
Conclusions
This paper presents a discrete-time volatility model in which the underlying follows a process with conditional variance driven by the new Simplied Component GARCH process. It is a more parsimonious model than the CJOW one and allows us to derive sucient conditions for non-negativity of the conditional variance. Maximum likelihood estimation of the model is discussed. We provide an empirical illustration, applying the model to the S&P 500 index data. The results are consistent with our economic intuition. We propose an option pricing method consistent with our model
9As
K=
$1,200 options
in Table 8. for
We believe these are numerical artifacts which result from the slow
convergence rate of the MC procedure while we have also obtained "0" prices even
K = $1,100 when using 1,000 MC iterations (which only took 4.3 s to compute), 120 min per table (each with 5 distinct strikes and 5 distinct maturities,
Typically, S&P 500 data set containing 500 distinct
the results we report here have been obtained with 100,000 MC iterations, which require over giving
25
distinct options).
options per day (with 100 distinct strikes ranging from $900 to $1,900 and 5 distinct maturities of 15, 45, 75, 165, and 345 days) would require approximately
= 40
500 25 2 h h per 1 iteration of the optimization algorithm (since each iteration requires
day of
data. This underscores the fact that using single-core CPU is not enough for an empirically relevant estimation (with empirically relevant sample sizes) when using the MC pricing methods.
22
MATT P. DZIUBINSKI
The performance of the pricing method across moneyness and maturity is compared with that of the Heston-Nandi GARCH and Black-ScholesMerton models. The results of the comparison suggest that our model might constitute a particularly interesting choice for the valuation of the ATM options. Several of the future research directions and possible extensions to this work are worth consideration regarding to the advanced gridgeneration techniques and optimization algorithms and applications of GPUs allowing for more advanced pricing and forecasting applications. We provide a number of approaches to achieve that in the respective sections of this paper.
(2010):
The
Shape and Term Structure of the Index Option Smirk: Why Mul-
Christoffersen, P., K. Jacobs, B. Feunou, and N. Meddahi Christoffersen, P., K. Jacobs, and K. Mimouni
Index Options, (2010): The Economic Value of Realized Volatility,
Science,
Management
55, 19141932.
Working Paper.
(2006): An
Christoffersen, P., K. Jacobs, C. Ornthanalai, and Y. Wang (2008): Option valuation with long-run and short-run
volatility components, 297.
Christoffersen, P., K. Jacobs, and Y. Wang Cochrane, J. (2002): Asset Pricing, Duan, J.-C. and J.-G. Simonato Dziubinski, M. P.
Search (CUFGS), Simulation for Asset Prices, (2010): CIRANO Working Papers 2004s-56, CIRANO.
(2004):
tion Valuation with Long-run and Short-run Volatility Components, Princeton University Press. (1998): Empirical Martingale 44, 12181233.
Management Science,
Cointegration, Causality, and Forecasting: A Festschrift in Honor of Clive W.J. Granger, University Press, 475497.
23
(1996): Ana-
Jour-
Econometrica,
Nonlinear
871907.
Metrika,
67,
Hayashi, F. (2000): Econometrics, Princeton University Press. Heston, S. and S. Nandi (2000): A closed-form GARCH option Hull, J. C. and A. D. White IEEE Task P754 Joshi, M.
97107. (2010): valuation model,
13, 585625.
Point Arithmetic.
Journal of Finance, 42, 281300. IEEE 754-2008, Standard for FloatingWilmott Journal,
2,
Rmetrics.
Software
Collection-
Economics Letters,
CHAPTER 2
25
method (together with an algorithm for choosing the starting values) pertinent to the constrained optimization problem where the variables have to satisfy a sequentially dependent set of constraints. In practice, these arise (for instance) in the estimation of the models with inequality constraints, in particular GARCH models such as the Engle and Lee (1999) GARCH model and the Simplied Component GARCH (SCGARCH) model. We also provide algorithms for the objective function and analytical gradient computation for SCGARCH. The method improves upon ad hoc modications of unconstrained optimization algorithms (such as BFGS with a penalized objective function) prevalent in practice and is found to be more eective in nding the solution.
JEL Classication.
1. Introduction In this paper we present a numerical optimization method applicable to constrained optimization problems, where the variables have to satisfy a sequentially dependent set (SDS) of constraints (i.e., the one involving constraints involving multiple variables, where a variable constrained by a given relation is then involved in a subsequent constraint for another set of variables). The method, the Conditionally-Uniform Feasible Grid Search (CUFGS), is essentially a particular kind of a random grid search coupled with a constrained feasible Sequential Quadratic Programming (SQP) algorithm.
ence under constraints, nonlinear programming, performance of numerical algorithms, SCGARCH, sequential quadratic programming. We acknowledge nancial support by the Center for Research in Econometric Analysis of Time Series, CREATES, funded by the Danish National Research Foundation.
26
27
We apply it to the estimation of the Engle and Lee (1999) GARCH model (hereafter referred to as EL) and the Simplied Component GARCH Model (SCGARCH) of Dziubinski (2011). One of the reasons for developing it are the problems encountered when using non-specialized (unconstrained) gradient-based optimization algorithms due to constrained feasible space requirement and scaling. For example, in relation to the Broyden-Fletcher-GoldfarbShanno (BFGS) algorithm, popular with econometricians, Nocedal and Wright (2006) write:
(1)
(2)
BFGS updating is generally less eective for constrained problems than in the unconstrained case because of the requirement of maintaining a positive denite approximation to an underlying matrix that often does not have this property. SQP methods are most ecient if the number of active constraints is nearly as large as the number of variables, that is, if the number of free variables is relatively small. They require few evaluations of the functions, in comparison with augmented Lagrangian methods, and can be more robust on badly scaled problems.
Note that observation (2) in particular applies to the SCGARCH model, with
constraints and
variables.
As an unconstrained optimization algorithm, BFGS does not apply to constrained optimization problems per se. However, it is often used in conjunction with ad hoc modications (such as penalized objective function), which are generally less eective. In addition, it is also not a good t for the badly scaled or highly constrained models (i.e., ones where the number of active constraints is nearly as large as the number of variables) which is also often the case for the GARCH models with non-negativity inequality constraints. In contrast, our algorithm is specically designed for the SDS-constrained optimization problems. We illustrate its superior performance in two applications (estimation of two dierent GARCH models), both in a Monte Carlo study (with a known DGP, data-generating process) and on a real-world data (S&P 500 index returns). Another advantage of the CUFGS algorithm is that it be coupled with another algorithm other than FSQP (which extends it to the cases where the other optimization method is known, but when one still faces the problem of generating feasible starting values consistent with the constraints-SDS). In this study we focus on the CUFGS-FSQP coupling for empirical applications.
28
MATT P. DZIUBINSKI
The paper proceeds as follows. In Section 2 we present the SCGARCH model and the Engle and Lee (1999) GARCH model. We discuss estimation in Section 3. In Sections 4 and 5 we analyze the estimation algorithms and show the results. Section 6 contains our conclusions. In the Appendix, we also provide the objective function and analytical gradient computation algorithms for SCGARCH, useful for the practical implementation purposes.
2. The Models For reference we present the SCGARH and the EL models; for details see Dziubinski (2011) and Engle and Lee (1999), respectively. observed time-series (e.g., log return on the spot asset price), (over time steps of length (physical) probability measure The
follows
1) P,
rt+1 log
St+1 = t+1 + t+1 St vt+1 wt+1 t+1 = xt+1 = mx + px (xt mx ) + ix ux,t vt+1 = xt+1 + pv (vt xt ) + iv uv,t
with
1) 2gv vt wt 1)
w GW N (0, 1)
where
rf
, vt is the conditional variance of the log return between t 1 and t, with v P . The process w is a Gaussian white P P noise with mean 0 and variance 1, i.e., wt N (0, 1) t T and w W N (0, 1) (under probability measure P , w has mean 0 and covariance function (s, t) = |ts| , where h := 1{0} (h) is the Kronecker delta).
interval of length This model is a simplied specication of the Christoersen et al. (2008) model, solving the problem of ensuring non-negativity of the conditional variance. volatility components The sucient conditions for non-negativity of
and
29
SCGARCH := (rf , , nx , iv , ix , pv , px , gv )T
and the restricted parameter space by
(2.9)
(2.11)},
Now, we can obtain the Engle and Lee (1999) GARCH model from the SCGARCH model, replacing (2.5), (2.6), and (2.7) above with the following:
(2.12)
2 t 2 t
vt
xt
(2.13) (2.14)
are:
EL := (rf , nx , iv , ix , pv , px )T
and the restricted parameter space by
where
EL := {EL EL :
(2.15)
(2.17)},
px
i v + pv px < 0
pv
its own constraint. Constraints given by (2.9)-(2.11) involve three sequentially dependent 2 constraints (one non-linear and two linear, respectively): iv gv pv < 0, ix + iv nx < 0, pv px < 0 here, iv aects the rst two constraints (needed for
pv
and
nx ),
while
pv
px .
3. Maximum Likelihood Estimation A statistical method used for estimating the models is the Maximum Likelihood Estimation (MLE), which involves maximizing an objective function (called the (log)likelihood function) in order to obtain the estimates.
30
MATT P. DZIUBINSKI
Dziubinski (2011) shows that we can state our optimization problem as a constrained minimization problem:
= arg min Q N ()
N
(3.1)
N () = Q
t=0
lt ()
(3.2) (3.3)
2 lt () = log(vt ) + wt .
where and to
is a generic parameter specialized to EL for the EL model SCGARCH for the SCGARCH model (with the rest of the entities
specialized analogously).
4. A Simulation Study of Estimation Method Choice 4.1. Overview. In this section we provide an overview of some of the methods to estimate the models under consideration. We simulate each model using the coecient values given in Table 1 (they are interesting in practice, since they have the same magnitude as those in Table 1 of Dziubinski (2011)), omitting inapplicable parameters for the EL model, and estimate the parameters using the simulated data and the following algorithms : (1) NM - Nelder-Mead (2) FR - Fletcher-Reeves (3) PR - Polak-Ribire (4) BFGS - Broyden-Fletcher-Goldfarb-Shanno (5) SA - Simulated Annealing (6) PG - Projected Gradient (7) SPG - Spectral Projected Gradient (8) FSQP - Feasible Sequential Quadratic Programming
The Nelder-Mead (also known as downhill simplex) algorithm is a derivative-free optimization method, FR and PR are nonlinear conjugate gradient methods, BFGS is a quasi-Newton method. PG is an extension of the steepest descent method for unconstrained minimization, where a line search is performed over the direction of a projected gradient. SPG is similar to PG, except accelerated convergence due to the choice of the spectral step-length; see Birgin et al. (2000) for details.
1We
use
the
implementations for
thereof
provided
by in
O2scl C++
an see:
object-oriented
http://o2scl.sourceforge.net/.
library
numerical
programming
31
rf nx iv ix pv px gv n,SCGARCH Q n,EL Q
study. 15,000. model.
Simulation (DGP) 1
1.000 10 2.000 10+0 8.000 106 1.000 106 2.000 106 6.000 101 9.000 101 4.000 10+2 8.427097 10+3 125, 348
9.000 10 1.800 10+0 7.200 106 9.000 107 1.800 106 5.400 101 8.100 101 3.600 10+2 6.228086 10+3 115, 476 NEL
=
Sample sizes
NSCGARCH
= 1,000 and
SA is a probabilistic metaheuristic (convergence to an optimal solution is not guaranteed) for the global derivative-free optimization. timization problems. FSQP is a quadratic programming method applicable to the constrained opFor the rst four algorithms, see Nocedal and Wright (2006), for SA see Henderson et al. (2003) or Dreo et al. (2005), for PG see Kelley (1999) and for SPG see Birgin et al. (2000). FSQP is based on Sequential Quadratic Programming (SQP), modied so as to generate feasible iterates. Sequential quadratic programming (SQP) methods model the general constrained optimization problem by a quadratic programming subproblem at each iterate and dene the search direction to be the solution of this subproblem. It is important to design the quadratic subproblem so that it yields a good step for the nonlinear optimization problem, see Nocedal and Wright (2006). A description of FSQP can be found in Lawrence et al. (1997). There are two FSQP algorithms: FSQP-AL and FSQP-NL. In the FSQP-AL (monotone line search), an Armijo type arc search (hence the "A" in "FSQP-AL") is used with the property that the step of unit length is eventually accepted, which is a requirement for superlinear convergence. In the FSQP-NL algorithm the same eect is achieved by means of a nonmonotone search (hence the "N" in "FSQP-NL") along a straight line. In other words, in the FSQP-AL algorithm the objective function decreases at each iteration, while in FSQP-NL we have a decrease of the objective function within at most four iterations. For details, see Lawrence et al. (1997). For reproducibility purposes, we use the default seed for the Simulated Annealing. To minimize the warm-up period we use the unconditional
32
MATT P. DZIUBINSKI
mean of values
v and x to initialize v0 and x0 , that is, we set the to mx derived from the starting values of nx and px .
starting
Note, that McCullough and Renfro (1999) and Brooks et al. (2001) stress the importance of reporting the initial values provided to the GARCH estimation software. Without them, any conditional heteroscedasticity model is only partially specied, because the elements on which the likelihood is conditioned are not specied if the initial values are not given. McCullough and Renfro (1999) also report that the initialization of the series, though often overlooked, can substantially aect the solution produced by the software.
4.2. Constraints. In practice we may actually need a constrained minimization algorithm in order to ensure that the non-negativity conditions hold. Only the last three algorithms in our list are of this type (note, that PG and SPG only allow for hypercubic constraints, while FSQP allows for nonlinear functional constraints). A solution commonly used in practice, see for example in Rouah and Vainberg (2007), is to implement a penalty, so that violating nonnegativity conditions generates a large value of the objective function. For the methods requiring this modication, we adjust the algorithms as follows:
Penalized-Objective-Function Input: a parameter vector
l(N, , w, v, x, r) v0 , x0
TRUE
l
for
l(N, , w, v, x, r) Input: a parameter vector , starting values r0 , v0 , x0 Output: the gradient g l ( ) 1 if Positivity-Conditions( ) = TRUE
Penalized-Gradient
2 3 4 5 6 7 8
33
(2007), Penalty is assumed constant and large enough to have the order of magnitude larger than the objective function evaluated at starting values.
4.3. Results with the known DGP optimization study for the
SCGARCH model. For the SCGARCH model the maximum number of iterations is 1,000 in all of the cases The sample size
N = 1, 000.
PR 2
r f n x iv i x p v p x g v
Time (s)
) n ( Q
= 9.000 10 2.800 10+0 = 7.200 106 = 9.000 107 = 1.800 106 = 5.400 101 = 8.100 101 = 3.600 10+2 0.0 10+0 6.244133 10+3 999, 999.999.
NM 2
= 9.000 10 = 1.800 10+0 7.673 106 4.946 1013 2.786 106 = 5.400 101 = 8.100 101 = 3.600 10+2 1.3439 10+2 6.747140 10+3 =
FR 2
= 9.000 10 = 1.800 10+0 7.671 106 3.834 1014 2.748 106 = 5.400 101 = 8.100 101 = 3.600 10+2 1.34516 10+2 6.741407 10+3
the estimation study using NM, FR and PR. Penalty set to Symbol indicates that the values are equal to the initial values.
We use the starting values reported in the third column (denoted Estimation Starting Values) of Table 1. The results are reported in Tables 2 and 3. We notice that almost every unconstrained optimization algorithm performs unsatisfactorily. This may be due to the diculty ensuring that the estimates remain in
this case introduces non-smoothness of the objective function, whereas many of the algorithms require smoothness for convergence. Similarly, Zivot (2009) reports that the GARCH log-likelihood function is not
2The
reason for choosing this criterion as opposed to, say, maximum elapsed
time, is to ensure the reproducibility of the results, independent of the performance of the computer hardware.
34
MATT P. DZIUBINSKI
SA
PG
SPG
r f n x iv i x p v p x g v
Time (s)
) n ( Q
9.008 102 1.800 10+0 3.917 105 1.079 106 3.702 106 5.398 101 8.102 101 3.600 10+2 1.72 101 7.625983 10+3 999, 999.999.
= 9.000 102 = 1.800 10+0 7.666 106 7.248 10233 2.670 106 = 5.400 101 = 8.100 101 = 3.600 10+2 2.937 100 6.729235 10+3 =
9.003 102 = 1.800 10+0 5.381 105 4.167 106 2.960 107 = 5.400 101 = 8.100 101 = 3.600 10+2 2.75 100 7.609727 10+3
the estimation study using SA, PG and SPG. Penalty set to Symbol indicates that the values are equal to the initial values.
r f n x iv i x p v p x g v
Time (s)
) n ( Q
9.908070 10 1.557766 10+1 3.017787 105 3.068260 107 2.191491 106 2.525282 101 6.252131 101 3.583449 10+2 1.88 101 8.432699 10+3
FSQP-AL 2
1.084674 10 9.986293 10+1 3.007165 105 1.509107 107 2.295927 106 1.805605 101 6.280134 101 3.623615 10+2 7.66 101 8.435838 10+3
FSQP-NL 1
always well behaved, which may cause diculties for standard optimization techniques, especially when one takes into account the need to ensure that the positive variance and stationarity constraints hold. Simulated annealing, as a local search algorithm (metaheuristic) capable of escaping local optima by allowing hill-climbing moves, fares better in our benchmark. As it is not gradient-based, smoothness is not an important requirement. One may also notice that in some cases the only arguments changing signicantly are
nx , i v
ately high sensitivity of the objective function to those three arguments. In particular, compare the value of
n Q
for
35
in Table 1 to the one returned in case of the NM algorithm in Table 2 a change in one variable () is enough to signicantly alter the value of the objective function. We suspect similar behavior in case of and
nx , iv
ix
Below we present the summary of the estimation study results for the rst seven algorithms: (1) NM: only the value of
has changed,
(2) FR: the value of ix has improved compared to (1) (which probably led to an improvement in the objective function value), the
iv
(3) PR: very similar to FR, other than iv is an order of magnitude further from the true value, (4) BFGS: BFGS failed to produce results dierent from the start+2 ing values after 2.628280 10 s, (5) SA: better than all of the above, (6) PG: exhibits a problem with
iv ,
formance (on par with NM, FR, PR), (7) SPG: nonmonotone line search combined with spectral choice of step length lead to a signicant improvement compared to PG;
iv
nx
, pv , px , gv
Finally, we discuss the results in Table 4 for FSQP algorithm and consider two variants of it. What is striking is that the objective function values have improved signicantly compared to those obtained by the previously discussed algorithms. estimation. The estimate that is worth of attention is
This suggests that FSQP might be the best choice for the
eter appears dicult, possibly less so in the FSQP-AL case. However, the problem observed in all of the above algorithms, is that the estimates quite often remain equal to their initial values. This case is common in the component GARCH models estimation issues were also reported by Christoersen et al. (2008). Furthermore, it is worth noting that in the optimization settings we have used a very large interval (-100, 100) for the allowed values of the
estimate compared
3In
fact, they are slightly better than those corresponding to the DGP we
believe the reason is the nite sample of our simulation study and the eects of conditioning on the initial values not dying o.
36
MATT P. DZIUBINSKI
reasonable (and realistic) range, depending on the beliefs and the experience of the researcher. For example, limiting it to (1, 3), we obtain an estimate of 1.72 by FSQP-AL and 1 by FSQP-NL (note the corner case). This seems to suggest that FSQP-AL deals better with this problem than FSQP-NL. We have also chosen to report the more negative results so as to not create an impression of an unfair treatment compared to the other algorithms. As for the other results, the ones for the estimates of sample
rf , nx , ix
and
pv
and
px
pv ).
N = 1, 000.
In order to nd out whether rescaling the parameters to the same magnitude would yield an improvement, we simulate the model with DGP parameters set to 1.0e-001,
rf
= 1.0e-001,
= 5.0e-001,
nx
= 4.0e-001,
iv
ix
= 2.0e-001,
pv
= 2.5e-001,
px
= 7.5e-001,
gv
= 1.0e+000.
However, we experience similar diculties as in the original case. Furthermore, one may argue that place.
ex ante
4.4. Results with the known DGP optimization study for the
EL model. In this section we discuss the results for the Engle and
Lee (1999) GARCH model. In the following, we present the summary of the estimation study results for all the algorithms (here, FSQP denotes the FSQP-NL variant) using the articially generated data with known DGP:
(1) BFGS: failed to produce results dierent from the starting values, RMSE: (2)
(3)
) = 115476, n ( 0.00695651, Q FR: other than rf and nx , failed to produce results dierent ) = 125113, n ( from the starting values, RMSE: 0.00148766, Q ) = n ( FSQP: all parameters estimated, RMSE: 0.000641846, Q 125335, ) = 125107, n ( 0.00156566, Q
(5) PG: similar to FR, RMSE: (7) SA: exactly as in BFGS (8) SPG: RMSE:
) = 17938. n ( 0.786488, Q
4Note,
that the starting values, reported in the third column (denoted Esti-
mation Starting Values) of Table 1, were chosen to be very close to the DGP parameter values, reported in the second column (denoted Simulation (DGP)) i.e.,
90%
37
Here, it is clear that FSQP is the algorithm of choice, both given the RMSE and the objective function value. 4.5. Choices. There are several directions one could investigate as far as estimation is concerned:
using smoothly changing constraint-violation penalty function for instance, as in cont_constraint in O2scl , using constrained minimization algorithms allowing for constraints of type (2006), ods allowing for inequality constraints in Nocedal and Wright
using the augmented Lagrangian method, see Conn et al. (1991), using constrained NLP, as in Altay-Salih et al. (2003).
