Sie sind auf Seite 1von 10

JOT

The Voices of Influence | iijournals.com

W W W. I I J O T. C O M

FALL2012VOLUME7NUMBER4

Trade Cost:
Handicapping on PAR
VLAD RASHKOVICH AND ARUN VERMA

Sponsored by:

Goldman Sachs
Morgan Stanley
D R ASHKOVICH
UBS
global
business

ger of trade anaat Bloomberg, L.P.


w York, NY.

he most commonly used trade cost


models are based on the research of
Almgren et al. [2000, 2003, 2005]
on the market impact of an equity
trade. Almgren has made great contribu-

segmented the traders job into two main


tasks: momentum management and liquidity
sourcing. See Rob Shapiros work at Morgan
Stanley Investment Management (Mehta
[2009]).

Article Abstract
Trade Cost: Handicapping on PAR
Vlad Rashkovich and Arun Verma

We analyze the prevailing approach to estimating implicit trade costs and suggest improvements
that create greater precision. Our model studies buy-side parent orders to prove that Size/ADV is
a sublinear factor, which means that doubling an order does not double the trade cost. The same
precision cannot be achieved from the child order data available to the sell side. After improving
on existing methods, we cast doubt on whether it is realistic to delineate permanent and temporary costs when numerous market participants influence prices simultaneously. As a result, we
introduce a new highly accurate pre-trade cost model with predictive power (R2) of up to 26%.

Trade Cost:
Handicapping on PAR
Vlad Rashkovich and Arun Verma

Vlad R ashkovich
is the global business
manager of trade analytics at Bloomberg, L.P.
in NewYork, NY.
vrashkovich1@bloomberg.net

A run Verma
is a senior quantitative
analyst at Bloomberg,
L.P. in New York, NY.
averma3@bloomberg.net

he most commonly used trade cost


models are based on the research of
Almgren et al. [2000, 2003, 2005]
on the market impact of an equity
trade. Almgren has made great contributions to the field of trade cost analysis, and
the popularity of his model demonstrates its
credibility. By reconsidering two assumptions
that Almgren made about timing and reversion, however, we have developed a dynamic
approach that adjusts based on the participation
rate of the trade and estimates costs during the
trade instead of after it is complete.
We also question whether it is necessary, or even possible, to distinguish between
temporary and permanent costs, and we offer
a simplified version of our method that looks
at trade cost as a whole. This new framework
results in higher predictive power when estimating trade costs.
It is worth noting that although
Almgrens model is the most popular, it is not
the only one. Other models attempt to look at
the market microstructure by analyzing each
stock on a tick-by-tick basis. This approach
requires a large capital investment to implement and shows no clear indication of greater
predictive power.
Estimating trade Costs

To assess what type of predictive power


we should seek from a trade cost model, we
Fall 2012

segmented the traders job into two main


tasks: momentum management and liquidity
sourcing. See Rob Shapiros work at Morgan
Stanley Investment Management (Mehta
[2009]).
To manage momentum, a trader must
assess the favorability of trading conditions and
make a decision about the execution speed. To
source liquidity, a trader must find the other
side(s) of the trade based on the desired execution speed.
Because momentum cannot be measured before the trade, cost models measure
the liquidity sourcing component, which can
be assessed before the trade occurs by looking
at order size, stock volatility, bidask spread,
and other factors.
Our research shows that momentum
management accounts for roughly 83% of
implementation shortfall (Perold [1988]),
while liquidity sourcing is responsible for
about 17%. Thus, trade cost models are trying
to estimate a fairly weak signal (the traders
ability to source liquidity) in a sea of noise
called momentum. As a result, we can expect
any trade cost model to have a predictive
power of much less than 17%.
almgrens tribute

Almgren et al. [2000, 2003 and 2005]


conducted detailed studies to delineate trade
cost into the permanent and temporary
The Journal of T rading

impacts of the trade. Our research confirmed the 2005


studys findings that such models have a predictive
power, or R 2, of a few percent.
Almgren found that after 30 minutes, price reversion reaches a point when the effects of a temporary
liquidity draw have dissipated. So, he chose to measure
the permanent impact of a trade 30 minutes after the
last fill.
As shown in Exhibit 1, Almgren takes half of the
permanent impact, because the earlier parts of the trade
experienced a much smaller impact and later parts experienced the entire impact.
At the end, the model calculates the temporary
impact as the arrival price versus the average execution
price, which it tries to solve for, minus half the permanent impact.
To express Almgrens model in a conceptual
formula:
Temporary Impact = Market Impact
0.5 Permanent Impact (1)
Most broker models also add an instant component, which estimates the depth of crossing into the
bidask spread.
REconsidering Almgren

Acknowledging Almgrens great contribution to


this field, we reconsider two assumptions in his model:
timing and constant reversion.

Exhibit 1

Framework of Almgrens Model

T rade Cost: H andicapping on PAR

With respect to timing, Almgrens model waits the


entire time one has been trading, plus an additional 30
minutes after the last fill, to measure the signal.As we mentioned above, the liquidity sourcing signal is much weaker
than the momentum management one, so we would rather
try to capture this signal throughout the trade, not 30
minutes after its end.
With respect to constant reversion, whether the
duration of an order was two minutes and the participation rate was 1%, or the duration was five hours with 30%
participation, Almgrens model, in both cases, waits 30
minutes for the temporary impact to dissipate. We have
found that a trades temporary impact varies based on the
participation rate, so a one size fits all approach reduces
the models accuracy.
participation arrival reversion (PAR)

To address the drawbacks of Almgrens assumptions about timing and constant reversion