Our solution is in the spirit of the last one, since we use specialized algorithms designed for optimization with inequality constraints, i.e. FSQP-AL and FSQP-NL. In addition, we have also considered Simulated Annealing, since it is reasonably fast and performs relatively well in the class of the unconstrained optimization algorithms.
5. Starting Values Selection Algorithm As our next step, we consider a practical problem in model estimation: choosing the starting values. We consider an optimization method which might be very useful in practice. It belongs to the class of grid search methods, which means optimizing without choosing the initial parameter values. First, we consider the naive grid generation, using unconditionally uniformly distributed initial values and box constraints for the support; for the SCGARCH model:
rf R, R, nx > 0, iv > 0, pv (0, 1), px (0, 1), gv R, and analogously for the EL model. We nd that the
ing feasible starting values this way takes a very signicant amount of
naive grid search method performs very poorly in practice. Generattime (> 1.0e+3 s). Even though the FSQP algorithm we use allows for non-feasible starting values and begins by searching for feasible ones, it takes too much time to use this method in practice. None of the algorithms was able to get close to the results obtained with the starting values in Table 1. Therefore, we introduce a grid search method which we shall refer to as the Conditionally-Uniform Feasible Grid Search (CUFGS). The idea behind it is to solve the main weakness of the naive method: the unacceptably long amount of time it takes to generate feasible initial
5See
http://o2scl.sourceforge.net/o2scl/html/minimize_8h.html.
38
MATT P. DZIUBINSKI
values. The way to solve this problem is to generate the starting values sequentially, starting by obtaining the parameters which are not aected by constraints
bution. Next, the constrained parameters are generated conditioning on the ones associated with their constraints. The choice of the conditional distributions ensures that the values of the generated parameters satisfy the constraints by denition. First, we provide the generic algorithm, applicable to arbitrary constrained optimization problems (for an arbitrary model Model), where the variables have to satisfy a sequentially dependent set of constraints. Subsequently, we provide two specializations, for the SCGARCH model and for the Engle-Lee (1999) GARCH model, respectively. The concept underlying (and thus required by) our generic algorithm is sequential dependence of the constraints (and acyclicity thereof, as we discuss in the following section). generic programming. 5.1. Sequentially Dependent Constraints Terminology. We have shown two practical examples of sequentially dependent constraints: a set See Dos Reis and Jrvi (2005) for details on
SCGARCH
EL
correspond-
ing to (2.15)-(2.17). Before proceeding with the development of the optimization algorithm, it is useful to introduce a few basic notions from graph theory and formalize the notion of SDS-constraints see Harary (1991) for details on graph theory. A graph is an ordered pair and a set
G = (V, E )
comprising a set
of vertices
V ).
for instance,
px
is a vertex.
D = (V, A)
with
V x
a is
a = (x, y )
is directed from
to
y, y
a.
dependence e.g., the fact that in order to verify whether its constraint one rst needs to know is expressed by the presence of an arrow The degree of a vertex
px
satises
pv
(pv , px )
v,
denoted
deg(v ),
dent to it. An isolated vertex is a vertex with degree zero (in our case, this corresponds to a parameter which does not have to satisfy any constraints dependent on the other parameters). The indegree of a vertex
6Inequality
39
deg (v ), is the number of the adjacent head endpoints. A vertex with deg (v ) = 0 is called a source. The outdgree of a vertex v , denoted deg+ (v ), is the number of the adjacent tail endpoints. A + vertex with deg (v ) = 0 is called a sink. In our case, we know that we v,
denoted have to know the parameters corresponding to the sources in order to verify the constraints of the parameters corresponding to the sinks. A path is a list of vertices of a graph where each vertex has an edge from it to the next vertex. A directed acyclic graph (DAG), is a directed graph with no directed cycles i.e., with no path that starts and ends at the same vertex. In a (nite) DAG there exists at least one vertex with no incoming edges and one vertex with no outgoing edges. A topological ordering of a DAG is a linear ordering of its vertices such that, for every edge
a, b, a
precedes
b.
graph has no directed cycles that is, if it is a DAG. We assume that SDS-constraints correspond to a DAG expressing the dependence. An assumption of a DAG is quite natural in our context, since it corresponds to no circular dependencies in the constraints imposed on the parameters. A topological ordering is precisely the one we need for the complete generation-verication cycle for all the constrained parameters involved in a given model. For any DAG there exists at least one topological ordering, and there exist linear-time topological ordering algorithms for arbitrary DAG. Now, in the context of sequentially dependent constraints given by the set
SCGARCH
spond to the parameters and the sequentially dependent constraints correspond to the arrows, we have the following DAG Figure 1):
DSCGARCH
(see
VSCGARCH = {rf , , iv , ix , gv , pv , px , nx }
EL
DEL
40
MATT P. DZIUBINSKI
rf
gv
iv
ix
2 pv > iv gv
2 pv > iv gv
nx > iv + ix
nx > iv + ix
pv
nx
px > pv
px
Figure 1. Directed acyclic graph (DAG)
DSCGARCH
SCGARCH
corresponding to (2.9)-(2.11),
where the vertices correspond to the parameters and the arrows correspond to the sequentially dependent constraints.
41
rf
nx
iv
ix
pv > ix
px > iv + pv
pv
px > iv + pv
px
Figure 2. Directed acyclic graph (DAG)
DEL
resulting
EL
tices correspond to the parameters and the arrows correspond to the sequentially dependent constraints.
42
MATT P. DZIUBINSKI
5.2. The Algorithm. We now introduce the generic algorithm, generating the CUF parameters consistent with the SDS of constraints:
Generate-CUF-Parameters<Model> Input: sequentially dependent constraints
M odel
0 1 2 3 4 5 6 7 8
M odel
Sort all the remaining variables in accordance with topological ordo for each group of variables, in the order determined by the topodo Generate variables corresponding to the sources while there remain variables corresponding to the sinks with no Generate the variables corresponding to the sinks until last (as given by the topological ordering) variable obtained return
Recall that in a (nite) DAG there exists at least one vertex with no incoming edges (which ensures the initiation of the algorithm) and that for any DAG there exists at least one topological ordering (which ensures no cycles will prevent the termination of the algorithm).
5.2.1.
EL
Output: the set of CUF parameters satisfying positivity conditions
EL
1 2 3 4 5 6 7 8 9 10
Generate do
Generate ix while do
pv )
U (bliv , buiv ), pv U (blpv , bupv ) blpx iv + pv (feasible lower bound for pv ) while blpx > bupx Generate px U (pv , bupx ) return EL := (rf , nx , iv , ix , pv , px )T
This procedure ensures that the generated initial values satisfy (2.15)(2.17). It is also much faster than searching for the feasible region by
43
optimization, as it only relies on the very fast (conditionally-)uniform pseudo-random number generator. 5.2.2.
SCGARCH model.
is as follows:
Generate-CUF-Parameters<SCGARCH> Input: lower (bl) and upper (bu) box bounds, inequality constraints
SCGARCH
Output: the set of CUF parameters satisfying positivity conditions
SCGARCH
Generate do
1 2 3
Generate iv
4 5 6 7
blpv blnx while blpv > bupv OR blnx > bunx Generate nx U (blnx , bunx ), pv U (blpv , bupv ), px U (pv , bupx ) return := (rf , , nx , iv , ix , pv , px , gv )T
U (bliv , buiv ), ix U (blix , buix ), gv 2 iv gv (feasible lower bound for pv ) iv + ix (feasible lower bound for nx )
This procedure ensures that the generated initial values satisfy (2.9)(2.11). It is also much faster than searching for the feasible region by optimization, as it only relies on the very fast (conditionally-)uniform pseudo-random number generator. 5.2.3.
Search a given number of times. Each time we obtain the starting values by using Generate-CUF-Parameters and passing them to the optimization algorithm. We keep track of the best objective function value achieved in each iteration. Finally, we return the overall best result. This allows us to optimize by only specifying (very rough) lower and upper bounds for the parameters, instead of manually choosing the starting values. Table 5 contains a summary of the CUFGS estimation study results. The best objective function value achieved and the number of iterations as well as the computation time in seconds are reported.
It is seen from the table that the dierence between the specialized and non-specialized algorithms is quite pronounced. Only the FSQP manages to achieve the results comparable to those obtained when the optimization process is started from manually-chosen starting values close to the DGP ones. It is also interesting to note the trade-o between the convergence properties (number of iterations) and the computational cost of each iteration. It is not always the case that the algorithms with
7To
44
MATT P. DZIUBINSKI
Iterations
Time (s)
999, 999.999.
faster convergence properties (a lower number of necessary iterations) perform best, since each iteration may be suciently costly to oset any gains in accuracy. The results for the FSQP CUFGS optimization are shown in more detail in Table 6. We note that FSQP-AL and FSQP-NL, FSQP-AL, in the case of FSQP-NL, for case of
rf
nx and iv are estimated slightly better by ix FSQP-AL is much closer to DGP than
and
pv FSQP-AL gets very close to DGP, whereas in the px FSQP-NL is quite close to DGP. Finally, the estimates of gv are not as good as the ones resulting from manually chosen gv .
starting values that were close to the DGP. Our conclusion is that it is worth to try both FSQP-AL and FSQP-NL and to experiment with the choice of lower and upper bounds for and Prior experience with the estimation of those two parameters might lead to a choice of the narrower bounds containing only "reasonable" values, which then leads to good parameter estimates. In our case we have allowed the to vary from -10 to 10 and
gv
We also note that some of the nal iterations in both cases (AL and NL) lead to numerically close objective function values (indicating possible atness of the objective function), so looking at several sets of the estimates (instead of only the best ones) might also be advisable if the values are suciently close to being numerically indistinguishable.
45
FSQP-AL
FSQP-NL
r f n x iv i x p v p x g v
Time (s)
) n ( Q
1.012 101 1.000 10+1 2.718 105 2.189 106 8.009 107 6.594 101 6.625 101 3.523 10+0 4.7218 10+1 8.43392 10+3
1.012 101 1.000 10+1 8.294 108 5.399 108 4.307 1037 9.921 101 9.990 101 9.986 10+2 2.025 10+1 8.4345 10+3
5.2.4.
In Table 7 we
present the results of the Monte Carlo (MC) estimation study (where in each MC iteration we simulate and then estimate a model with sample size
N = 15, 000) for the CUFGS in conjunction with all the algorithms
(here, FSQP denotes the FSQP-NL variant; all the algorithms were allowed to run for 36 h, we report the averages over the best results for each MC iteration) using the articially generated data with known DGP (as in Section 4.4).
Objective fn. value -41,466.75 -124,260.4 -124,896.5 -54,131.2 -123,933.1 -124,448.1 -68,404.34
We can see that CUFGS-FSQP leads to average RMSE which is an order of magnitude better than the next-best results of FR, PG, and PR all of which are also signicantly slower (by several orders of magnitude) than CUFGS-FSQP.
46
MATT P. DZIUBINSKI
6. Empirical Estimation Results As a practical example, we use estimation of both models under consideration on the S & P 500 index data. Due to results in the previous section we have chosen FSQP and SA to estimate the model parameters for SCGARCH. In this section we provide a brief summary of the methodology and the results from the numerical perspective and refer the reader to Dziubinski (2011) for further details. 6.1. Results for the SCGARCH model. For the purposes of research reproducibility, we report the starting values in column Estimation Starting Values in Table 1. 811. The estimates obtained using the SA optimization algorithm, reported in Table 8, are quite stable over the sampling frequency. One of the issues, however, is that some of the coecients are statistically insignificant as their standard errors are large. Because of that, we do not put trust in those results also since there are several scaling issues when using the SA algorithm for optimization; we note, that the objective function values are inferior to those obtained using FSQP, reported in Tables 1011. Next, we consider the estimates obtained using FSQP-AL CUFGS optimization algorithm, reported in Table 10. The objective function This strengthens values improve and large standard errors (indicating insignicance) are practically not encountered in FSQP optimization. deal in this regard. The results for the FSQP-NL CUFGS optimization algorithm, reported in Table 11, are mostly similar to those discussed above, except that in this case the estimate of our belief that the choice of the optimization algorithm matters a great
gv
is more accurate.
In order to examine the scaling issues with the SA algorithm, we also perform the optimization with rescaling. The procedure is as follows: (1) rescale the
ders of magnitude from the FSQP CUFGS results (note that this requires prior estimation using FSQP CUFGS), (2) use starting values equal to
1.0
(3) perform SA CUFGS optimization (grid search is necessary, since we nd that rescaled non-grid optimization fails to converge with given starting values),
8Note,
that as Zivot (2009) reports, poor choice of starting values can lead to an
ill-behaved log-likelihood and cause convergence problems this is why we use the starting values that satisfy the non-negativity conditions.
47
(4) store the results of the SA CUFGS optimization, (5) use the stored results in non-scaled non-grid SA optimization to obtain nal results. We report the results in Table 9. We note the improvements in the However, we
objective function value in case of low-frequency data. CUFGS for the following reasons:
(1) prior estimation using dierent algorithm (such as FSQP CUFGS) is still necessary for this method we nd SA to be very sensitive to scaling, (2) it is impractically slow not only due to prior estimation requirements, but also because SA CUFGS takes a considerable amount of time to nd feasible solutions (> 5 min for lowfrequency data, > 20 min for high-frequency data) and then to improve upon them in the nal step (> 2.5 min, > 10 min, respectively). In comparison, FSQP-AL needed < 4 min to converge for high-frequency data.
48
MATT P. DZIUBINSKI
Daily Data
5-minute Data
T r f n x i v i x p v p x g v ) Qn (
8.621144 10 1.800082 10+0 1.315838 103 1.482922 106 3.957780 104 5.413485 101 8.099383 101 3.599995 10+2
14, 984 (3.693460 102 ) (4.730217 10+0 ) (1.369405 102 ) (7.798768 103 ) (6.343736 103 ) (3.478335 10+1 ) (1.644191 10+0 ) (1.912336 10+6 ) 5.334414 10+4
8.639511 10 1.800325 10+0 1.118116 103 8.790822 107 9.997710 104 5.414670 101 8.099021 101 3.599994 10+2
523, 068 (3.500509 102 ) (4.542920 10+0 ) (1.689276 102 ) (1.806808 102 ) (1.763490 102 ) (1.278797 10+2 ) (1.846139 10+0 ) (7.410604 10+6 ) 1.862679 10+6
ses) obtained for the S&P 500 data using SA optimization algorithm (without rescaling).
Daily Data
5-minute Data
T r f n x iv i x p v p x g v ) n ( Q
4.399291 10 6.048738 101 4.985880 106 2.501975 106 2.473879 106 9.419734 101 9.419895 101 2.138208 10+2
14, 984 (7.354550 105 ) (1.174593 10+0 ) (6.485218 107 ) (2.524714 107 ) (3.701944 107 ) (3.543022 103 ) (7.269373 103 ) (3.375485 10+1 ) 1.295084 10+5
523, 068 3.192352 10 (1.518713 106 ) 8.046584 10+0 (1.453661 10+0 ) 1.116417 107 (1.222634 109 ) 1.116159 107 (7.681806 1010 ) 1.260360 1013 (3.870117 1010 ) 9.108052 101 (6.632709 104 ) 9.109176 101 (8.665883 104 ) 1.035382 10+2 (4.420274 10+1 ) 6.716776 10+6
15
ses) obtained for the S&P 500 data using SA optimization algorithm (with rescaling).
49
Daily Data
5-minute Data
T r f n x i v i x p v p x g v ) Qn (
14, 984 4 5.604 10 (1.048 104 ) 7.024 101 (1.392 10+0 ) 4.744 106 (5.998 107 ) 2.320 106 (4.376 107 ) 2.396 106 (9.433 107 ) 9.375 101 (3.896 103 ) 9.375 101 (7.828 103 ) 2.183 10+2 (N aN ) 1.29423 10+5
523, 068 2.685 10 (1.548 106 ) 4.392 10+0 (1.475 10+0 ) 1.047 107 (8.741 109 ) 2.742 108 (1.785 109 ) 7.714 108 (6.401 109 ) 9.157 101 (6.563 103 ) 9.157 101 (6.652 103 ) 1.595 10+1 (1.247 10+2 ) 6.7172 10+6
10
ses) obtained for the S&P 500 data using FSQP-AL CUFGS optimization algorithm.
Daily Data
5-minute Data
T r f n x iv i x p v p x g v ) n ( Q
14, 984 7.028 10 (1.174 104 ) 4.007 101 (1.488 10+0 ) 4.040 106 (8.989 107 ) 3.474 106 (1.910 108 ) 5.336 107 (1.097 107 ) 9.460 101 (4.777 103 ) 9.460 101 (4.312 102 ) 1.022 10+2 (4.339 10+0 ) 1.29362 10+5
4
523, 068 2.906 10 (1.540 106 ) +0 3.913 10 (1.465 10+0 ) 1.049 107 (2.714 1010 ) 2.018 1010 (8.008 105 ) 1.047 107 (8.008 105 ) 9.155 101 (1.350 101 ) 9.155 101 (1.784 104 ) 5.031 10+2 (9.509 10+1 ) 6.71722 10+6
29
ses) obtained for the S&P 500 data using FSQP-NL CUFGS optimization algorithm.
50
MATT P. DZIUBINSKI
6.2.1.
the estimation study results for all the algorithms (here, FSQP denotes the FSQP-NL variant) on the S&P 500 data: (1) BFGS: failed to produce results dierent from the starting values, RMSE: (2)
(3)
) = 120226, n ( 0.010051, Q FR: other than rf and nx , failed to produce results dierent ) = 128694, n ( from the starting values, RMSE: 0.0123022, Q ) = n ( FSQP: all parameters estimated, RMSE: 0.0133402, Q 129904, ) = 128694, n ( 0.0123018, Q
(5) PG: similar to FR, RMSE: (7) SA: exactly as in BFGS (8) SPG: RMSE:
) = 17917.7. n ( 0.795278, Q
While the RMSEs are comparable, FSQP once again achieved the best objective function value.
6.2.2.
mation study results for the CUFGS in conjunction with all the algorithms (here, FSQP denotes the FSQP-NL variant; all the algorithms were allowed to run for 96 h, we report the best results for each) on the S&P 500 data: (1) BFGS: Found minimum (2) FR: Found minimum (3) (4) (5) (6) (7) (8)
44, 551.8 after 410 iterations (329, 032s), 128, 567 after 1, 281 iterations (149, 390s), FSQP: Found minimum 129, 906 after 5 iterations (14.71s), NM: Found minimum 79, 527.7 after 719 iterations (2.71s), PG: Found minimum 128, 577 after 1, 281 iterations (74, 237.9s), PR: Found minimum 128, 681 after 812 iterations (95, 326.1s), SA: Found minimum 99, 565.2 after 117, 122 iterations (440.78s), SPG: Found minimum 93, 315.9 after 719 iterations (74, 115.1s).
Again, both the achieved minimum and the estimation time clearly favor the CUFGS-FSQP algorithm note that it's especially impressive that CUFGS-FSQP managed to improve upon the results of FSQP, even though FSQP was given manually-tuned starting values (based on Table 1, corresponding to the available estimates of component GARCH model also using the S&P 500 index data), while CUFGSFSQP requires no starting values (since the CUFGS itself is used to obtain them).
51
7. Conclusions In this paper we present and evaluate numerical optimization methods (together with an algorithm for choosing the starting values) pertinent to the constrained optimization problem where the variables have to satisfy a sequentially dependent set of constraints. In practice, these arise (for instance) in the estimation of the models with inequality constraints, in particular GARCH models such as the Engle and Lee (1999) GARCH model and the Simplied Component GARCH (SCGARCH) model of Dziubinski (2011). We also provide algorithms for the objective function and analytical gradient computation for SCGARCH. We discuss and provide a benchmark of several optimization routines applied to the estimation problems arising in the above-mentioned GARCH models. We nd that the FSQP algorithms (FSQP-AL and FSQP-NL) have the best performance of the algorithms we consider for all the models under consideration both in the study using a known DGP and in the empirical study (using the S&P 500 index data). We conclude that the CUFGS algorithm can be protably used in the selection of starting values for the estimation. We nd that the FSQP algorithms coupled with CUFGS perform reasonably well and signicantly better than the non-specialized optimization algorithms. We believe that our algorithm is applicable to a wide-range of nonlinearly constrained optimization problems, where there is a sequential functional dependence among the constraints on the variables. The algorithm is especially useful in practical econometric modeling, where the choice of starting values is often not an obvious one, especially for the end-users unfamiliar with the given modeling framework, or where the model is new and there is no well-known range of parameters' bounds, or when dealing with new (previously unencountered) data sets.
52
MATT P. DZIUBINSKI
Con-
SIAM
International Jourand
Ornthanalai,
ally Convergent Augmented Lagrangian Algorithm for Optimization with General Constraints and Simple Bounds,
Proceedings of the First International Workshop of Library-Centric Software Design (LCSD '05). An OOPSLA '05 workshop.
in
Dziubinski, M. P. (2011):
per 2011-9.
Cointegration, Causality, and Forecasting: A Festschrift in Honor of Clive W.J. Granger, University Press, 475497. Harary, F. (1991): Graph theory, Addison-Wesley.
tory Component Model of Stock Return Volatility, in
Henderson, D., S. H. Jacobson, and A. W. Johnson (2003):
Metaheuristics,
Handbook of
Academic Publishers, International Series in Operations Research & Management Science, chap. 10, 287320.
Kelley, C. T. (1999): The Gradient Projection Algorithm, in
Iter-
53
for CFSQP Version 2.5: A C Code for Solving (Large Scale) Constrained Nonlinear (Minimax) Optimization Problems, Generating Iterates Satisfying All Inequality Constraints, Tech. rep., Institute for Systems Research, University of Maryland, Technical Report TR94-16r1.
McCullough, B. D. and C. G. Renfro (1999): Benchmarks and
software standards: A case study of GARCH procedures, Journal of Economic and Social Measurement, 25, 5971. Nocedal, J. and S. Wright (2006): Numerical optimization, Springer: Springer, 2. ed. ed.
Rouah, F. D. and G. Vainberg (2007): Zivot, E. (2009):
Springer,
54
MATT P. DZIUBINSKI
Appendix
7.1. Objective Function Computation. Here we provide an algorithm to compute the objective function for the SCGARCH model.
We can compute the summands (3.3) of (3.2) using the specication given in (2.1)(2.8). The following procedures implement the computation given in these equations:
Summand
l0 (, w, v, x, r)
starting values
1 2 3
v 0 , x0
Summand
1 2 3 4 5 6 7
lt+1 , values wt+1 , 2 ux,t (wt 1) 1/2 2 1) 2gv vt wt uv,t (wt xt+1 nx + px xt + ix ux,t vt+1 xt+1 + pv (vt xt ) + iv uv,t 1/2 wt+1 (rt+1 rf vt+1 )/vt+1 2 lt+1 log(vt+1 ) + wt +1 return lt+1
Finally, the procedure Objective-Function implements the computation of the objective function (3.2):
l(N, , w, v, x, r) Input: a parameter vector , starting values r0 , v0 , x0 N Output: the objective function l t=0 lt ( ) 1 l Summand l0 (, w, v, x, r) 2 for t 0 to (N 1) do 3 l l + Summand lt+1 (, w, v, x, r) 4 return l
Objective-Function
7.2. Analytical Gradient. To implement MLE in practice it is useful to have the analytical gradient. There are at least two reasons for that. First, in case of GARCH models estimation using gradient-based optimization the analytical gradient is more accurate than its numerical approximation (see (Zivot, 2009, Section 5.1) and Brooks et al. (2001)).
9The
reference
w, v, x
is
pass-by-
55
Second, it may also be applied for computing the outer-product gradient (OPG) estimate of the information matrix. Here is the pseudocode for the procedures implementing the computation of the analytical gradient for the objective function in the SCGARCH model:
Summand
l0 (, w, v, x, r)
observation
r0 ,
values
w 0 , v 0 , x0
1 2 3 4 5 6 7 8
lt+1 (, w, v, x, r) , observation rt+1 , values wt , vt , xt Output: summand lt+1 , values wt+1 , vt+1 , xt+1 2 1 ux,t wt 2 xt+1 px xt + ix ux,t 3 nx xt+1 nx xt+1 + 1 4 ix xt+1 ix xt+1 + ux,t 5 px xt+1 px xt+1 + xt 1/2 1/2 2 6 uv,t wt + gv wt vt vt + 2 gv vt wt 1/2 7 gv uv,t gv uv,t + 2 vt wt 8 vt+1 xt+1 + pv (vt xt ) + iv uv,t 9 iv vt+1 iv vt+1 + uv,t 10 pv vt+1 pv vt+1 + (vt xt ) 1 11 wt+1 1 vt +1 wt+1 vt+1 2 1/2 12 rf wt+1 rf wt+1 vt+1 1/2 13 wt+1 wt+1 vt+1 2 14 wt +1 2 wt+1 wt+1 1 2 15 lt+1 vt +1 v0 + wt+1 16 return lt+1
Summand Input: a parameter vector
Finally:
l(N, , w, v, x, r) Input: a parameter vector , starting Output: the gradient g l ( ) 1 l Summand l0 (, w, v, x, r) 2 g Summand l0 (, w, v, x, r ) 3 for t 0 to (N 1) do
Gradient
values
r0 , v0 , x0
56
MATT P. DZIUBINSKI
4 5 6
return g
Note, that the calculation of the gradient still requires the computation of
w, v
and
t T.
CHAPTER 3
57
CHRISTIAN BACH AND MATT P. DZIUBINSKI We introduce tractable models for commodity derivatives pricing with inventory and volatility eects, and illustrate with applications to the oil market. We contribute to the existing literature in several respects. First, whereas the previous literature uses futures data for investigating the relationship between inventory and volatility, we use the information available in options traded on futures. Second, performance assessment in the previous literature has primarily evolved around explaining moments of data or forecasting prices of futures. Instead, we asses the performance of our model by considering both the ability of explaining prices in-sample and out-of-sample assessing both the pricing-performance and the hedging-performance of the models. Third, we model the futures surface rather than the spot price process, and we limit the number of parameters to calibrate. We introduce a new, maturity-wise calibration method compatible with this modeling methodology. Fourth, we use actual data on inventories rather than a proxy. Fifth, our model is very exible and allows for analyzing several dierent types of relationships between inventory and volatility. JEL Classication. C51, C52, G12, G13, Q40.
Abstract.
IF
Introduction
he role of inventory in explining ommodity spot prie voltility is signi(ntF sn prtiulr onsidering the oil mrketD qemn @PHHSA writesX
CREATES, Department of Economics and Business, Denmark Business and
Secondary 62M99, 90B05, 91B70, 91G60. Key words and phrases. Energy futures and options markets, energy price volatility, commodities, crude oil, stochastic volatility, stochastic inventories, inventories, option pricing, scarcity. We acknowledge nancial support by the Center for Research in Econometric Analysis of Time Series, CREATES, funded by the Danish National Research Foundation.
58
E-mail addresses: matt@creates.au.dk. Date : August 7, 2012. 2010 Mathematics Subject Classication. Primary 62P05, 91B24, 91B25, 91G20;
59
F F F whenever there is downwrd djustment of the esE timted oil reserves in the or nother regionD the voltility of oil pries inreses shrply @nd pries s wellAF he role of inventories in explining prie nd voltility of ommodiE ties hve een studied in severl ppersF frennn @IWSVA nd elser @IWSVA re erly studies on the e'et of the level of inventory on griulE turl ommoditiesD ut the inventory e'et hs lso een doumented for metls @xg nd irrong @IWWRAA nd oil nd nturl gs mrkets @qemn nd yhn @PHHWAAF hese studies use proxies for inventories rther thn tul inventory dtF his pproh is ommon sine inE ventories of most ommodities re generlly not known to the mrketsF e ommon proxy used is the spred etween the IEmonth nd the IPE month futures ontrts @seeD eFgFD pm nd prenh @IWVUA nd pm nd prenh @IWVVAAF qemn nd yhn @PHHWA show tht this proxy for inventory is of mixed qulity depending on the ommodity of interestF snsted of relying on proxy for inventory dt we use weekly dt on oil inventoriesF qemn nd xguyen @PHHSA onstrut dtse of soyen inventory over IHEyer period nd show tht voltility n e written s n ext inverse funtion of inventoryF sn this pper we do not otin suh ler reltionship lthough we see strong signs of the reltionship etween inventory nd voltility eing negtiveF everl simple models hve een proposed to utilize the informtion in ommodity inventory levelsF wost of this literture inorportes the informtion in inventories in the prie dynmis of the spot prie @seeD eFgFD qison nd hwrtz @IWWHAD hwrtz @IWWUA nd outledge et lF @PHHHAA ndD in ordne with vst empiril reserhD models spot pries s menEreverting proessF1 enother wy to model the e'et of inventories is to let onveniene yield e funtion of inventoryF uldor @IWQWA nd orking @IWRWA suggest tht the reltionship is inverse @the so lled uldorEorking hypothesisAF sn the models we onsider in this pper we let inventory e seprte proess nd model the reltionship etween futures pries nd inventory through the orreltion strutureF2 wost ppers ssess model performne y onsidering the futures mrketD in terms of either explining the moments of futures @outledge et lF @PHHHAAD explining oserved futures or forwrds urves @qemn nd xguyen @PHHSAA or the performne of the model in foresting the prie of futures @qison nd hwrtz @IWWHAAF e ssess model performne oth y onsidering the ility to prie options left out of the oservtion set used for lirtion nd y the ility to forest option pries one dy hedF
e.g., Gibson and Schwartz (1990), Brennan (1991), Cortazar and Schwartz (1994), Bessembinder et al. (1995), and Ross (1997). 2In particular, through the non-zero instantaneous correlations of the Wiener Processes driving the price, volatility and/or inventory stochastic processes.
1See,
60
yur lirtion proedure is inspired y tht of ghristo'ersen et lF @PHHWA who lirte their model to weekly oservtions of pries on iuE ropen style optionsF e lirte ll our models on dily frequeny nd onsider oth the setting where one prmeterEset exists for ll fuE tures on given dy nd the setting where prmeterEset is lirted for eh futureF he ltter mounts to otining prmeterEset for eh di'erent future mturityF viquidity in rude oil futures nd emerin options is very highF rE ditionllyD iuropen options on rude oil re not s liquid s emerin options nd this poses n importnt trdeEo' in the hoie etween the option exerise typeF emerin option pries re omputtionlly signi(ntly more expensive to lulte nd when we seek to lirte on dily sisD the time required eomes too longF his neessittes onversion of emerin pries to iuropen priesD similrly to rolle nd hwrtz @PHHWAF e lirte to thus otined iuropen style option priesD overing the time period from eptemer PQD IWVV to 1 wy SD PHIID yielding over 22 2 yers of dt used in the lirtionF e ontriute to the existing literture in severl respetsF pirstD wheres the previous literture uses futures dt for investigting the reltionship etween inventory nd voltilityD we use the informtion ville in options trded on futuresF gonsidering the ft tht opE tions hve strikeEdimension it is plusile tht they ontin inreE mentl informtion over futuresF eondD performne ssessment in the previous literture hs primrily evolved round explining moE ments of dt or foresting pries of futuresF snstedD we ssess model performne y onsidering the ility oth of explining pries inE smple nd outEofEsmpleF snEsmple performne ssessment is done y using only suset of the options dt on eh dy for lirtion nd ttempting to forest the prie of the remining optionsF yutE ofEsmple performne ssessment is done y mking oneEdyEhed preditions of option priesF hirdD we onsider models for the futures pries rther thn spot pries @thusD modeling the futures surfeAD nd from the noEritrge reltionship etween spot nd futures pries we limit the numer of prmeters to lirteF pourthD we use tul dt on inventories rther thn proxyF pifthD our model is very )exE ile nd llows for nlyzing severl di'erent types of reltionships etween inventory nd voltilityF sn prtiulrD it is strightforwrd to only llow for n inverse reltionship etween inventory nd voltilityD when inventory is elow its verge levelD s proposed in qemn nd yhn @PHHWAF ixthD we onsider two di'erent lirtion pprohes nd show tht oth priing performne nd foresting performne n e improved gretly y lirting the prmeter set to eh fuE tures ontrt on eh dyF his is prtiulrly true for short mturity options nd fr outEofEtheEmoney optionsD wheres for long mturity options nd fr inEtheEmoney options oth methods fre eqully wellF
61
he pper proeeds s followsF sn etion P we present the dt nd motivte the use of inventory dt in modeling voltility of futuresF sn etion Q we present the models onsidered in the pperF etion R disusses lirtion methodsF etion S presents empiril resultsD etion T ompres the priing nd hedging performne ross models nd lirtion methodsD etion U ontins the results of enhmrks ginst the flk @IWUTA modelD nd etion V onludesF PF
Data and Initial Analysis
sn this setionD we (rst desrie the dt on optionsD futures nd inventory nd then present desriptive sttistisF e further study the reltionship etween inventory on the one hnd nd option prie nd voltility on the otherF PFIF Futures Data. sn the ommodity derivtives mrket the underE lying is usully not the ommodity itselfD ut rther futures ontrt written on the ommodityF e onsider the est exs sntermedite @sA rude oil futures trded on the xew ork werntile ixhnge @xwiA3 sine these re historilly the most liquid rude oil fuE tures nd we will refer to the prie of the front future s the oserved spot prieF putures trding on xwi ws initited in IWVQ nd the ontrts re stndrdizedD speifyingD eFgFD the minimum qulity of delivered rude oilD in terms sulfur ontent nd grvityD nd the ple of delivery @gushingD yklhomAF gontrts urrently trde TH onE seutive monthsD with longEdted heemer nd tune futures with up to W yers to mturity eing ville s wellF es is pprent from pigure PFID most of the volume is in ontrts with only few months to mturityD ut futures with up to I yer to mturityD s well s tune nd heemer ontrtsD in generlD re resonly liquidF rding in futures ontrt termintes t the lose of usiness on the third usiness dy prior to the PSth lendr dy of the month preeding the delivery monthD or the preeding usiness dy if the PSth is not usiness dyF por instneD the ontrt with I month to delivery in pigure PFI is future with mturity in perury nd the lst dy of trding in tht future is round the PPnd of tnuryF helivery tkes ple etween the (rst nd lst usiness dy of the delivery monthF he futures dt series is from gf rderF PFPF Options Data. e use dily losing quotes on emerin rude oil options trded on xwi during the period from eptemer IWVV to wy PHIIF4 he option dt series is from gf rder nd the xwi rude oil derivtives mrket is the most liquid derivtives mrE ket in the worldF ih option is on one futures ontrt with expiry in
3As of 2008 a subsidiary of the CME Group. 4European-style crude oil futures options have been trading on NYMEX since
2008, but liquidity is still substantially smaller than for American options.
62
x 10
Volume
10 15 Maturity (months)
20
the sme month s the optionF rding in n option ontrt termintes t the lose of usiness on the third usiness dy prior to expirtion of the underlying futures ontrtF he option ontrts trded re with similr mturity s for the futures ontrtsF he strike prie rule for the option series is to lwys hve twenty strike pries in inrements of 6HFS per rrel ove nd elow the tEtheEmoney @ewA strike prieF st furthermore hs the next IH strike pries in inrements of 6PFS per rrel oth inEtheEmoney @swA nd outEofEtheEmoney @ywAF his is lredy sustntil mount of strikes trded t eh mturityF es time progresses the oil prie hnges nd new ontrts re introdued suh tht the strike prie rule reminsF hereforeD t eh dy one n esily oserve more thn IDHHH option ontrts with di'erent mtuE rity nd strikeF st lso implies tht one often oserves fr sw nd fr yw optionsF he strike rule hs hnged through time nd in erly prts of the dtset fewer ontrts in oth the strike nd mturity dimension were trdedF
63
fefore (ltering the options nd futures dt onsists of 4, 845, 069 dily losing pries on options nd 238, 814 dily losing pries on fuE turesF5 ih option quote is mthed with the prie of the underlying futures ontrtF pollowing fkshi et lF @IWWUA we exlude ll options with less thn U dys to mturityD s well s options with extreme moneyness @PH7 in nd out of the moneyAF purthermoreD sine option pries re quoted in inrements of I entD we exlude options with prie of less thn 6HFSF6 yn eh dy there re options with severl di'erent mturities nd the riskEfree rte for eh of these mturities is lulted y interpolting the ville Eill rtesF PFPFIF Further Filtering Rules. riing of emerinEstyle options is sigE ni(ntly more omplex thn priing iuropenEstyle optionsF sn prtiE ulr when models eome more omplex thn the stndrd oneEftor geometri frownin motion @qfwA modelD lirting to emerinE style options is very time onsumingF hereforeD we tke stndrd pproh @seeD eFgFD frodie et lF @PHHUA nd rolle nd hwrtz @PHHWAA nd onvert the oserved emerin option pries to pries of orresponding iuropenEstyle optionF heres rolle nd hwrtz @PHHWA pply the formul y froneEedesi nd hley @IWVUA in the onversion we pply the tu nd hong @IWWWA formulD otined y introduing orretion terms to the froniEedesi nd hley formulD sine it is more urte for intermedite mturitiesF pirstD we k out the implied voltility from the tu formulF vet = (K, , F, r)D C JU (, ) e the emerin ll prie lulted with the tu formul with prmeterEset {, }D cBS (, ) the orreE sponding iuropenEstyle ll option nd C the orresponding oserved emerinEstyle ll optionF r is the riskEfree rteD F is the prie of the futures ontrtD K is strike prie of the optionF hen the implied voltility is otined @using the xewtonEphson methodA suh tht C = C JU (, )F eondD we otin the orresponding iuropenEstyle ll option prie s cBS (, )D where is the implied voltility from the (rst stepF he pproh is nlogous for put optionsF sn this fshion we onvert ll our oservtions to orresponding iuropenEstyle option pries nd n prie them s suhF e present desriptive sttistis using reltively detiled moneyE ness (lter @inluding QS7 sw nd yw optionsA nd mturity (lter @inluding options with QTH ` D where denotes timeEtoEmturityAF es mentioned erlierD nd s is lso seen from le PFID options with
Settlement prices are determined by the "Settlement Price Committee" at the end of regular trading hours. They are a very accurate measure of the true market prices at the time of close and we expect them to have no relevant inuence on our results. 6This implies that a price error of 2 percent might still remain due to discrete prices.
5We follow Trolle and Schwartz (2009) and base the analysis on settlement prices.
64
long mturities re widely villeF roweverD s ws seen in the preE vious setionD liquidity is thin in futures with more thn IP months to mturity nd s result we limit ourselves to onsidering options with no more thn IP months to mturity in the lirtion exeriseF prom les PFP nd PFQ the expeted inrese in ll pries in moneyE ness nd mturity is oservedF gompring the iuropenEstyle pries of le PFP with the emerin style pries of le PFQ we note tht the emerinEstyle premium is prtiulrly pronouned for short mE turity options slightly in the money nd long mturity options fr in the moneyF le PFR shows the verge implied voltility for eh of the moneyE ness nd mturity tegoriesF gontrry to ghristo'ersen et lF @PHHWA who doument voltility smirk for options on 8 SHH futuresD we oserve more or less perfet smile from the implied voltilities ross moneynessF sn prtiulr for ew nd sw options on 8 SHH fuE tures the implied voltility is onstnt ross mturityF his pttern is very di'erent from the one oserved in le PFRD where implied voltility is deresing shrply in mturityF por instneD for ew opE tions verge implied voltility is 37 perent for options with less thn I month to mturityD 31 perent for options with IVH to QTS dys to mturityD nd 25 perent for options with more thn I yer to mturityF his shows tht even onsidering ew options more )exile models thn the stndrd oneEftor qfw model re neededF he implied voltilities re otined from the flk @IWUTA model C = er F N (d1 ) KN d1
d1 =
ln (F/K ) + ( 2 /2) ,
where N is the umultive distriution funtion @ghpA of the stndrd norml distriutionD C is the ll option prieD r is the riskEfree rteD F is the prie of the futures ontrtD K is strike prie of the option nd the voltilityF gonsidering nlogous tles for put options in les PFS E PFV simiE lr onlusions re mdeF es expetedD the put option prie is deresE ing in moneyness nd inresing in mturityF gompring the iuropen style put option pries in le PFT to the emerin style ones in le PFU we see tht for shortEmturity options the emerinEstyle premium is iggest for slightly inEtheEmoney options while for put options with long time to mturity the premium is iggest for fr inEtheEmoney opE tionsF hen onsidering the implied voltility @sA surfe for put options in le PFV it is immedite tht lso for put options the stnE drd oneEftor qfw model is not su0ient for mthing oserved priesF por ompleteness the tles in this setion present desriptive stE tistis for options QS7 in nd out of the money s well s options with
65
more thn I yer to mturityF sn the lirtion we limit the smple to inlude options PH7 in nd out of the money nd we only inlude options with more thn I yer to mturity if they re written on either the heemer or tune ontrtsF
14 30 1,706 3,001 5,318 7,923 9,794 5,355 5,553 5,720 5,443 4,892 4,425 3,420 2,203 2,181 1,025 554 306 143 68,962
30 < 90 7,462 13,342 19,281 25,237 30,499 16,825 17,748 18,494 19,285 18,241 17,099 15,836 14,531 23,357 14,753 8,126 4,588 3,011 287,714
90 < 180 8,718 14,049 19,412 26,112 34,406 19,999 21,793 23,105 24,580 23,458 21,559 19,456 17,594 30,784 24,766 17,865 11,605 7,634 366,895
180 < 360 9,996 13,511 17,691 23,227 31,319 19,334 22,061 24,558 26,910 25,248 21,648 19,906 17,599 29,586 23,976 18,482 14,330 10,913 370,295
360 < 15,131 19,371 24,477 30,001 32,278 16,602 17,287 18,009 18,784 17,002 14,550 12,983 11,581 20,558 18,591 16,530 14,833 13,796 332,364
All 43,013 63,274 86,179 112,500 138,296 78,115 84,442 89,886 95,002 88,841 79,281 71,601 63,508 106,465 83,111 61,557 45,662 35,497 1,426,230
14 30 23.99 20.19 15.33 10.72 6.35 3.56 2.01 1.39 1.68 1.97 1.51 1.23 1.13 1.09 0.98 0.84 0.75 0.71 5.87
30 < 90 22.20 17.48 11.87 6.83 3.05 1.92 1.97 2.31 2.82 3.09 2.63 2.20 1.74 1.53 1.33 1.28 1.28 1.22 4.50
90 < 180 11.94 10.00 5.42 2.60 2.32 2.80 3.21 3.70 4.30 4.43 3.80 3.34 2.96 2.45 1.94 1.67 1.57 1.53 3.54
180 < 360 6.53 3.89 3.05 3.27 3.89 4.36 4.84 5.26 5.76 5.85 5.23 4.74 4.32 3.76 3.43 2.59 2.32 2.13 4.23
360 < 5.11 4.24 5.34 6.18 7.14 8.06 8.70 9.26 9.91 10.35 9.90 9.37 8.85 8.24 7.45 6.71 6.16 5.68 7.45
All 10.54 8.99 6.97 5.21 4.25 4.17 4.42 4.81 5.37 5.56 4.93 4.47 4.09 3.70 3.39 3.24 3.26 3.30 4.94
Table 2.2. Average European call option price for dierent maturi-
ties and strike, 1988-2011. The European price is obtained by applying the Ju formula conversion from American-style options.
66
14 30 24.73 20.31 15.58 11.44 7.95 5.72 4.47 3.37 2.56 1.99 1.51 1.23 1.13 1.09 0.98 0.84 0.75 0.71 6.82
30< 90 26.26 20.59 15.88 12.02 8.79 6.77 5.59 4.58 3.73 3.12 2.63 2.20 1.84 1.53 1.33 1.29 1.28 1.22 6.80
90< 180 26.39 21.14 16.98 13.25 10.12 8.25 7.08 6.12 5.28 4.49 3.80 3.34 2.96 2.45 1.94 1.67 1.57 1.53 7.16
180< 360 28.04 23.45 19.18 15.12 12.07 10.06 8.94 7.79 6.84 5.98 5.25 4.74 4.32 3.77 3.13 2.59 2.32 2.14 8.53
360 < 29.97 26.19 22.79 19.51 16.56 14.69 13.59 12.38 11.34 10.55 9.93 9.38 8.87 8.27 7.46 6.72 6.16 5.69 14.11
All 27.94 23.02 18.75 14.90 11.62 9.58 8.41 7.34 6.45 5.65 4.94 4.47 4.10 3.71 3.40 3.25 3.26 3.30 9.05
Table 2.3. Average American call option price for dierent matu-
Average call IV
0.65 < K/F < 0.7 0.7 < K/F < 0.75 0.75 < K/F < 0.8 0.8 < K/F < 0.85 0.85 < K/F < 0.9 0.9 < K/F < 0.925 0.925 < K/F < 0.95 0.95 < K/F < 0.975 0.975 < K/F < 1 1 < K/F < 1.025 1.025 < K/F < 1.05 1.05 < K/F < 1.075 1.075 < K/F < 1.1 1.1 < K/F < 1.15 1.15 < K/F < 1.2 1.2 < K/F < 1.25 1.25 < K/F < 1.3 1.3 < K/F < 1.35 All
14 30 0.61 0.54 0.48 0.44 0.41 0.39 0.38 0.37 0.37 0.38 0.38 0.40 0.44 0.52 0.63 0.72 0.81 0.88 0.43
30< 90 0.46 0.44 0.42 0.39 0.38 0.37 0.36 0.36 0.35 0.36 0.36 0.37 0.37 0.39 0.42 0.48 0.55 0.61 0.39
90< 180 0.40 0.39 0.37 0.36 0.35 0.34 0.34 0.33 0.33 0.33 0.33 0.34 0.35 0.35 0.36 0.38 0.41 0.43 0.35
180< 360 0.36 0.35 0.35 0.33 0.32 0.31 0.31 0.31 0.30 0.30 0.30 0.31 0.31 0.31 0.32 0.33 0.34 0.35 0.32
360 < 0.30 0.29 0.28 0.27 0.27 0.26 0.26 0.26 0.25 0.25 0.26 0.26 0.26 0.26 0.27 0.27 0.28 0.28 0.27
All 0.37 0.37 0.36 0.35 0.33 0.33 0.32 0.32 0.31 0.32 0.32 0.33 0.33 0.34 0.34 0.35 0.36 0.36 0.34
Table 2.4. Average IV for call options for dierent maturities and
strike, 1988-2011.
PFQF Inventory Data. everl inventories dt re villeD ut s rgued y qillon @IWWIA the relevnt series for derivtives priing is the ommeril inventories seriesF his series is updted on weekly sis nd ville from the FF inergy snformtion edministrtionF es mentioned erlierD it is widely doumented in the literture tht for ommodities in generl nd oil in prtiulr inventories in)uene pries
67
14 30 7 42 185 592 1812 1962 3305 4386 4870 5259 5228 4630 4094 6348 4601 3141 2003 1341 53806
30< 90 899 2644 5857 12666 23074 14754 16026 17160 18630 17629 15749 13493 11589 18097 12752 9083 6372 4705 221179
90< 180 5978 11449 21650 33082 39711 21339 22567 23873 24699 21031 17133 13990 11216 16622 11538 8155 5707 4612 314351
180< 360 15676 24957 33033 38063 42456 23412 25358 26459 27007 21881 15837 12688 9577 14347 10145 7286 6038 5487 359707
360 < 18231 21394 25552 30658 32813 16912 17442 17632 17966 15461 12186 10405 8691 14569 12698 11534 11022 10517 305683
All 40791 60486 86277 115060 139866 78379 84698 89510 93172 81261 66133 55206 45167 69983 51734 39199 31142 26662 1254726
14 30 0.56 0.60 0.69 0.81 0.99 1.07 1.18 1.47 1.98 1.85 1.34 1.56 2.62 4.58 7.10 10.05 13.29 16.96 3.75
30< 90 0.72 0.80 0.98 1.12 1.37 1.72 2.11 2.57 3.09 2.99 2.40 2.02 1.84 2.31 4.03 6.44 8.85 10.82 2.77
90< 180 1.05 1.17 1.29 1.55 2.17 2.82 3.29 3.82 4.50 4.55 3.96 3.60 3.36 3.15 2.94 3.36 4.18 5.41 2.98
180< 360 1.43 1.56 1.87 2.39 3.21 3.91 4.46 5.04 5.83 6.07 5.82 5.64 5.45 5.19 4.81 4.87 4.50 4.45 3.90
360 < 3.07 3.99 5.06 6.06 7.08 7.99 8.75 9.56 10.38 10.45 10.19 9.77 9.63 9.25 8.60 7.76 7.08 6.49 7.47
All 2.09 2.31 2.61 2.98 3.49 4.01 4.46 4.96 5.60 5.57 4.97 4.67 4.55 4.75 5.33 6.18 6.81 7.17 4.33
Table 2.6. Average European put option price for dierent maturi-
ties and strike, 1988-2011. The European price is obtained by applying the Ju formula conversion from American-style options to Americanstyle options.
nd voltilities of priesF sn this pper we model the e'et y introduE ing inventories s ftor in)uening the prie of oil futuresF pigure PFP shows the time series of inventory nd verge implied voltilityF
68
14 30 0.61 0.63 0.73 0.83 1.00 1.08 1.18 1.48 1.99 2.57 3.32 4.30 5.40 7.19 9.88 12.97 16.46 19.77 5.37
30< 90 0.75 0.85 1.02 1.15 1.39 1.73 2.12 2.58 3.09 3.69 4.43 5.32 6.33 7.99 10.62 13.92 17.33 20.63 5.02
90< 180 1.09 1.23 1.35 1.60 2.21 2.86 3.32 3.84 4.52 5.26 6.05 7.05 8.15 10.01 12.73 16.07 19.68 23.15 5.08
180< 360 1.46 1.68 2.03 2.56 3.37 4.04 4.56 5.12 5.90 6.80 7.97 9.21 10.59 12.65 15.93 19.53 23.07 26.61 5.95
360 < 3.09 4.07 5.12 6.09 7.12 8.02 8.77 9.58 10.42 11.50 13.01 14.40 16.28 18.93 22.46 25.55 28.84 23.07 11.38
All 2.12 2.40 2.70 3.06 3.56 4.07 4.50 4.99 5.64 6.35 7.19 8.28 9.52 11.63 14.97 18.75 22.89 26.77 6.87
Table 2.7. Average American put option price for dierent matu-
Average put IV
0.65 < K/F < 0.7 0.7 < K/F < 0.75 0.75 < K/F < 0.8 0.8 < K/F < 0.85 0.85 < K/F < 0.9 0.9 < K/F < 0.925 0.925 < K/F < 0.95 0.95 < K/F< 0.975 0.975 < K/F < 1 1 < K/F < 1.025 1.025 < K/F < 1.05 1.05 < K/F < 1.075 1.075 < K/F < 1.1 1.1 < K/F < 1.15 1.15 < K/F < 1.2 1.2 < K/F < 1.25 1.25 < K/F < 1.3 1.3 < K/F < 1.35 All
14 30 1.14 0.99 0.88 0.71 0.56 0.47 0.42 0.39 0.38 0.37 0.36 0.37 0.38 0.40 0.44 0.48 0.53 0.60 0.42
30< 90 0.68 0.60 0.53 0.46 0.41 0.39 0.38 0.37 0.36 0.35 0.35 0.36 0.36 0.38 0.41 0.44 0.46 0.49 0.39
90< 180 0.50 0.45 0.41 0.37 0.36 0.35 0.34 0.34 0.33 0.33 0.34 0.34 0.35 0.37 0.38 0.40 0.43 0.44 0.37
180< 360 0.39 0.37 0.35 0.34 0.33 0.32 0.31 0.31 0.30 0.31 0.32 0.32 0.33 0.34 0.35 0.36 0.37 0.37 0.33
360 < 0.30 0.30 0.29 0.28 0.27 0.26 0.26 0.26 0.25 0.26 0.26 0.27 0.27 0.28 0.28 0.28 0.28 0.28 0.27
All 0.38 0.37 0.36 0.35 0.34 0.33 0.33 0.32 0.32 0.32 0.32 0.33 0.34 0.35 0.36 0.37 0.38 0.38 0.34
Table 2.8. Average IV for put options for dierent maturities and
strike, 1988-2011.
smplied voltility tends to inrese shrply fter shrp deline in inE ventoriesD while implied voltility tends to drop s inventories inreseF his e'et is lso re)eted y negtive orreltion of 0.08F hile the level of inventories ertinly hs n e'et on voltility of oil priesD there is lmost no e'et on the prie of the futuresF sn pigure PFQ we show the timeEseries of inventory nd rude oil spot prieF prom the (gure it might seem like spot pries re reting to hnges
69
x 10
0.5
1990
1995
2000 year
2005
2010
Figure 2.2. Weekly time series plots of oil inventory and average
implied volatility. Average implied volatility is the average over option contracts with dierent strike and dierent maturity.
in inventories when the hnges re very lrgeD nd otherwise the two series seem unreltedD s lso indited y orreltion oe0ient of 0.04F sn le PFW we show orreltion oe0ients etween mesures of voltility nd the inventory levelF por ny mesure of orreltion nd ny mesure of voltility orreltion is negtiveF st is ler tht emE piril dt show support for the uldorEorking reltionshipD ut it is lso ler tht the reltionship is not s strong s the one found y qemn nd xguyen @PHHSA for the soyen mrketF e re ertinly not le to write voltility s n ext inverse funtion of inventoryF7
7Note that this analysis applies mostly to the results obtained using rank
correlation statistics, i.e., Spearman's and Kendall's . Since they measure (dis)agreement on the ranking one can consider rank-inversion as having an effect of changing the correlation sign. At the same time, conventional Pearson's correlation coecient far from +/- 1 allows us to exclude the case of a perfect linear relationship where applicable.
Inventory
70
x 10
Barrels in thousands
3.5
100
$ Price
50
2.5
1985
1990
1995 year
2000
2005
2010
Figure 2.3. Weekly time series of oil inventory and spot oil price.
e lso oserve tht when inventory is sre then the erson orE reltion drops s low s EHFIWD whih indites tht when inventory is sre voltility rets more to hnges in inventoryF
Corr(Spot,Inventory) Corr(RV,Inventory) Corr(IV,Inventory) Corr(IV,Inventory if scarcity(below mean)) Pearson -0.04 -0.11 -0.08 -0.19 Spearman -0.12 -0.24 -0.22 -0.18 Kendall -0.09 -0.17 -0.14 -0.12
Table 2.9. The table shows Pearson, Spearman and Kendall corre-
lation coecients for how inventory is correlated with the futures spot price, futures realized volatility and option implied volatility.
QF
The Models
sn this setion we desrie the vrious models used in the lirtionF o void unneessry omplexityD nd without loss of generlityD we develop ll our models under the riskEneutrl mesure QF he (rst thing to notie is the generl result tht there is no drift in the futures
71
proesses under QF his is esily estlished y onsidering simple proess for the underlying spot prie
dSt = (r y ) dt + dWt , St
nd the stndrd reltionship etween the spot nd future prie
F = Se(ry) ,
where r is the instntneous riskEfree rteD is the spot voltility nd y is the onveniene yieldF prom st o9s lemm the proess for the futures ontrt is given y dFt = dWt . Ft e will only work with models for the futures with no driftD ut we will inrese omplexity of the voltility of the futuresF sn the followingD we de(ne n inventory vrile s
It =
where Invt is oserved ommeril inventories in rrelsF QFIF FV model. e onsider simple reston twoEftor @pA model for futures
dFt =
Vt Ft dWtF , Vt dWtV ,
V0 = v0 @lirted prmeterA, = F V t,
where aV determines the speed of men reversion to bV in stohsE ti voltility nd determines skewness in the distriution of futures priesF sf = 1 the model ollpses to the stndrd reston modelF ith = 1 the model lso enompsses the prtilly very relevnt stti ef @4stohsti lphD etD rho4A modelF8 sn this pper we only onsider the simple se of = 1F his is done for oth simpliity nd omputtionl eseD howeverD the generliztion should theoretiE lly e strightforwrd for of HD 1 /2 D nd IF emong those vlues the skewness of I seems to e the empirilly relevnt hoieD see qemn @PHHSAF
eectively only piecewise-constant (not constant) and therefore time-varying. This is reected in the (re)calibration methods used in the paper. In order to keep the notation light, we do not introduce this additional temporal dependence in most of the formulas other than the ones used when discussing the calibration.
8The model is a dynamic SABR model in the most general case since F V is
72
QFPF FI model. e simple modi(tion of the reston twoEftor model gives us model utilizing the informtion in inventoryF e still model the prie of the future s in the reston modelD ut insted of ugE menting the eqution y n eqution for voltility we ugment it y n eqution for inventoryF he model is formulted s9
dFt =
It Ft dWtF , It dWtI ,
I0 = i0 @oserved dtA, = F I t,
where the interprettion of the oe0ients now pply to inventory rther thn voltilityD ut is otherwise the smeF sn prtie this model is di'erent from the reston p model in only one spetF ther thn lirting initil voltility we tke initil inventory from dt @heneD resulting in more prsimonious modelAF his llows for muh fster lirtion ut lso puts severe restrition on the modelF e note tht euse of the restrition the model will y onstrution not e le to mth dt s well s the stndrd p model @whih hs more degrees of freedomAF roweverD s will e shown in etion SD the liE rted model prmeters re less voltile whih is desirle for hedgingF he ps model only performs slightly worse thn the p model in outE ofEsmple priing nd in some settings it is le to gretly outperform the p model in terms of foresting option pries one dy hedF sing the stohsti di'erentil eqution @hiA pproh together with lirtion o'ers lot of )exiility in terms of modelingF sn this pper we only onsider twoEftor modelsD ut it is not theoretiE lly muh more omplited to work with threeEftor modelsF xtE url models to pply from the lss of threeEftor models inlude the douleEvoltility model of ghristo'ersen et lF @PHHWAD model ugE menting the hi for futures pries with oth voltility nd inventory s proposed in eppendix fD p nd ps modelsD respetivelyD or model ugmenting the hi for futures pries with voltility nd ltent ftor only in)uening futures pries nd voltility through the orreltion struture @s proposed in eppendix fD pv modelAF emiElosedEform solutions n still e otined for suh models @see eppendix fA nd priing is lmost s fst s in twoEftor modelsF
I cI It dWt , where I is a functional transformation of I . For instance, one could = 1 . Here, we focus on modeling the inverse-inventories (scarcity) levels where I I have chosen to focus on modeling the inverse relationship via the (non-zero) instantaneous correlations of the Wiener Processes driving the price F and inventory I . However, we have also performed an assessment of the FI model where the inventory data was replaced with inverse-inventory data and found no improvement in model performance (and even observed a deterioration).
73
ht omplites the use of threeEftor models for empiril studies is the omputtionl omplexity when lirting the modelsF QFPFIF Specication Analysis. en empirilly interesting reserh quesE tion is whether the proposed model (ts the dt for the inventory levels well ! in prtiulrD whether the distriutionl ssumptions mde re resonleF sn order to investigte itD we (rst note tht the ontinuousE time hi @QFTA n e disretized s follows It+t It = aI (bI It ) t + cI I t t D whih is equivlent to It+t It aI b I t I t t + cI t D where t is the error termF = a I rt It his eqution n e used for liner regressionF e run the regresE sion nd perform dignosti nlysis of the residuls ! in prtiulrD we illustrte the E plot for the smple of dt versus theoretil distriution in pigure QFIF e onlude tht the model (ts the dt remrkly wellD leding to residuls lmost ompletely within WS7E on(dene level pointEwise on(dene envelopeF
74
QQ Plot
3
q qqq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q
Studentized Residuals(fit)
0 t Quantiles
E plot for studentized residuls from the smple of dt (tted with liner model ginst theoretE il quntiles of omprison tEdistriutionD long with WS7Eon(dene level pointEwise on(dene envelopeF
Figure 3.1.
RF
Calibration methods
wny suggestions hve een mde in the literture s to how the reston nd restonElike models n e lirtedF he (rst importnt distintion is etween wonteEgrlo nd losedGsemiElosed form soluE tions using pourier methodsF wonteEgrlo methods re in generl esE ier to implement nd muh more )exileD ut lso very slow ompred to losedEform solutionsF es result most of the literture fouses on models for whih losedEform or semiElosedEform solutions existF e follow similr pproh in our prtil implementtionF he requireE ment of existene of losedEform solution often limits the )exiility of the modelsD whih is not lwys desiredF sn etion RFI we desrie how we lirte the models dilyD otining lirted set of prmeters for eh dyF sn etion RFP we desrie how we lirte the models
75
for eh mturityD dilyD yielding lirted set of prmeters for eh distint mturity on eh dyF xturllyD the ltter pproh is muh more )exile thn the former ut it lso requires very liquid dtsetD with severl strikes trded in eh mturityF purthermoreD the ltter pproh is not gurnteed to @nd will lmost surely notA yield liE rted prmeter vlues tht re onsistent ross mturities on eh dyF sn prtiulrD the lirted vlue of V0 ould e di'erent for eh mturity on eh dyF st eing sustntilly di'erent ross mturiE ties on eh dy might e sign of model misspei(tionF e similr interprettion pplies to the longErun vlueD b @bI or bV depending on the modelAF roweverD we lso note tht futures ontrts with di'erE ent mturity n hve oth di'erent spot voltility nd long run level of voltility nd s suh these prmeters re llowed to show some vrition ross mturities on the sme dyF por oth the dily nd mturityEwise lirtion method we use 80% of the oservtions for lirtion nd hek the outEofEsmple performne on the remining 20% of the oservtionsF xote thtD similrly to stndrd pproh found in numerous rtiE les sed on rossEsetionl lirtionD suh s n impliedEprmeter estimtion proedure in fkshi et lF @IWWUAD the lirtion proedure is repeted for every oservtion dte nd thus llows the prmeters to )utute through timeF hile one n lso see studies relying on sttisE til estimtion under the physil mesure P mking the ssumption of onstnt prmeter vlues throughout the entire smpling periodD our min fous is derivtives priing nd relisti performne ssessE mentD not sttistil infereneF reneD it is germne to enhmrk the models using the methodology omptile with the one used in the industry nd it is importnt to onform to the previling mrket onE ventions in order to provide fir nd prtilly relevnt ssessmentF his is lso emphsized in the sequentil nd dptive lirtion litE ertureD see vindstrm et lF @PHHVAF RFIF Daily Calibration. sn our dily lirtions we follow most of the existing literture nd ssume one prmeterEset for ll options trded on the given dyF his set n e otined in numer of wysF sn this pper we follow fkshi et lF @IWWUA nd ghristo'ersen et lF @PHHWA nd tret the spot voltility V0 s n extr prmeter to e lE irtedF sn ghristo'ersen et lF @PHHWA the spot vrine is lirted for eh dyD while the rest of the prmeters in the prmeterEset re ssumed onstnt eh yer10F e llow more )exiility in our t denote lirtion setup nd let ll prmeters vry eh dyF vet i the prmeterEset t time @dyA tY ot, the oserved iuropen option prie t tD with mturity D nd strikes i = 1, 2, ..., I F purtherD denote
Jing-Zhi and Wu (2004).
10Christoersen et al. (2009) implement this using the iterative procedure of
76
s the model iuropen option pries given t F hen the gE o i t, t gregte sum of squred priing errors optimiztion prolem is of the form t = arg min
t
1 IN
i=1 =1
oi i t, o t, t
t = 1, 2, ..., T, @RFIA
where T is the numer of dys in the dtset nd N is the option with the longest time to mturityF RFPF Daily Maturity-wise Calibration. prom the desriptive sttisE tis of implied voltility from s options it is ler tht the muelson e'et @voltility delines with timeEtoEmturityA is present in the dtF st is lso very likely tht the options re heterogeneous in other respets ross the mturity dimensionF o llow for suh heterogeneity in our models we lso lirte them using wht we ll dily mturityEwise lirtion pprohF st is similr to the dily lirtion exept tht within eh dy we further lirte set of prmeters for eh mtuE rityF he ggregte sum of squred priing errors optimiztion prolem is of the form
t, = arg min
t,
1 I
@RFPA sing this lirtion proedure we n nlyze how the lirted pE rmeters vry in the mturity dimensionF sn the flkEholes world the model performne n e ssessed y onsidering the implied voltility surfeF sn similr fshion this lirtion proedure llows us to ssess the vlidity of our model y onsidering prmeters ross mturityF por exmpleD for the reston model the lirted vlue of spot voltility V0 nd longEterm voltility b should not vry too muh ross mturityF sn the p model b should e onstnt ross mturityD sine the long run inventory should e independent of mturityF SF
Results
i=1
oi i t, o t, t,
, = 1 , 2 , FFF, N , t = 1, 2, ..., T.
sn this setion we present results ross the models nd lirtion methodsF he llEdominting gol of the lirtions is to investigte whih models hve the est outEofEsmple (tF e lso investigte how the prmeters of the models hve hnged historillyD s well s the orreltion struture etween prmeters nd (tsF sn the ps models it is possile to investigte the importne of inluding inventory dt in the nlysisF sn the mturityEwise lirtions it is further possile to investigte the di'erene in prmeters s funtion of mturity of the optionsF es suh the di'erent models nd di'erent lirtions llows for nlysis of the models in vrious dimensionsF
77
SFIF FV with Daily Calibration. e refer to the stndrd reston @IWWQA model @QFIAE@QFRA s the p modelF hen the model is liE on eh dy nd the rted dily we otin the prmeter vetor verge prmeters in eh yer re shown in le SFIF sn generlD the speed of men reversion a nd the voltility of voltility c hve delined throughout the smple periodD while the long run voltility b hs een unhngedF hese ptterns re illustrted in pigure SFIF he orreltion etween the futures pries nd voltility hs delined sustntilly from verge vlues round zero in the period from IWVW to PHHH to onsistently negtive vlues etween 67% to 22%F his is lso on(rmed in pigure SFI whih lso shows tht the orreltion oe0ient ws not lirted to very high positive vlue sine round PHHHD nd is in ft very onsistently negtive sine thenF prom le SFI the inEsmple (t is 0.0186 on verge per optionF wore interestinglyD from le SFI it is seen tht while the reston model hs historilly produed good (tsD sine PHHS the model hs produed sustntilly lrger lirtion errors of more thn 0.05 on verge per option onE trtF his indites tht the reston model previously performed very wellD while the model does not seem to e desriing the mrkets so well reentlyF his inrese in priing errors hs lso resulted in lrger outEofEsmple errors in more reent yers ompred to the erlier peE riod in the smpleF elso from pigure SFP it is seen tht the deline in (t qulity in reent yers hs resulted oth in higher errors in generl nd spikes in errors9 mgnitude for some dys in prtiulrF he lst olumn in le SFI shows tht in reent yers there hs een enough dt ville to lirte the model on lmost every dy of the yerF st lso shows tht historilly mny oservtions hve een ville eh yer with ISU oservtions in IWWV s low @ignoring PHIIAF pigure SFP shows tht there re reltively short nd onentrted peE riods with very poor (tsF he (rst nd most ler of them is relted to the ersin qulf r @P eugust IWWH E PV perury IWWIAF he oil prie strted to inrese reltively shrply sine PHHR nd this inrese ould potentilly e prt of the explntion why the model (ts hve een poor in reent yersF hen pries re inresingD ontrts furE ther inEtheEmoney will e prt of the set of options with open interestF hisD together with reltively high implied voltility through some of the reent yers @see pigure PFPAD ensures tht some of the lirted options hve reltively high nominl prieF ith high nominl pries of the options n error mesured in ents is likely to e higher even though the reltive error ould e lower for these optionsF he poor (ts otined in reent yers ould lso e explined y the turmoil in the (nnil mrketsF he top right plot in pigure SFI lerly shows tht the spot voltility in this period not only exhiits very high spikeD ut remins high for n extended periodF his is lso on(rmed y the very lrge men nd medin vlues of spot voltility in PHHV nd PHHW
78
Year 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Mean
-0.2969 0.0479 0.0564 -0.0756 -0.2944 -0.0854 0.0963 -0.0837 0.0931 0.0716 -0.0793 -0.1344 -0.4575 -0.6159 -0.6723 -0.2224 -0.2828 -0.5210 -0.5325 -0.2286 -0.6611 -0.6052 -0.4941 -0.2677
0.0647 0.2845 0.1295 0.0604 0.1943 0.1589 0.0694 0.2248 0.3429 0.2800 0.3087 0.1896 0.2108 0.2222 0.2429 0.1958 0.1478 0.1326 0.1039 0.3318 0.3579 0.1281 0.1114 0.1935
In-sample t 0.0004 0.0271 0.0025 0.0008 0.0022 0.0019 0.0019 0.0059 0.0151 0.0140 0.0120 0.0034 0.0088 0.0127 0.0085 0.0135 0.0109 0.0223 0.0260 0.0710 0.0552 0.0615 0.0710 0.0186
Out-of-sample t 0.0005 0.0283 0.0031 0.0010 0.0024 0.0019 0.0024 0.0057 0.0197 0.0132 0.0126 0.0033 0.0094 0.0120 0.0084 0.0145 0.0109 0.0228 0.0271 0.0721 0.0552 0.0623 0.0708 0.0191
Obs. 207 213 228 247 231 234 240 209 187 157 181 242 223 232 232 245 251 250 249 242 250 252 85 221
Table 5.1. Mean of the parameter values and ts for each year
using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-ofsample t is the out-of-sample equivalent. The results are for the daily calibrated FV model.
s reported in les SFI nd SFPF sn ft PHHW hs y fr the lrgest medin spot voltility of 28% whih is very high for oil futuresF prom le SFPD whih reports medins s opposed to the men vlues onsidered oveD severl importnt points n e mdeF he speed of menEreversion of voltility aD longErun voltility bD nd voltility of voltility c ll hve outliers nd the reported medins re sustntilly lower thn the men vluesF he orreltion etween futures pries nd voltility on the other hnd hs lrgely similr medin vlues s the mensF wedins of (ts of the model re sustntilly lower thn their mens reported ove in le SFIF his indites tht the reston model is in generl etter thn wht ws suggested y the ove men vluesF roweverD in reent yers the medin vlues hve een just s poor s the men vluesD supporting the notion tht the reston model hs een reltively worse in reent yersF le SFQ shows the stndrd devition of the prmeter estimtes nd (tsF hile the model (t ws reltively poor in reent yers in terms of men nd medin inEsmple nd outEofEsmple (ts stndrd devitions of (ts re only lrge in PHHV nd PHHWD when the risis ws t its pekF sn PHIH nd PHII the stndrd devition is muh in line with wht n e onsidered normlF he sme pttern is seen in the stndrd devition of spot voltility V0 F xot surprisinglyD the stndrd
79
800 600 400 200 0 1990 1995 2000 2005 2010 Year 10 8 6
c a b (blue), v0 (black)
1.5
0.5
0 1990 1995 2000 2005 2010 1 0.5 0 0.5 1 1990 1995 2000 2005 2010
model, 1988-2011.
devition of model (t nd spot voltility is lso lrge during the qulf r in IWWHF hile the ig risis periods n e seen from results of the reston model it is unfortunte in the sense of @worseA performne s this implies tht the reston model is not sophistited enough to del with situtions of gret turmoilF st is interesting to investigte the reltionship etween the lirted prmeters nd model (t in more detil nd we do this y onsidering their orreltions in le SFRF e report erson @permnA orE reltions in the lower @upperA tringulr prtF here is strong posiE tive reltionship oth etween the menEreversion prmeter a nd the voltility of voltility prmeter c nd etween menEreversion nd spot voltility V0 F es is seen for exmple in le SFI the longErun voltility b is generlly lowD so with high spot voltility the results imply tht reversion to the lower voltility hppens fster thn if spot voltilE ity hd een low @the drift of the proess depends on the distne etween b nd its urrent vlueAF feuse menEreversion hppens fst it mkes voltility more voltile implying the positive reltionship etween menEreversion nd voltility of voltilityF he erson orE reltion etween menEreversion nd orreltion is only 0.11 while the permn orreltion is 0.41F his implies tht when menEreversion is
80
FV model, 1988-2011. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent.
lrge then futures nd voltility re more strongly orreltedD ut more so in the nonEliner permn orreltion thn in the liner erson orreltion senseF st is further interesting tht the longErun vrine prmeter b does not seem to e orrelted with the other prmeters nd (t mesures in the liner erson senseD ut is reltively strongly orrelted with ll the remining prmeters nd oth (t mesures in the nonEliner @rnkE ingA permn senseF xot surprisinglyD the orreltion etween the inE nd outEofEsmple (t is lmost unity implying tht when the model hs poor inEsmple (t it lso hs poor outEofEsmple (tF snE nd outE ofEsmple (t re strongly orrelted with the prmeters of the model in terms of permn orreltionF sf voltility is highD s indited y spot voltility ndGor high longErun vrineD then the model performs worse in explining dtF his ould imply tht the reston model is not the orret model for underlying voltility nd tht one should look for riher modelF roweverD it ould lso imply tht the lirtion proedure is too strit nd indeed we show in etion SFQ tht when we
81
Obs. 207 213 228 247 231 234 240 209 187 157 181 242 223 232 232 245 251 250 249 242 250 252 85 232
-0.2880 -0.0690 0.0646 -0.0569 -0.2477 -0.1876 0.0707 -0.1530 0.1786 0.1296 -0.1003 -0.1456 -0.4111 -0.6297 -0.6524 -0.1892 -0.1709 -0.4832 -0.5524 -0.1808 -0.6621 -0.6219 -0.5298 -0.2650
0.0602 0.1299 0.0536 0.0419 0.0594 0.1062 0.0520 0.1371 0.1114 0.2199 0.1869 0.1708 0.1484 0.1837 0.1503 0.1656 0.1428 0.0861 0.0949 0.2183 0.2810 0.1262 0.1042 0.1293
In-sample t 0.0002 0.0007 0.0007 0.0003 0.0006 0.0007 0.0006 0.0016 0.0109 0.0029 0.0041 0.0018 0.0028 0.0035 0.0048 0.0073 0.0071 0.0139 0.0209 0.0537 0.0424 0.0577 0.0707 0.0042
Out-of-sample t 0.0002 0.0009 0.0007 0.0005 0.0006 0.0008 0.0007 0.0015 0.0088 0.0026 0.0039 0.0017 0.0029 0.0033 0.0049 0.0075 0.0076 0.0145 0.0211 0.0553 0.0430 0.0574 0.0660 0.0039
Table 5.2. Medians of the parameter values and ts for the daily
calibrated FV model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent. The results are for the FV model calibrated daily.
pply the mturityEwise pproh to lirtion we otin muh etter (ts nd the model explins the dt well oth in high nd low voltility regimesF le SFR further shows tht if menEreversionD voltility of voltility ndGor orreltion etween futures pries nd voltility re high then the model explins dt etterF SFPF FI with Daily Calibration. por our slight modi(tion of the p @or restonD IWWQA modelD the ps modelD we present medin pE rmeter vlues nd model (ts in le SFSF snsted of lirting the spot voltility we tke I0 s eing the inventory oserved from dtF hile this gives us one prmeter less to lirteD this simpli(tion omes t ost in terms of model (tF gompring model (t in le SFS to the (t in le SFI the ps model on verge produes more thn 4 times lrger errors nd this is prtiulrly pronouned in the more reent dtF gonsidering I0 two likely explntions for this n e menE tionedF pirstD I0 in the ps model is remrkly onstnt ompred to V0 in the p model implying tht the more voltile V0 provides etter desription of dtF eondD ompring the level of the two seriesD V0 in the p model is in generl smller thn I0 in the ps model nd sine the long run level of voltility in p nd inventory in the ps model re more or less equlD the ps model requires stronger men reversion
82
Year 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Mean
0.1736 0.3700 0.2694 0.1245 0.2522 0.3704 0.2014 0.3580 0.3864 0.4465 0.2971 0.1745 0.2156 0.2407 0.2105 0.3481 0.2589 0.1526 0.2413 0.1489 0.1130 0.0648 0.1266 0.2371
0.0258 0.3076 0.2071 0.1084 0.2570 0.1893 0.1028 0.2990 0.4073 0.2437 0.3672 0.0722 0.1604 0.1496 0.2117 0.1872 0.0238 0.2108 0.0281 0.3067 0.1975 0.0200 0.0200 0.1760
In-sample t 0.0007 0.0877 0.0098 0.0027 0.0047 0.0038 0.0122 0.0124 0.0211 0.0396 0.0178 0.0070 0.0291 0.0308 0.0105 0.0210 0.0128 0.0268 0.0240 0.1009 0.0885 0.0258 0.0319 0.0270
Out-of-sample t 0.0012 0.0875 0.0151 0.0032 0.0049 0.0043 0.0150 0.0130 0.0336 0.0312 0.0207 0.0071 0.0337 0.0313 0.0104 0.0270 0.0151 0.0276 0.0298 0.1014 0.0848 0.0288 0.0333 0.0287
Obs. 207 213 228 247 231 234 240 209 187 157 181 242 223 232 232 245 251 250 249 242 250 252 85 221
Table 5.3. Standard deviations of the parameter values and ts for
the daily calibrated FV model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent. The last row shows the mean in the full sample period. The results are for the FV model calibrated daily.
a b c
V0
model as well as in- and out-of-sample ts are in the lower triangular part and Spearman correlations are in the upper triangular part. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent. The results are for the FV model calibrated daily.
prmeterD whih in turn lso implies lrger voltility of voltility prmeterF his indites tht the restrition put on I0 is severe nd might not e desirleF snterestinglyD oth models produe lrgely simE ilr orreltion prmetersF elsoD the optimiztion lgorithm involved with the lirtion of the ps model onverges to n interior solution more often thn the p modelD whih is prtiulrly pronouned in IWWTEIWWWF
83
Obs. 170 201 227 246 223 250 240 242 236 196 243 247 224 238 231 244 251 250 246 198 229 250 83 225
-0.6922 -0.0937 0.1026 -0.2497 -0.5951 -0.0562 0.1763 0.0598 0.1131 0.1549 0.0073 -0.1316 -0.2895 -0.4439 -0.6559 -0.2296 -0.3830 -0.5059 -0.4821 -0.2495 -0.5641 -0.7086 -0.4870 -0.2621
0.3382 0.3664 0.3450 0.3366 0.3409 0.3344 0.3189 0.3045 0.3098 0.3334 0.3223 0.2920 0.3046 0.3066 0.2810 0.2886 0.3153 0.3347 0.3284 0.3043 0.3484 0.3532 0.3495 0.3230
In-sample t 0.0009 0.0505 0.0061 0.0012 0.0018 0.0037 0.0040 0.0083 0.0170 0.0115 0.0139 0.0095 0.0211 0.0212 0.0233 0.0325 0.1233 0.1102 0.1727 0.3365 0.2347 0.4371 0.5147 0.0822
Out-of-sample t 0.0008 0.0524 0.0059 0.0014 0.0024 0.0042 0.0053 0.0075 0.0208 0.0117 0.0159 0.0095 0.0219 0.0220 0.0242 0.0337 0.1231 0.1113 0.1734 0.3355 0.2347 0.4415 0.5083 0.0830
Table 5.5. Mean of the parameter values and ts for the daily
calibrated FI model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent.
sn pigure SFQ we show time series of the prmeters in the modelF prom the top left plot it is immedite tht s for the p model the menEreversion prmeter is muh lrger nd muh more voltile in the erly prt of the smple period nd muh more stle on lower level in more reent dtF sn the ps model this is even more pronounedF sn the topEright setion of the plot we n see tht inventory I0 is on muh higher level thn the prmeter for longErun inventory bF his suggests tht the interprettion of the ltent eqution in the reston @IWWQA model s eing inventory rther thn voltility might not e trueF yn the other hnd it is remrkle how stle the prmeter estimte of long run inventory is ompred to inventoryF his seems muh more plusile thn the errti ehvior thereof oserved in the p modelF his stility is lso desirle for foresting nd hedging s will e pprent lterF gompred to the p model the voltility of inventory is muh more voltile erly in the smpleD while it ehves muh more similrly in the reent dtF es noted ove orreltion etween futures pries nd inventory seems to ehve similrly in the two modelsF imilrly to the p modelD the ps model produes reltively lrge priing errors during the qulf r nd in prtiulr during the (nnil risis @see pigure SFRAF purtherD sine the (nnil risis the ps model
84
0.8 0.6 0.4 0.2 0 1990 1995 2000 2005 2010 1 0.5 0 0.5 1 1990 1995 2000 2005 2010
400 200 0 1990 1995 2000 2005 2010 Year 20 15 10 5 0 1990 1995 2000 2005 2010
c
Figure 5.3. Time series of the calibrated parameters (a, b, c, I0 , and ) in the daily calibrated FI model.
hs performed sustntilly nd onsistently worse thn the p model in terms of oth inE nd outEofEsmple (tF roweverD the misEpriing in the ps model seems to go up shrply lredy in PHHS @in ontrst with the p modelD pigure SFPA nd therefore the worse performne of the ps model does not seem to e relted solely to the (nnil risisF therD it ws round PHHS tht the oil pries strted to inrese shrplyD nd mye this @possily with ssoited rise in voltilityA presents possile explntionD sine the inventory dt does not seem to e the driver of the inreseF hen onsidering the medins in le SFT rther thn mens it is immedite tht the medin of the menEreversion prmeter a nd voltility of voltility c in the full smple re muh lower thn the menF hen onsidering medin (ts it is worth noting tht the ps model 4only4 produes out three times lrger errors thn the p modelD while mens were out RES times higherF egin the reson for this lies in the ft tht the ps model exhiits sustntil misEpriing only in reent yers nd this misEpriing hs lrger impt on men (ts s ompred to medin (tsF
85
is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent.
he stndrd devitions reported in le SFU on(rm wht ws lso noted from pigure SFQF he menEreversion prmeter nd voltility of voltility prmeter re muh less stle in the ps thn in the p modelF gontrry to thisD the longErun level of inventory is muh more stle thn the longErun level of vrineF st is tempting to onlude tht this implies tht the p model is le to identify men reversion nd voltility of voltility with more ertinty thn the ps model while the reverse holds for the longErun vlue of either voltility or inventoryF roweverD sine the oserved inventory level is lmost onstnt nd the long run vlue of inventory is muh lower thn this vlueD interpreting b s the longErun vlue of inventory does not seem orretF gonsidering the orreltion mtrix in le SFV we fous on the difE ferenes etween results for the ps nd the p modelF he reltionship etween voltility of inventory @voltilityA nd orreltion etween fuE tures pries nd inventory @voltilityA in the ps @pA model hs hnged sign s ompred to the p modelF he most signi(nt di'erene eE tween the orreltion mtrix in the ps nd tht in the p model is tht
86
Year 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Median
-0.7132 -0.2183 0.1418 -0.1956 -0.7451 -0.0922 0.1723 0.0224 0.1238 0.1899 -0.0177 -0.1393 -0.3201 -0.4658 -0.6349 -0.1534 -0.3871 -0.4918 -0.4957 -0.2187 -0.6414 -0.7073 -0.4083 -0.3030
0.3393 0.3704 0.3451 0.3336 0.3411 0.3348 0.3192 0.3050 0.3104 0.3366 0.3301 0.2905 0.3074 0.3150 0.2804 0.2924 0.3191 0.3352 0.3284 0.3026 0.3473 0.3577 0.3489 0.3256
In-sample t 0.0006 0.0011 0.0012 0.0006 0.0016 0.0027 0.0038 0.0047 0.0142 0.0046 0.0135 0.0088 0.0166 0.0183 0.0191 0.0230 0.1191 0.0997 0.1565 0.2832 0.1563 0.4401 0.4934 0.0137
Out-of-sample t 0.0005 0.0013 0.0013 0.0007 0.0015 0.0032 0.0040 0.0038 0.0132 0.0042 0.0137 0.0083 0.0173 0.0183 0.0187 0.0247 0.1156 0.1033 0.1543 0.2732 0.1493 0.4457 0.4746 0.0137
Obs. 170 201 227 246 223 250 240 242 236 196 243 247 224 238 231 244 251 250 246 198 229 250 83 238
Table 5.6. Medians of the parameter values and ts for the daily
clibrated FI model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent.
the permn orreltion etween inventory nd (t is lmost zero in the ps model ompred to roughly 48% etween spot voltility nd (t in the p modelF hile this ould imply tht the (t is un'eted y the inventory level it is worth noting oth tht the erson orreltion is in ft higher thn in the p modelF edditionllyD the orreltion etween the longErun level of inventory nd inEsmple (t is roughly twie s high in the ps model s ompred to the orreltion etween longErun level of voltility nd inEsmple (t in the p modelF SFQF FV with Maturity-wise Calibration. he mturityEwise lE irtion requires more dt thn the dily lirtionF feuse of this we re only rrely le to lirte the model in the eginning of the dtset nd for IWWP nd IWWS for no dys t llF his results in disonE tinuous series of lirted prmeters in oth our tles nd (guresF ine IWWW plenty of dt hs een ville nd s is seen from the lst olumn of le SFW more thn PDHHH lirtions hve een performed for eh yer sine IWWW nd sine PHHS more thn QDHHH lirtions hve een performed eh yerF gonsidering the spot voltility V0 nd longErun voltility bD the qulf r nd (nnil risis re very lerly identi(edD resulting in very high prmeter vluesF his result is even more ler thn in the dily lirtion of the p modelF hile the
87
Obs. 170 201 227 246 223 250 240 242 236 196 243 247 224 238 231 244 251 250 246 198 229 250 83 225
0.2910 0.5213 0.4349 0.4039 0.3657 0.4450 0.3696 0.2965 0.3127 0.4055 0.2443 0.1629 0.2644 0.1905 0.1230 0.3409 0.1698 0.1013 0.2194 0.2208 0.3120 0.0954 0.3015 0.2828
0.0075 0.0166 0.0074 0.0097 0.0106 0.0055 0.0114 0.0056 0.0099 0.0114 0.0157 0.0073 0.0129 0.0164 0.0065 0.0120 0.0119 0.0085 0.0147 0.0106 0.0126 0.0114 0.0091 0.0107
In-sample t 0.0012 0.1252 0.0245 0.0028 0.0014 0.0026 0.0017 0.0091 0.0129 0.0486 0.0072 0.0058 0.0142 0.0168 0.0228 0.0236 0.0513 0.0468 0.0698 0.1931 0.1814 0.0666 0.1182 0.0420
Out-of-sample t 0.0011 0.1282 0.0234 0.0035 0.0024 0.0032 0.0043 0.0089 0.0227 0.0416 0.0125 0.0060 0.0154 0.0260 0.0232 0.0287 0.0527 0.0503 0.0770 0.1987 0.1811 0.0716 0.1283 0.0447
Table 5.7. Standard deviations of the parameter values and ts for
the FI model calibrated daily, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent.
a b c
model as well as the in- and out-of-sample ts in the lower triangular part and Spearman correlations in the upper triangular part. In-sample t is the minimized value of criterion function Eq. (4.1) and out-ofsample t is the out-of-sample equivalent. The results are for the FI model calibrated daily.
longErun voltility ws sustntilly elow spot voltility on verge in the dily lirtion this is not so in the mturityEwise lirtion where these vlues seem of the sme level in generlF his mkes the interprettion of the ltent ftor in the reston @IWWQA model s eE ing voltility more plusileD nd the model more redileF hile the results in the dily lirtion implied tht the menEreversion prmE eter nd the voltility of voltility prmeter deresed throughout the smple period this is not replited in the mturityEwise lirtionF
88
wenEreversion nd voltility of voltility seem either onstnt or perE hps even inresing throughout the smple periodF roweverD one hs to e somewht utious in the interprettion of menEreversion nd longErun voltility in mturityEwise lirtion sine the options re lirted for one mturity t timeD whih elimintes one dimension from the modelF sn prtie investors re out how the model (ts the dt nd in prtiulr out how it performs in terms of outEofEsmple (tF prom le SFW it is seen tht the mturityEwise lirtion improves suE stntilly over the dily lirtion method in terms of oth inE nd outEofEsmple (tF he verge inEsmple (t in the mturityEwise lE irtion is 0.0054 ompred to 0.0186 in the dily lirtionF ht isD the dily lirtion results in more thn three times higher errorsF his is prtiulrly impressive sine the ulk of oservtions in the mturityEwise lirtion is in the end of the smpleD where the reE ston model hs the worst performneF st is lso worth noting tht even though the mturityEwise lirtionD in generlD lso produes the lrgest errors in the lst yers of the smpleD this feture is muh less pronouned thn ws the se in the p model with dily liE rtionF e nturl explntion for this is tht reently more ontrts with long mturities re trdingD whih inreses the need for more )exile model nd therefore fvors the mturityEwise lirtion relE tively more in this periodF he outEofEsmple wi is lso improved from 0.0191 in the dily lirtion to 0.0071 in the mturityEwise lE irtionD n improvement of 63%F hen the medins in le SFIH re onsidered insted of mens the lignment in the level of spot voltility nd longErun voltility is minE tined nd the longErun voltility is still slightly higher thn the spot voltilityF sn medins the onvergene rte a nd voltility of voltility c re now either stle or slightly deresing through the smple periodF his is di'erent from the onlusion drwn from onsidering the mens ut more in line with wht hs een oserved in the dily lirtion of the p modelF wore importntlyD the mturityEwise lirtion is still strongly outperforming the dily lirtion oth in terms of inE smple wi nd outEofEsmple wiF his improvement mounts to roughly 60% lower mgnitude of errorsF he strong performne of the mturityEwise p model reltive to the dily p model is even more pronouned in reent yersF hus the mturityEwise p model produes verge wi from PHHU to PHII of 0.004 while the dily p model produes wi of 0.049 or redution of the wi of more thn 90%F prom le SFII it is seen tht ll prmeters of the model exept spot voltility hve higher stndrd devition in the mturityEwise lirted p model thn in the dily lirted p modelF his is not surprising sine eh mturity on eh dy is in priniple ontrt
89
Obs. 431 303 74 152 717 495 1025 2037 2573 2351 2485 2277 2709 3612 3706 3573 3726 3659 3848 1253 2050
0.3155 0.2888 -0.3849 0.0045 -0.0082 0.1476 0.1526 -0.0898 -0.1689 -0.3288 -0.4931 -0.5210 -0.1804 -0.1761 -0.2876 -0.3607 -0.1405 -0.4961 -0.4991 -0.2754 -0.2796
0.4520 0.2613 0.0877 0.0971 0.1203 0.1266 0.1538 0.1585 0.1671 0.1849 0.1578 0.1160 0.1439 0.1297 0.0850 0.1115 0.2595 0.2389 0.0850 0.0966 0.1538
In-sample t 0.0313 0.0023 0.0003 0.0003 0.0028 0.0026 0.0020 0.0021 0.0017 0.0023 0.0028 0.0027 0.0027 0.0032 0.0043 0.0067 0.0122 0.0077 0.0069 0.0118 0.0054
Out-of-sample t 0.0414 0.0030 0.0005 0.0006 0.0065 0.0083 0.0027 0.0023 0.0020 0.0041 0.0089 0.0036 0.0033 0.0040 0.0055 0.0074 0.0173 0.0089 0.0076 0.0131 0.0071
Table 5.9. The mean of the parameter values and ts for each year
using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-of-sample equivalent. The last row shows the mean in the full sample period. The results are for the FV model calibrated for each maturity on each daily.
Year 1990 1991 1993 1994 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Median a 6.2494 4.2678 5.8979 5.3089 4.3281 6.7790 4.3707 3.7365 6.8550 5.2409 3.8265 5.0761 3.9077 5.0268 2.7680 3.8751 3.3370 3.4065 3.4187 2.0433 4.1453 b 0.4404 0.2021 0.1191 0.1075 0.1183 0.0582 0.0904 0.1044 0.0884 0.1307 0.1378 0.1323 0.1322 0.1204 0.0933 0.0842 0.1744 0.2269 0.1348 0.1198 0.1230 c 2.3284 1.1704 0.4662 0.5161 0.8659 0.4139 0.4241 0.4138 0.9230 1.0065 0.8443 1.0150 0.9704 0.9041 0.7248 0.8033 1.1290 1.2402 0.9857 0.7996 0.9070 v0 0.3882 0.1257 0.0876 0.0888 0.0822 0.1025 0.1398 0.1337 0.1393 0.1346 0.1442 0.0837 0.0882 0.0981 0.0537 0.0539 0.1233 0.1479 0.0667 0.0537 0.0955 In-sample t 0.0013 0.0005 0.0002 0.0002 0.0003 0.0001 0.0001 0.0002 0.0004 0.0006 0.0004 0.0006 0.0007 0.0006 0.0011 0.0024 0.0060 0.0033 0.0034 0.0051 0.0011 Out-of-sample t 0.0016 0.0005 0.0003 0.0002 0.0003 0.0001 0.0002 0.0003 0.0004 0.0006 0.0005 0.0006 0.0007 0.0006 0.0012 0.0026 0.0062 0.0034 0.0034 0.0056 0.0011 Obs. 431 303 74 152 717 495 1025 2037 2573 2351 2485 2277 2709 3612 3706 3573 3726 3659 3848 1253 2314
0.3085 0.2294 -0.3585 0.0990 0.0998 0.0992 0.1101 -0.0624 -0.1226 -0.2782 -0.4463 -0.5134 -0.0914 -0.1163 -0.2309 -0.3255 -0.0790 -0.4539 -0.4837 -0.3622 -0.2692
Table 5.10. The median of the parameter values and ts for the
maturity-wise calibrated FV model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-ofsample equivalent.
90
in itself with its own set of prmeters whih ould eD nd indeed isD di'erent from the ontrts with other mturities on the sme dyF sn the dily lirtion one (nds prmeters for the sme type of ontrt eh dyF his ontrt is weighted verge of ll mturities on eh dyD ut is not likely to vry too muh from dy to dy nd the lirted prmeters will therefore hve lower stndrd devitionF his type of rgument n lso e pplied to the stndrd devition of wiF sn tht ontext it is somewht surprising tht while the stndrd devition of inEsmple wi is lower for the mturityEwise lirted p model thn for the dily lirted p model it is higher for outE ofEsmple wiF e elieve this result is present minly due to the wi eing muh higher outEofEsmple thn inEsmple in IWWUD PHHP nd PHHV ! see pigures SFR nd SFTF
Year 1990 1991 1993 1994 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Mean a 13.1208 28.3876 7.1529 13.7997 23.2164 18.5136 80.6161 68.1876 38.2668 40.9422 22.2400 40.0187 131.2668 75.4589 148.8299 199.4782 292.3869 20.1198 21.6243 241.6098 98.6876 b 0.3560 0.3091 0.2278 0.2497 0.2559 0.2061 0.1591 0.1688 0.1578 0.1622 0.1913 0.2596 0.2238 0.1036 0.1414 0.1469 0.2254 0.2110 0.1438 0.1960 0.1802 c 1.6384 1.8589 0.7068 0.5260 0.6983 0.4756 0.6125 0.7504 0.8085 0.8921 0.7675 0.9062 1.3877 1.1805 1.4515 1.5911 2.5837 1.0385 0.7854 2.2907 1.2400
v0
0.3229 0.3002 0.2926 0.4139 0.3833 0.2608 0.2893 0.3103 0.2872 0.2517 0.2645 0.2778 0.3619 0.2351 0.2524 0.2811 0.2418 0.1970 0.1861 0.3002 0.2622
0.3299 0.3451 0.0519 0.0611 0.1957 0.1667 0.1000 0.1443 0.1566 0.1701 0.1227 0.1319 0.2112 0.1511 0.1502 0.2063 0.3195 0.2473 0.0947 0.1762 0.1800
In-sample t 0.1125 0.0133 0.0002 0.0008 0.0177 0.0097 0.0240 0.0117 0.0067 0.0136 0.0151 0.0299 0.0185 0.0089 0.0144 0.0326 0.0184 0.0142 0.0101 0.0150 0.0172
Out-of-sample t 0.1600 0.0180 0.0005 0.0026 0.0585 0.1017 0.0279 0.0149 0.0211 0.0519 0.1396 0.0499 0.0228 0.0123 0.0161 0.0431 0.1078 0.0161 0.0101 0.0181 0.0416
Obs. 431 303 74 152 717 495 1025 2037 2573 2351 2485 2277 2709 3612 3706 3573 3726 3659 3848 1253 2050
for the maturity-wise calibrated FV model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-ofsample equivalent.
pigure SFS shows the prmeter verges ross mturities for eh dy through the smple periodF ine the mturityEwise model does not hve enough dt to e lirted often enough in the erly smple period it is not ler whether it is le to identify the impt of the qulf r on the oil pries ut the (nnil risis seems to e identi(ed lerly in terms of higher voltility of voltility nd high spot voltilityF ell prmeters of the model seem reltively voltileD whih is similr to wht ws otined in the dily lirtionF he orreltion oe0ient lso shows the sme pttern s in the dily lirtion with delines
91
etween IWWS nd PHHSD followed y shrp inrese nd gin period with delines until the end of PHIHF
1.5
0.5
0 1990 1995 2000 2005 2010 1 0.5 0 0.5 1 1990 1995 2000 2005 2010
10
c
Figure 5.5. The average of the calibrated parameters (a, b, c, v0 , ) across maturity in the maturity-wise calibrated FV model.
hen onsidering the inE nd outEofEsmple wi in pigure SFT the mturityEwise lirtion pproh generlly produes sustntilly smller verge wi ll through the smple periodF iven the lrge spikes re mostly sustntilly lower thn the spikes oserved in the dily lirtionF st is lso noted tht the lirtion methods lrgely gree on the periods in whih poor (ts re otined nd tht the (t is generlly worse in reent yersF pinllyD s in the dily lirtion there is reltionship etween inE nd outEofEsmple wiD ut it is not s lerEut s in the dily lirtion nd will e investigted further elowF les SFIP E SFIR ontin the menD medin nd stndrd devitionsD respetivelyD of the lirtion results ross mturity of the option ontrtsF hese tles show severl interesting resultsF pirstD the muelson e'et is lerly oserved oth in the longErun voltility b nd the spot voltility V0 D whih oth show ler delining pttern in mturityF his is in ordne with the results for implied voltility
92
0.1
In sample fit
0.08 0.06 0.04 0.02 0 1990 0.1 1995 2000 2005 2010
FV model. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-of-sample equivalent.
oserved in les PFR nd PFVF eondD while the stndrd devition of the longErun vrine is lso delining in mturity it is onstnt ross mturity for the spot voltilityF hirdD the orreltion oe0ient is negtive for ll mturities whih is somewht surprisingF pourthD the orreltion is deresing in mturity implying tht longEmturity futures re more negtively orrelted with voltility thn the short onesF egin this is somewht surprising sine one would expet longer mturity futures to e more positively orrelted with futures thn short mturity futuresF pifthD the vrine of vrine oe0ient c is delining in mturityF his result is expeted sine for long mturity option the spot voltility of the futures ontrt is of less importne reltive to the longErun voltility of the futures ontrtF pinllyD it is seen from the (ts tht the reston model together with mturityEwise lirtion is produing impressively low inE nd outEofEsmple wiD in prtiulr for the shortEmturity optionsF xturllyD the inresing wi in mturity is expeted oth euse option pries re inresing in mturity nd they re possily posing hrder hllenge for the
93
model @eFgFD the impt of interest rtes @nd stohstiity thereofA might gin higher importne for longer mturitiesAF
Maturity 14 - 30 31 - 90 91 - 180 181 - 360 361 - 720 721 a 14.2604 14.1051 23.3084 20.1342 16.5323 16.3133 b 0.2014 0.2403 0.2237 0.1798 0.1228 0.0987 c 1.2395 1.3249 1.3721 1.2134 0.9066 0.7044
V0
Table 5.12. Means of the parameter values and ts for maturity
intervals. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-of-sample equivalent. The results are for the FV model calibrated for each maturity, daily; 1988-2011.
V0
Table 5.13. Medians of the parameter values and ts for maturity
intervals. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent. The results are for the FV model calibrated for each maturity, daily; 1988-2011.
V0
for maturity intervals. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent. The results are for the FV model calibrated for each maturity, daily; 1988-2011.
sn generlD most results re s expetedD ut it is quite surprising tht the orreltion oe0ient is persistently negtive nd even more so for long mturity ontrtsF his implies one of the following three thingsF iither the reston model is not pplile to this mrketD the lirtion method is not generl enough or the stndrd intuition in the literture is simply not orret @there might e dynmi termE struture presentD s opposed to sttiD onstntD reltionshipAF le SFIS shows the erson @lower tringleA nd permn @upE per tringleA orreltions etween lirted vluesF gompred to the
94
dily lirtion of the reston model the most remrkle di'erene is tht the orreltions etween the prmeter vlues nd oth inE nd outEofEsmple (ts re generlly muh loser to zeroF his is prtiulrly true for permn orreltionsF his ould imply tht the mturityE wise lirtion method is more pplile thn the dily lirtion sineD eFgFD we do not oserve the ounterEintuitive high negtive orreE ltion etween wi nd voltility of voltilityF yn the other hnd the very plusile positive orreltion etween wi nd oth long term voltility b nd spot voltility V0 is no longer oservedF
a b c
V0
model as well as the in- and out-of-sample ts are in the lower triangle and Spearman correlations are in the upper triangle. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-of-sample equivalent. The results are for the FV model calibrated for each maturity on each day, 1988-2011.
SFRF FI with Maturity-wise Calibration. e lso perform the mturityE wise lirtion of the ps model nd show men nd medin results in les SFIT nd SFIU respetivelyF gompred to the dily lirE tion of the ps model the mturityEwise lirtion lerly shows muh more plusile vlues of longErun inventoryD lthough they re still low reltive to wht ould e expetedF sn the dily lirtion the men nd medin longErun inventory is HFHRP nd HFHQWD respetivelyD while in the mturityEwise lirtion they re HFIHP nd HFHUQD respetivelyF eltive to oth the mturityEwise p nd the dily ps lirtion results the orreltion oe0ient is more negtive nd the voltility of the inventory oe0ient is higher thn the omprle oe0ient in the mturityEwise p model nd dily ps modelF sn terms of wi the mturityEwise ps model is performing sustntilly etter thn oth the dily lirted reston model nd the dily lirted ps modelF es expeted from the derese in the degrees of freedomD it does not perform s well s the fully )exile mturityEwise p modelD ut it does require one prmeter less to lirte nd is therefore sustnE tilly fster to lirteF he reent inrese in wi does not seem to e mitigted y using the mturityEwise ps modelD ut onsidering the stndrd devitions in le SFIU they re onsiderly lower thn in the dily ps model nd of the sme level s for the mturityEwise p modelF
95
Obs. 476 314 73 147 725 480 1012 2032 2537 2316 2331 1955 2578 3425 3021 2929 3711 3530 3673 1162 1921
0.3020 0.3575 -0.3811 -0.0034 0.0834 0.1679 0.2030 -0.1076 -0.2158 -0.4345 -0.5862 -0.6326 -0.1994 -0.2711 -0.4073 -0.4614 -0.1852 -0.6543 -0.7624 -0.4172 -0.3671
0.3584 0.3363 0.3403 0.3322 0.3034 0.3067 0.3343 0.3203 0.2916 0.3048 0.3063 0.2802 0.2895 0.3158 0.3347 0.3273 0.3063 0.3469 0.3532 0.3496 0.3195
In sample t 0.0380 0.0024 0.0004 0.0003 0.0029 0.0025 0.0014 0.0020 0.0017 0.0024 0.0030 0.0025 0.0035 0.0044 0.0075 0.0128 0.0162 0.0114 0.0132 0.0254 0.0081
Out of sample t 0.0476 0.0031 0.0006 0.0005 0.0068 0.0114 0.0020 0.0021 0.0020 0.0043 0.0095 0.0029 0.0039 0.0052 0.0085 0.0131 0.0207 0.0129 0.0145 0.0275 0.0100
Table 5.16. Means of the parameter values and ts for the
maturity-wise calibrated FI model, 1988-2011, using the number of observations shown in the last column. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent.
Year 1990 1991 1993 1994 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Median a 19.4043 40.4067 55.1889 69.1561 39.0844 43.8047 19.4456 18.2893 20.7928 21.9648 16.3658 22.9859 16.0027 18.6844 31.8064 25.9830 13.1673 15.1589 16.5357 14.7857 20.1660 b 0.4471 0.0886 0.0326 0.0427 0.0501 0.0386 0.0513 0.0724 0.0653 0.0879 0.0991 0.0836 0.0700 0.0802 0.0543 0.0504 0.0923 0.1554 0.0803 0.0577 0.0730 c 3.9256 2.1032 1.7483 2.1728 1.4209 1.0842 0.5158 0.6956 1.2940 1.5648 1.3811 1.7779 1.3300 1.4029 1.6040 1.4730 1.5082 2.0187 1.5803 1.2886 1.5123 In sample t 0.0017 0.0006 0.0004 0.0002 0.0005 0.0002 0.0002 0.0003 0.0005 0.0009 0.0006 0.0010 0.0017 0.0015 0.0033 0.0059 0.0112 0.0047 0.0081 0.0146 0.0021
0.3017 0.3594 -0.4455 0.0691 0.1144 0.0969 0.1620 -0.0696 -0.1875 -0.4217 -0.5915 -0.7088 -0.0970 -0.1886 -0.3703 -0.4176 -0.1043 -0.6490 -0.8138 -0.4089 -0.3659
0.3520 0.3361 0.3398 0.3325 0.3031 0.3056 0.3366 0.3244 0.2905 0.3075 0.3121 0.2793 0.2930 0.3191 0.3351 0.3253 0.3059 0.3445 0.3578 0.3489 0.3204
I0
Out of sample t 0.0019 0.0007 0.0003 0.0002 0.0006 0.0002 0.0002 0.0003 0.0005 0.0009 0.0006 0.0010 0.0016 0.0016 0.0033 0.0059 0.0106 0.0048 0.0081 0.0158 0.0021
Obs. 476 314 73 147 725 480 1012 2032 2537 2316 2331 1955 2578 3425 3021 2929 3711 3530 3673 1162 2174
Table 5.17. Medians of the parameter values and ts for the
maturity-wise calibrated FI model, 1988-2011, using the number of observations shown in the last column. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent.
96
Year 1990 1991 1993 1994 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Mean
0.2899 0.3240 0.4107 0.4443 0.4525 0.3353 0.3680 0.3806 0.3561 0.2903 0.2735 0.3119 0.4147 0.3061 0.3370 0.3389 0.2712 0.2328 0.2012 0.4476 0.3125
0.0154 0.0072 0.0081 0.0050 0.0056 0.0111 0.0102 0.0158 0.0071 0.0124 0.0162 0.0066 0.0114 0.0116 0.0085 0.0149 0.0106 0.0126 0.0113 0.0090 0.0113
In sample t 0.1189 0.0131 0.0003 0.0002 0.0171 0.0124 0.0146 0.0093 0.0061 0.0099 0.0127 0.0051 0.0169 0.0127 0.0234 0.0375 0.0232 0.0213 0.0214 0.0325 0.0192
Out of sample t 0.1618 0.0176 0.0007 0.0008 0.0598 0.1315 0.0161 0.0139 0.0210 0.0513 0.1452 0.0053 0.0081 0.0138 0.0217 0.0400 0.1061 0.0247 0.0230 0.0359 0.0420
Obs. 476 314 73 147 725 480 1012 2032 2537 2316 2331 1955 2578 3425 3021 2929 3711 3530 3673 1162 1921
for the maturity-wise calibrated FI model, 1988-2011, using the number of observations shown in the last column. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent.
les SFIW E SFPI ontin the mensD medins nd stndrd deviE tions of the lirted prmeters nd wis ross option mturityF he longErun inventory prmeter b is delining in mturityF his is pttern similr to the one oserved for the longErun voltility in the mturityEwiseElirted p modelF elso similrly to wht ws found in the mturityEwise lirtion of the p modelD the orreltion eE tween futures pries nd inventory is deresing in mturity nd is negE tive in oth men nd medin for ll mturity intervls onsideredF eginD this oservtion is muh more in line with intuitionD ompred to the result for the p modelD sine we expet inventory to e muh more importnt for longerEmturity options thn shorterEmturity onesF e further note tht the inE nd outEofEsmple wis re roughly 50% higher in terms of mens nd 100% higher in terms of medin thn in the mturityEwise p model for ll mturitiesF pinllyD from the lst olumn it is seen tht ompred to the mturityEwise p model the most oservtions re lost for high mturityF his implies tht the mturityEwise ps model fils to e lirted more often for long mturity ontrts thn the mturityEwise p modelF his is someE wht surprising sine we expet inventory dt to e reltively more importnt for longer mturity options s ompred to shorter mturity optionsF he stndrd devition of longErun inventory is sustntilly lower for longEmturity options thn the longErun vlue of vrine in the mturityEwise p modelF eginD this is likely used y inventory
97
I0
Table 5.19. Mean of the parameter values and ts for maturity
intervals. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent. The results are for the FI model calibrated for each maturity, daily.
I0
Table 5.20. Medians of the parameter values and ts for maturity
intervals. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent. The results are for the FI model calibrated for each maturity, daily.
I0
for maturity intervals. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent. The results are for the FI model calibrated for each maturity, daily.
pigure SFU shows the prmeter verges ross mturities for eh dy through the smple periodF es the mturityEwise p model the mturityEwise ps model lerly identi(es the (nnil risis with high spot voltilityF prom the top right plot we lerly see tht the inventory is muh more stle thn the spot voltility lirted in the mturityE wise p modelF es ws seen in the dily lirtionsD the longErun inventory level b in the mturityEwise ps model is muh more stle thn the longErun voltility vlue in the mturityEwise p modelF his stility is n dvntge for the ps modelD ut is not replited for ny of the other prmetersF es in the dily lirtion when ompring the
98
time series properties of the mturityEwise ps model with the mturityE wise p model they re lrgely similr for aD b nd F elso onsidering the timeEseries properties of the wi in pigure SFV we note tht the wi timeEseries is shifted down ompred to the dily ps model in pigure SFR nd still seems to produe high wi in similr periods s the dily ps nd the mturityEwise p modelF @pigure SFTA
1.5
0.5
0 1990 1995 2000 2005 2010 1 0.5 0 0.5 1 1990 1995 2000 2005 2010
20
c
10
rameters across maturity in the FI model calibrated for each maturity on each day.
le SFPP for the mturityEwise ps model orrespond to le SFIS for the mturityEwise p modelF he results re lrgely similr to those oserved in the mturityEwise p model ut there is one importnt di'ereneF snitil inventory is positively orrelted with wi while spot voltility is only very wekly orrelted with wiF he positive orreltion etween inventory nd wi suggests tht the model is etter for low inventory thn high inventory @iFeFD the model (ts etter in the se of srityAF his oservtion is in ordne with the theory suggesting tht inventory is more importnt for option pries when it is low thn when it is highF
99
0.1
In sample fit
0.08 0.06 0.04 0.02 0 1990 0.1 1995 2000 2005 2010
the maturity-wise calibrated FI model. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-ofsample equivalent.
a 1.0000 0.0126 0.5308 0.0010 0.0180 0.1004 0.0383 b -0.0140 1.0000 0.4936 0.0744 0.0868 0.0293 0.0245 c 0.6718 0.5645 1.0000 0.0020 0.0729 0.0281 0.0135
I0
a b c
model as well as in and out of sample ts are in the lower triangle and Spearman correlations are in the upper triangle. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent. The results are for the FI model calibrated for each maturity on each day.
TF
Model Performance
sn this setion we present the performne of the models in two disE tint dimensionsF pirstD we leve out numer of oservtions when
100
a b c
model as well as in- and out-of-sample ts are in the lower triangular part and Spearman correlations are in the upper triangular part. Insample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-of-sample equivalent. The results are for the FI model calibrated for each maturity on each day. The averages of the parameters and ts are taken across maturities and the correlations are calculated from these daily averages.
lirting the modelF e then use the lirted prmeters to lE ulte the vlue of the options we left out nd ompre the model vlue with the mrket vlueF his pproh determines whether the model n e used for priing ssets @priing performneA not urE rently trded on the mrkets ! iFeFD in mrket mkingF por instneD one ould e interested in introduing n option with strike not urrently ville on the mrketD or one might wnt to determine the vlue of n option whih hs not een trded reentlyF eondD we onsider the oneEdyEhed forest errors @forestingGhedging performneY whih n lso e interpreted s modelEroustness ssessmentAF e do this using two di'erent settingsF sn the (rst one we use the lirted pE rmeters on dy t 1 to determine the option pries on dy tF sn the seond setting we use the verge of the prmeters otined on the pst 5 @the length of the usiness weekA trding dys to determine the option vlue on dy tF sn this setion we ompre the models y lulting root men squred reltive errors @wiAF e onsider reltive errors rther thn errorsD sine the reltive mesure is not 'eted y hnge of sleF ht isD 10% priing error in fr inEtheEmoney option is treted s 10% priing error in fr outEofEtheEmoney optionF xoteD tht oth the priing performne nd the foresting perforE mne re the forms of outEofEsmple performne nlysisD designed to ssess the potentil overE(tting onerns ssoited with the dily lirtion frequenyF TFIF Pricing Performance. es desried in etion PFPFI we leve out 20% of the oservtions when lirting the modelsF e then lulte the vlue of these options using the lirted prmeters nd ompre wi etween the modelsF le TFI shows tht the mturityEwise
101
lirtion method outperforms the dily lirtion methodF sn prE tiulr the p model lirted using the mturityEwise method proE dues low wi with n verge of 0.086D ompred to 0.124 for the mturityEwise ps model nd 0.138 for the dily p modelF he mturityEwise p model is performing prtiulrly well ompred to the dily p model during the ooming yers in the lte IWWHsF st is noteworthy tht the dily ps model performs lmost s well s the dily p modelD even though it is less )exileF his mkes it very vlid lterntive to the dily p model if one vlues lirtion speedF his result might further imply tht inventory dt n e useful for option priingF
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Mean Daily FV 0.026 0.126 0.137 0.090 0.148 0.161 0.141 0.299 0.298 0.332 0.324 0.091 0.102 0.146 0.136 0.082 0.038 0.062 0.107 0.069 0.068 0.078 0.110 0.138 Daily FI 0.174 0.173 0.142 0.072 0.155 0.077 0.119 0.139 0.173 0.268 0.128 0.078 0.133 0.121 0.145 0.107 0.155 0.145 0.191 0.157 0.165 0.230 0.233 0.151 Maturity-wise FV 0.111 0.076 0.063 0.016 0.119 0.140 0.088 0.072 0.061 0.096 0.101 0.136 0.123 0.072 0.086 0.114 0.071 0.053 0.051 0.072 0.086 Maturity-wise FI 0.107 0.076 0.027 0.107 0.095 0.150 0.183 0.085 0.090 0.112 0.164 0.221 0.143 0.097 0.188 0.216 0.064 0.090 0.115 0.150 0.124
smportnt nd very interesting oservtions n e mde when one onsiders the priing performne s funtion of strike @le TFPA nd s funtion of mturity @le TFQAF gompring the lirtion methods for the p modelD inEtheEmoney options re pried eqully wellD ut oth tEtheEmoney options nd outEofEtheEmoney options re pried more preisely using the mturityEwise lirtion methodF snE terestinglyD this oservtion is not similr for the ps modelsF foth lirtion methods perform eqully well for fr inEtheEmoney nd fr outEofEtheEmoney optionsD ut for tEtheEmoney options the mturityE wise method performs etterF gonsidering the priing performne ross mturity it is seen tht the mturityEwise p model gretly outperforms the dily p model for short nd medium term optionsD
102
while for longer term options performne is equlF gonsidering the ps model the mturityEwise lirtion outperforms dily lirtion for short mturity options while it is worse for long mturity optionsF he wek performne of the dily ps model for short mturity options is expeted sine inventories ply only smll role for short mturity optionsF gonsidering the p modelD these results suggest tht the mturityE wise lirtion method should lwys e preferred for priing short mturity options nd outEofEtheEmoney optionsD while for long mtuE rity options nd fr inEtheEmoney options one n use the omputE tionlly heper dily lirtion method without muh loss of priing performneF
Moneyness K/F < 0.8 0.8 < K/F < 0.9 0.9 < K/F < 1 1 < K/F < 1.1 1.1 < K/F < 1.2 1.2 < K/F Mean Daily FV 0.083 0.081 0.106 0.122 0.132 0.141 0.111 Daily FI 0.129 0.163 0.177 0.179 0.179 0.187 0.169 Maturity-wise FV 0.080 0.082 0.085 0.075 0.080 0.105 0.084 Maturity-wise FI 0.131 0.136 0.134 0.124 0.144 0.185 0.142
Table 6.2. Averages of RMSRE for each strike interval for both
Time-to-maturity < 30 30 < < 90 90 < < 180 180 < Mean
yverllD we onlude tht the dily ps model is fres well for fr sw or longEmturity options @with the less prsimonious p models performing even etterAD while the mturityEwise p model is de(nitely preferle for fr yw or shortEmturity optionsF his might re)et the roder presene of the fundmentl mrket prtiipnts trding these instruments ! with moneyness re)eting the intrinsi ontrt vlue nd longEmturity ontrts importnt forD eFgFD fuel hedging y the irlinesF TFPF Forecasting Performance. sn prtie the hoie etween modE els often ultimtely depends on their foresting performneF sn this setion we onsider oneEdyEhed foresting performne using oth the prmeterEset on the previous dy nd the verge of the prmeterE sets over the most reent (ve trding dysF sn le TFR we present the
103
verge wi in eh yer for oth models nd lirtion methE ods using the prmeterEset otined on the previous dyF pirstD it is noteworthy tht the wis re higher thn the ones in le TFIF his is expeted sine prmeters re timeEvryingD ut it my e surE prising tht the di'erene is not lrger thn it isF his re)ets the ft tht prmeters re generlly not moving muh from dy to dyD demonstrting the roustness of the models nd lirtion methodsF he ordering of models nd lirtion methods is similr to the orE dering in the previous setionF egin the mturityEwise p gretly outperforms the dily p nd gin this phenomenon is prtiulrly pronouned in the lte IWWHsF
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Mean Daily FV 0.036 0.137 0.149 0.091 0.129 0.158 0.140 0.298 0.289 0.332 0.322 0.099 0.105 0.146 0.139 0.088 0.046 0.064 0.112 0.079 0.078 0.084 0.115 0.141 Daily FI 0.185 0.182 0.155 0.078 0.151 0.077 0.115 0.142 0.172 0.267 0.131 0.083 0.136 0.122 0.147 0.110 0.156 0.144 0.195 0.160 0.168 0.234 0.235 0.154 Maturity-wise FV 0.132 0.114 0.072 0.034 0.125 0.128 0.093 0.087 0.073 0.105 0.107 0.141 0.128 0.078 0.088 0.122 0.083 0.069 0.061 0.091 0.096 Maturity-wise FI 0.129 0.125 0.037 0.130 0.096 0.154 0.179 0.104 0.101 0.119 0.168 0.226 0.144 0.104 0.190 0.208 0.076 0.100 0.117 0.148 0.133
Averages of RMSRE for each calibration method and each model, 19892011. Forecasts of option prices are calculated using the parameter-set from the previous day.
Table 6.4.
le TFS shows the wi ross strike nd le TFT ross mE turityF gonlusions re similr to those in the previous setionF he mturityEwise lirtion method should e used for priing short mE turity options nd outEofEtheEmoney optionsD while long mturity opE tions nd fr inEtheEmoney options n e forest using the dily lE irtion methodF xoteD tht the mturityEwise ps model outperforms the mturityEwise p model for the intermedite timeEtoEmturity opE tionsD iFeFD when 30 < < 180F le TFU shows the wi when option pries re forest using the verge of the prmeterEsets otined over the most reent S dysF sn generlD the wi re higher thn when using the prmeterEset
104
Averages of RMSRE for each strike interval for both calibration methods and each model. Forecasts of option prices are calculated using the parameter-set from the previous day.
Table 6.5.
Mean
Averages of RMSRE for each time to maturity interval for both calibration methods and each model. Forecasts of option prices are calculated using the parameter-set from the previous day.
Table 6.6.
from the pst dyF his suggests tht while the model prmeters do not hnge muh from dy to dyD they still hnge su0iently muh tht using informtion from s fr s (ve dys k is not relevntF his shows the importne of using the most reent informtion @nd underE sores the desirility of the dily lirtion frequeny in prtieAF he performne loss is prtiulrly pronouned for the mturityEwise lirtion methods nd proly re)ets the ft tht in the mturityE wise lirtion method prmeters re less stle thn they re in the dily lirtionD sine they re lirted to ontrts with spei( mturityF he dilyElirted ps model is now performing muh etE ter thn ny other modelF his re)ets the ft tht prmeters re more stle in the ps model s ompred to the p modelF his stE ility is desired feture for hedgingD sine the osts of relning re redued with less frequent relningF les TFV E TFW show wi ross the strike nd mturity diE mensionsF por the models lirted using the dily method the wE is re reltively stle ross the strike dimensionD while for the mturityEwise lirtion they re remrkly lower for tEtheEmoney options s ompred to inE nd outEofEthe money optionsF etD for ny model nd ny lirtion method the wi re s low s those oserved when using the prmeterEset from the previous dyF his is lso the se when onsidering the wi ross mturity for ll modelsF sn generlD these tles suggest tht rther thn using the verge of prmeterEsets over the most reent period one should use those lirted on the previous dyF
105
method and each model, 19892011. Forecasts of option prices are calculated using the average of the parameter set on the past 5 days.
K/F < 0.8 0.8 < K/F < 0.9 0.9 < K/F < 1 1 < K/F < 1.1 1.1 < K/F < 1.2 1.2 < K/F < 1.3 1.3 < K/F
Table 6.8. Averages of RMSRE for each strike interval for both
calibration methods and each model. Forecasts of option prices are calculated using the average of the parameter set on the past 5 days.
Mean
Averages of RMSRE for each time to maturity interval for both calibration methods and each model. Forecasts of option prices are calculated using the average of the parameter set on the past 5 days.
Table 6.9.
yverllD we onlude tht when the forests of option pries re lulted using the prmeterEset from the previous dy the mturityE wise models re preferred to dily models nd p models re in generl
106
preferred to ps modelsF roweverD we note good performne of the dily ps model for longEmturity optionsF hen the forests of option pries re lulted using the verge of the prmeterEset on the pst S dysD the dily models re preferred to mturityEwise modelsF yn vergeD the dily ps model performs the est hereD whih might e relted to fundmentl ftors mentioned in etion TFIF sn generlD foresting the option pries using the prmeterEset from the previous dy with the mturityEwise p model results in the est performneF UF
The Black-Scholes Model Benchmark
sn the previous setions we hve shown tht inventory dt n e useful for lirtion in terms of (tD foresting nd priingF sn order to investigte this furtherD we lso inlude the results for the flkE holesEwerton @fwA model @orD the in ontext of our underlyingD flk @IWUTA modelA s enhmrkF UFIF BSM with Maturity-wise Calibration. he men nd medin stndrd devitions re ehving s expeted ! they re round the expeted levels nd they lerly show oth the dotEom risis round PHHIEPHHP nd the reent redit risis round PHHVEPHHWF he inEsmple nd outEofEsmple (t errors re signi(ntly higher @out n order of mgnitudeA thn for ny of our other modelsD whih is of ourse expetedD due to the extr )exiility in the other models ompred to fwF his demonstrtes the reltive importne of stohsti voltility @or inventoryA over the onstnt voltility ssumptionF st is remrkle tht the inEsmple (t hs een very poor sine the (nnil risisF emong prtitionersD it hs een oserved tht mny of the stndrd tools hve not worked well in the reent yers ! perhps this is just on(rmtion of tht oservtionF yutEofEsmple (t is lso muh worse in reent yers whih is very lerly seen in the medin resultsF here is lrge di'erene etween men nd medin results for inE nd outEofEsmple (tsF his indites tht for lot of the oservtions the model is performing very poorlyF UFPF BSM with Daily Calibration. eginD the vlues re round the expeted levelF roweverD in the dily lirtion the vlues do not seem to pture the wellEknown inrese in the voltility in PHHVE PHHW nd only wekly pture the voltility inrese in the PHHIEPHHP risisF roweverD the PHHVEPHHW risis is lerly 'eting the wiF gompring results with those for the mturityEwise lirtion the wis re sustntilly higherD whih ws lso the se for ll the previously disussed modelsF xturllyD this is still prtilly result of the dded )exiility of the mturityEwise methodD ut it on(rms nd highlights tht using the mturityEwise lirtion one n otin
107
Year
1990 1991 1993 1994 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
0.6049 0.3545 0.2397 0.2436 0.2724 0.2206 0.2633 0.2884 0.3042 0.3614 0.3540 0.3137 0.3013 0.3117 0.2594 0.2453 0.3394 0.3982 0.2990 0.2803 0.3056
111.0000 138.0000 52.0000 126.0000 442.0000 366.0000 843.0000 1553.0000 1504.0000 1275.0000 1591.0000 1219.0000 1510.0000 2246.0000 2549.0000 2247.0000 1779.0000 1558.0000 2139.0000 847.0000 1204.7500
Table 7.1. The mean of the parameter values and ts for each year
using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.2) and out-of-sample t is the out-of-sample equivalent. The last row shows the mean in the full sample period. The results are for the BSM model calibrated for each maturity on each day.
Year
1990 1991 1993 1994 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
0.1633 0.1631 0.1080 0.0959 0.0869 0.1160 0.1502 0.1197 0.0843 0.1088 0.0685 0.0552 0.0616 0.0541 0.0274 0.0361 0.1293 0.1236 0.0339 0.0351 0.0740
111.0000 138.0000 52.0000 126.0000 442.0000 366.0000 843.0000 1553.0000 1504.0000 1275.0000 1591.0000 1219.0000 1510.0000 2246.0000 2549.0000 2247.0000 1779.0000 1558.0000 2139.0000 847.0000 1204.7500
Table 7.2. Standard deviations of the parameter values and ts for
the maturity-wise calibrated BSM model, 1988-2011, using the number of observations shown in the last column. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent.
108
Year
1990 1991 1993 1994 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
0.6189 0.3181 0.2626 0.2707 0.2870 0.2722 0.3055 0.3155 0.3076 0.3403 0.3524 0.3075 0.2972 0.3223 0.2641 0.2506 0.3009 0.3780 0.3047 0.2888 0.2981
111.0000 138.0000 52.0000 126.0000 442.0000 366.0000 843.0000 1553.0000 1504.0000 1275.0000 1591.0000 1219.0000 1510.0000 2246.0000 2549.0000 2247.0000 1779.0000 1558.0000 2139.0000 847.0000 1389.5000
Table 7.3. Medians of the parameter values and ts for the
maturity-wise calibrated BSM model, 1988-2011, using the number of observations shown in the last column. In sample t is the minimized value of criterion function Eq. (4.2) and out of sample t is the out of sample equivalent.
muh lower wi thn using the dily lirtionF st is importnt tht oth going from dily to mturityEwise lirtion nd from fw to more dvned models gretly redues wiF UFQF BSM Performance. sn this setion we onsider oneEdyEhed foresting performne using oth the prmeterEset on the previous dy @4ltest4A nd the verge of the prmeterEsets over the most reent (ve trding dys @4previousES4AF poresting performne using the 4ltest4 prmeterEset in the dily model is sustntilly worse thn for the p nd ps modelsF roweverD the simple fw model tully performs slightly etter thn oth the p nd ps model when it is used in the mturityEwise settingF his strongly illustrtes the dvntges of the )exiility o'ered y the mturityEwise lirtionF he results re lrgely similr when the 4previousES4 prmeterEset is usedF sn terms of priing performne the fw model lso performs suE stntilly etter using mturityEwise lirtion thn dily lirtionF roweverD oth for the mturityEwise nd the dily setting it is outperE formed y the p modelF he mturityEwise fw model still outperE forms the mturityEwise ps model in terms of priing errorsF
109
Year
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2010 2011
0.2251 0.3258 0.2246 0.1824 0.1753 0.2192 0.1596 0.1748 0.2072 0.2099 0.2942 0.2905 0.3143 0.3281 0.3019 0.2831 0.3009 0.2536 0.2275 0.2200 0.2758 0.2493 0.2494
201.0000 159.0000 198.0000 243.0000 173.0000 219.0000 234.0000 189.0000 139.0000 151.0000 161.0000 241.0000 221.0000 227.0000 226.0000 238.0000 250.0000 237.0000 220.0000 33.0000 114.0000 20.0000 186.0909
Table 7.4. Means of the parameter values and ts for the daily cli-
brated BSM model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent.
e see tht the fw is only performing quite well for tEtheEmoney optionsD while it performs the worst of ll models using ny lirtion method for the fr inEtheEmoney optionsF fw performs out eqully well in the mturityEwise setting indeE pendently of mturityF ine oth the p nd ps models re performing sustntilly etter for shortEmturity options thn for longEmturity optionsD these models re gretly outperforming the fw model for shortEmturity options while the fw model performs no worse thn the p model nd etter thn the ps model for longEmturity optionsF sn onlusionD fw works quite well for ew options @eqully well independent of timeEtoEmturityAF roweverD it performs signi(ntly worse thn the more dvned models for ll the other types of options ! in prtiulr for the fr inEtheEmoney optionsF his shows tht stoE hsti voltility orD lterntivelyD inventory dtD re useful for priing options outEofEsmpleD oth in the rossEsetion @priing performneA nd hed in time @foresting performneAF VF
Conclusion
e on(rm the empiril results in the literture nd (nd negtive reltionship etween inventory nd futures priesF et the sme timeD we
110
Year
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2010 2011
0.0757 0.1800 0.0581 0.0246 0.0259 0.0172 0.0117 0.0303 0.0201 0.0237 0.0372 0.0183 0.0463 0.0232 0.0277 0.0351 0.0160 0.0186 0.0090 0.0035 0.0144 0.0019 0.0335
201.0000 159.0000 198.0000 243.0000 173.0000 219.0000 234.0000 189.0000 139.0000 151.0000 161.0000 241.0000 221.0000 227.0000 226.0000 238.0000 250.0000 237.0000 220.0000 33.0000 114.0000 20.0000 186.0909
for the BSM model calibrated daily, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-ofsample equivalent.
(nd n interesting nonEliner timeEstruture of the stohsti inventory proessF sing mturityEwise lirtion pproh improves sustntilly on oth inE nd outEofEsmple priing errors ompred to dily liE rtionF sing the mturityEwise lirtion pproh the ps model only perE forms slightly worse thn the p modelD whih ontins one extr pE rmeter to lirte ! espeilly worth noting is n interesting trdeEo' etween prmeter stility @ndD onsequentlyD hedgingEperformneD whih we ttriute to prsimonyA nd priingEperformneF e (nd tht the ps model often performs well in ses ssoited with fundmentlsEdriven priesD iFeFD for the ontrts of high intrinsi vlue or with longEmturity @where inventories my e n importnt longEterm ftorAF e elieve tht the extendedD threeEftor models might o'er good pltform for further study of the neessry (neEtuning of the modeled reltionships etween the prieD voltilityD nd inventory in order to hieve the optiml performne for given mrket nd pplitionF
111
Year
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2010 2011
0.2362 0.2318 0.2033 0.1761 0.1737 0.2185 0.1587 0.1637 0.2110 0.2058 0.2783 0.2930 0.3119 0.3272 0.2912 0.2771 0.2998 0.2526 0.2284 0.2198 0.2696 0.2496 0.2429
201.0000 159.0000 198.0000 243.0000 173.0000 219.0000 234.0000 189.0000 139.0000 151.0000 161.0000 241.0000 221.0000 227.0000 226.0000 238.0000 250.0000 237.0000 220.0000 33.0000 114.0000 20.0000 210.0000
Table 7.6. Medians of the parameter values and ts for the daily
clibrated BSM model, 1988-2011, using the number of observations shown in the last column. In-sample t is the minimized value of criterion function Eq. (4.1) and out-of-sample t is the out-of-sample equivalent.
BSM 0.04 0.19 0.17 0.06 0.11 0.14 0.14 0.26 0.26 0.35 0.22 0.22 0.22 0.22 0.25 0.24 0.20 0.17 0.28 0.35 0.34 0.21 0.26 0.21
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Mean
Averages of RMSRE for BSM model, 19892011. Forecasts of option prices are calculated using the parameter-set from the previous day.
Table 7.7.
112
casts of option prices are calculated using the average of the parameter set on the past 5 days.
References
fkshiD qFD gF goD nd F ghen @IWWUAF impiril performne of lterntive option priing modelsF Journal of Finance 52 @SAD PHHQ! RWF froneEedesiD qF nd F iF hley @IWVUAF i0ient nlyti pproxiE mtion of merin option vluesF Journal of Finance 42 @PAD QHI!PHF fenhmouD iFD iF qoetD nd wF wiri @PHIHAF ime dependent heston modelF SIAM Journal on Financial Mathematics 1 @IAD PVWF fesseminderD rFD tF pF goughenourD F tF eguinD nd wF wF moller @IWWSAF wen reversion in equilirium sset priesX ividene from the futures term strutureF Journal of Finance 50 @IAD QTI!USF flkD pF @IWUTAF he priing of ommodity ontrtsF Journal of Financial Economics 3 @IEPAD ITU!IUWF frennnD wF tF @IWSVAF he supply of storgeF The American Economic Review 48 @IAD SH!UPF frennnD wF tF @IWWIAF The Price of Convenience and the Valuation frodieD wFD wF ghernovD nd wF tohnnes @PHHUAF wodel spei(E tion nd risk premiX ividene from futures optionsF Journal of Finance 62 @QAD IRSQ!IRWHF
of Commodity Contingent Claims. In: Lund, D., and Oksendal, B. (eds), Stochastic Models and Option Values. xorth rollndF
113
ghristo'ersenD FD F restonD nd uF tos @PHHWAF he shpe nd term struture of the index option smirkX hy multiftor stohsti voltility models work so wellF Management Science 55 @IPAD IWIR! IWQPF gortzrD qF nd iF F hwrtz @IWWRAF he vlution of ommodityE ontingent limsF The Journal of Derivatives 1 @RAD PU!QWF pmD iF pF nd uF F prenh @IWVUAF gommodity futures priesX ome evidene on forest powerD premiumsD nd the theory of storgeF The Journal of Business 60 @IAD SS!UQF pmD iF pF nd uF F prenh @IWVVAF fusiness yles nd the ehvior of metls priesF Journal of Finance 43 @SAD IHUS!WQF qillonD tF @IWWIAF he term strutures of oil futures priesF yxford snstitute for inergy tudiesF qemnD rF @PHHSAF Commodities and Commodity Derivatives : Modelling and Pricing for Agriculturals, Metals and EnergyF iley piE nneF qemnD rFD xF il urouiD nd tF gF ohet @IWWSAF ghnges of numrireD hnges of proility mesures nd priing of optionsF Journal of Applied Probability 32D RRQ!RSVF qemnD rF nd F xguyen @PHHSAF oyen inventory nd forwrd urve dynmisF Management Science 51 @UAD IHUT!IHWIF qemnD rF nd F yhn @PHHWAF porwrd urvesD srity nd prie voltility in oil nd nturl gs mrketsF Energy Economics 31 @RAD SUT!SVSF qisonD F nd iF F hwrtz @IWWHAF tohsti onveniene yield nd the priing of oil ontingent limsF Journal of Finance 45 @QAD WSW!UTF restonD F vF @IWWQAF e losedEform solution for options with stohsti voltility with pplitions to ond nd urreny optionsF Review of Financial Studies 6 @PAD QPU!RQF tingEhi nd vF u @PHHRAF pei(tion nlysis of option priing models sed on timeEhnged lvy proessesF The Journal of Finance 59 @QAD IRHS!IRQWF tuD xF nd F hong @IWWWAF en pproximte formul for priing merE in optionsF The Journal of Derivatives 7 @PAD QI!RHF uldorD xF @IWQWAF peultion nd eonomi stilityF Review of Economic Studies 7 @IAD I!PUF vindstrmD iFD tF trjyD wF frodnD wF iktorssonD nd tF rolst @PHHVD peruryAF equentil lirtion of optionsF Comput. Stat. Data Anal. 52D PVUU!PVWIF xgD F uF nd F gF irrong @IWWRAF pundmentls nd voltilityX torgeD spredsD nd the dynmis of metls priesF The Journal of Business 67 @PAD PHQ!QHF ossD F @IWWUAF redging long run ommitmentsX ixerises in inomE plete mrket priingF Economic Notes 26 @PAD WW!IQPF
114
outledgeD fF FD hF tF eppiD nd gF F ptt @PHHHAF iquilirium forwrd urves for ommoditiesF Journal of Finance 55 @QAD IPWU! IQQVF hwrtzD iF F @IWWUAF he stohsti ehvior of ommodity priesX smplitions for vlution nd hedgingF Journal of Finance 52 @QAD WPQ!UQF elserD vF qF @IWSVAF putures trding nd the storge of otton nd whetF Journal of Political Economy 66 @QAD PQQ!PSSF rolleD eF fF nd iF F hwrtz @PHHWAF nspnned stohsti voltilE ity nd the priing of ommodity derivtivesF Review of Financial Studies 22 @IIAD RRPQ!RRTIF orkingD rF @IWRWAF he theory of the prie of storgeF American Economic Review 39D IPSR!IPTPF huD tF @PHHHAF Modular Pricing of OptionsF pringerD xew orkF
115
WF
WFIF FVI model. sn the ove model formultion we implemented inventory in the prie dynmis through deterministi formulF en lterntive pprohD nd more )exile oneD is to inlude inventory s third ftor suh tht the model @psA n e formulted s
dFt =
Vt Ft dWtF , Vt dWtV ,
WF,WI WV ,WI
sn this model we let inventories in)uene the dynmis of the futures through the long run vlue of stohsti voltility nd through the orreltion dynmisF WFPF Heston model. he reston @IWWQA model ssumes tht S D the prie of the ssetD is determined yX
t St dWtS dt = ( t ) dt + t dWt
where D the instntneous vrineD is gs proessF xoteD tht the prtiulr form of the drift term in gs proess indues menE reversion of the voltility proess @whih (ts the empiril evidene etter thn nonEreverting nonEzero driftAF fenhmou et lF @PHIHA onsider timeEdependent reston modelD with timeEvrying {t } nd {t }F WFQF FVV model. e onsider model for the prie of future onE trt in reston type model ugmented y n dditionl proess for voltility @whih n lterntively e n inventory ftorD leding to nother vrition of ps modelAF sn the se for stoks the model hs previously een onsidered in very generl setup in hu @PHHHA nd is pplied for lirting stok options in ghristo'ersen et lF @PHHWA in less generl setupF es in ghristo'ersen et lF @PHHWA we mke simplifyE ing ssumptions on the orreltion struture nd formulte the model
116
dFt =
Q Q d[WF 1 , WV 1 ]t = F 1V 1 dt
dVt2 = aV 2 bV 2 Vt2 dt + cV 2
dVt1 = aV 1 bV 1 Vt1 dt + cV 1
Q Q d[WF 2 , WV 2 ]t = F 2V 2 dt Q Q d[WF 1 , WV 2 ]t = 0 Q Q d[WF 2 , WV 1 ]t = 0 Q Q d[WF 1 , WF 2 ]t = 0 Q Q d[WV 1 , WV 2 ]t = 0 Q Q Q Q he orreltion mtrix of the rndom soures WF 1 , WF 2 , WV 1 , WV 2 n e summrized s 1 0 F 1V 1 0 0 1 0 F 2V 2 . := F 1V 1 0 1 0 0 F 2V 2 0 1
iuropen ll options re vlued vi pourier nlysis nd formuls presented in ghristo'ersen et lF @PHHWAF WFRF FVL model. e onsider model for the prie of future onE trt in reston type model ugmented y ltent proessF he model is formulted in the riskEneutrl world s
dFt =
dVt = aV (bV Vt ) dt + cV
Q Q d[WF , WV ]t = F V dt, Q Q ]t = F L dt, d[WF , WL Q Q d[WV , WL ]t = V L dt.
@WFRA
dLt = aL (bL Lt ) dt + cL
WFRFIF The Option Valuation Formulas. he orreltion mtrix of the Q Q Q , WL is rndom soures WF , WV
:= V L F V
V L F V 1 F L , F L 1
117
nd introduing iener proesses W nd W oth mutully indepenE Q Q Q dent nd independent of WL , WV D nd WF the gholesky deomposiE tion yields
V L F V
0 1 V L 2 F L F V V L
1V L 2
0 0
(F L F V V L )2 V L 2 1
F V 2 + 1
@WFSA
F L F V V L 1 2 VL
Q WL +
Q Q = F V WV + WL + W ,
2 2 2 F L + F V + V L 2F L F V V L 1 W 2 VL 1
@WFTA
where
= =
1 2 V L,
2 2 2 F L + F V + V L 2F L F V V L 1 , 2 VL 1 F L F V V L . = 1 2 VL
1 dXt = Vt dt + 2
Q Vt dWF,t .
@WFUA
e de(ne the donExikodym derivtive of hnge from the riskE neutrl mesure Q to the mesure QF ssoited with numerire Ft Y see qemn et lF @IWWSA
F gt :=
r (s) ds
0
Ft , F0
@WFVA
where
t := exp
0
rds .
e know tht disounted prie proess of iuropenEstyle option on futures ontrt F with strike prie K is mrtingle under the
118
M (Q)F rene
C (F0 , T ; K ) = E Q = EQ
Q t vet Bt,T := Et denote the vlue of zeroEoupon ond t time T t nd mturity t time T F hen the donExikodym derivtive to hnge from the riskEneutrl mesure Q to the forwrd mesure QB is de(ned y dQB 0 BT,T B gt := |Ft = . @WFWA dQ T B0,T
xote BT,T = 1F e n now pply the donExikodym derivtive from @WFVA to the (rst term nd the donExikodym derivtive from @WFWA to the seond term
F B C (F0 , T ; K ) = E Q F0 gT 1(FT >K ) E Q B0,T gT K 1(FT >K ) F B = F0 E Q gT 1(FT >K ) B0,T KE Q gT 1(FT >K )
@WFIHA
where k = ln (K ) nd Q (XT > k ) is the proility tht the option (nishes inEtheEmoney @swAD under the Q mesureF sing the inverse pourier trnsformD the iuropenEstyle ll option vlution formul @WFIHA n e written s @where we insert the sw proilitiesA
where
Re F ()
exp(ik ) d i exp(ik ) d. i
nd
QB (XT > k ) =
Re B ()
119
imilrlyD the iuropenEstyle put option vlution formul eomes @where we insert the yw proilitiesA where
Re F ()
exp(ik ) d i exp(ik ) d. i
nd
QB (XT < k ) =
Re B ()
xote tht we n lso otin the priing formul for put options y using putEll prityF WFRFPF Characteristic function F . o lulte the option pries we re interested swGyw proilities under the QF mesureF he hrteristi funtion under the proility mesure QF is de(ned y sing the donExikodym derivtive @WFVAD we n otin the the hrE teristi funtion under the originl riskEneutrl mesure ine we ssume onstnt interest rte nd sustituting this into @WFIIA yields
F F () = E Q gT exp {iXT } . F gT = exp {rT + XT X0 } ,
F () := E Q [exp {iXT }] .
@WFIIA
= E Q [exp {rT X0 + (1 + i) XT }] . 1 2
T T
@WFIPA
XT = X0
Vt dt +
0 0
Q Vt dWF,t .
F () = E
Vt dt +
0 Q Vt dWF,t 0
Q Vt dWF,t
= E Q exp RF exp AF
0
Vt dt + B F
0
1 F where RF = iX0 rT D AF = 2 B D nd B F = (1 + i)F ustituting Q @WFTA for dWF,t yields T T Q AF 0 Vt dt + B F F V 0 Vt dWV,t Q F F () = E exp R exp T T Q +B F 0 Vt dWL,t + B F 0 Vt dW
120
T 0
T 0
exp R
exp
AF F +B
T Q F T V dt + B Vt dWV,t t V 0 0 T F T Vt dW + B Vt dW 0 0
T 0
Vt dW
nd sine W nd W re independent iener proesses nd I1 (T ) = T T Vt dW nd I2 (T ) = 0 Vt dW re mrtingles under Q mesure 0 it follows @where the E opertor ts solely nd seprtely under the
dependent stochastic variables, X and Y and dene X := E eX = eX . Further dene variables Zi , i = 1, 2, 3, mutually independent and independent of X and Y . Then
11To show this several observations need to be made. In general, consider two
eZi X |X
= E E eY X |X
3
eZi X
i=1
= E E eY X
i=1 3
eZi X |X
= E eY X
eZi X .
i=1
121
lws of W nd W A y the st o isometry tht Q F F T exp B exp R Vt dWV,t V 0 Q F () = E T 1 F 2 T F Vt dt + 1 exp AF 0 Vt dt + 2 B B 2 0 Q F T exp RF exp BV V dW t V,t 0 2 2 = EQ F F . (B ) +(B ) T F exp A + Vt dt 2 0
T T
T 0
Vt dt
VT V0 = aV bV T aV
T 0 Q Vt dWV,t =
Vt dt + cV
0 0 T 0
Q Vt dWV,t
V0 aV bV T aV VT + cV cV cV cV
Vt dt.
ustituting into the hrteristi funtion yields T VT aV bV T aV V0 F V dt exp RF exp BV + t cV cV cV cV 0 2 2 F () = E Q F F (B ) +(B ) T F exp A + Vt dt 2 0 F BV exp RF cV (V0 + aV bV T ) = EQ F 2+ BF 2 F (B ) ( ) T BV F F aV Vt dt + cV VT exp A + + BV cV 2 0
T
=E
exp R
sF 1
(V0 + aV bV T ) exp
sF 2
Vt dt + sF 1 VT
T
Vt dt + sF 1 VT
where
sF 1 =
F BV cV (1 + i) = (F V + V L ) cV
sF 2
= A + =
F B
F + B 2
2 F + BV
aV cV
(1+i)2 2
1 (1 + i) + + 2 2 2 aV aV + (1 + i) F V cV + (1 + i) V L cV
(1+i)2 2 2
= (1 + i)
aV 1 1 2 2 (F V + V L ) + (1 + i) 2 + cV 2 2
122
y (VT , T ) = E Q exp sF 2
Vt dt + sF 1 VT
eording to the peynmnEu theorem the expeted vlue stis(es the oneEdimensionl hi y y 1 2 2y V y + a ( b V ) = sF + V , V V 2 T V 2 V 2 with the oundry ondition
y (V0 , 0) = exp sF 1 V0 .
he solution to this hi is given y where
nd
1 =
2 F a2 V + 2 cV s 2
F 2 = 21 e1 T + aV + 1 c2 V s1
1 e1 T .
WFRFQF Characteristic function B . o lulte the option pries we re interested swGyw proilities under the QB mesureF he hrteristi funtion under the proility mesure QB is de(ned y
B we n otin the the hrteristi funtion under the origE sing gT inl riskEneutrl mesure Q B B () = E Q gT exp {iXT } ,
T 0 0 BT,T T = = 1, = 0 T B0,T T T
B () := E Q [exp {iXT }] .
@WFIQA
B () E Q [exp {iXT }] 1 2
T T
@WFIRA
XT = X 0
Vt dt +
0 0
Q Vt dWF,t .
123
B () = E Q exp i X0
1 2
Vt dt +
0 T 0
Q Vt dWF,t T
,
Q Vt dWF,t
Vt dt + i
0 0
B () = E
1 i 2
T 0
Q ustituting @WFSA for dWL,t yields T T Q 1 Vt dWV,t V 0 2 i 0 Vt dt + iF T Q Q B () = E exp {iX0 } exp +i V L 0 Vt dWV,t T T +i 0 Vt dW + i 0 Vt dW T 1 i V dt t 2 0 T Q Q = E exp {iX0 } exp + (iF V + i V L ) 0 Vt dWV,t +i T V dW + i T V dW 0 t 0 t T T Q B AB 0 Vt dt + BV Vt dWV,t 0 = E Q exp {iX0 } exp , B T B T +B Vt dW + B Vt dW 0 0 B B B where AB = 1 iD BV = iF V + i V L D B = i D nd B = 2 i F e n now rewrite the hrteristi funtion s Q B T exp { iX Vt dWV,t 0 } exp BV 0 Q B () = E T B T B T exp AB 0 Vt dt + B Vt dW + B Vt dW 0 0
T nd sine W nd W re frownin motions nd I1 (T ) = 0 Vt dW T nd I2 (T ) = 0 Vt dW re mrtingles with I2 (0) = 0 it follows tht Q B T exp { iX Vt dWV,t 0 } exp BV 0 Q B () = E T 1 B 2 T B 2 T exp AB 0 Vt dt + 1 B Vt dt + 2 B Vt dt 2 0 0 Q B T exp {iX0 } exp BV Vt dWV,t 0 Q =E B 2+ BB 2 . (B ) ( ) T B Vt dt exp A + 2 0
124
VT V0 = aV bV T aV
T Q Vt dWV,t =
Vt dt + cV
0 0 T
Q Vt dWV,t
VT V0 aV bV T aV + Vt dt. cV cV cV cV 0 0 ustituting into the hrteristi funtion yields T aV bV T aV V0 B VT exp {iX0 } exp BV cV cV cV + cV 0 Vt dt B () = E Q B 2+ BB 2 (B ) ( ) T B Vt dt exp A + 2 0 B V0 B aV bV T exp iX0 BV B V cV cV 2 2 = EQ B B B B + B ( ) ( ) T B a B V V exp AB + + B V dt + V t T V 2 cV cV 0
T Q = exp iX0 sB exp sB 1 (V0 + aV bV T ) E 2 0
Vt dt + sB 1 VT
where
sB 1 =
B BV cV i (F V + V L ) = cV
sB 2
= A +
B B
B + B 2
2 B + BV
aV cV
y B (VT , T ) = E Q exp sB 2
Vt dt + sB 1 VT
eording to the peynmnEu theorem the expeted vlue must stE isfy the oneEdimensionl hi
125
where
B H1 =
nd
B B 1 T B B 2 = 21 e + aV + 1 c2 V s1
1 e1 T .
2011:13 2011:16
2011:22
2012-1
2012-11
ISBN 9788790117931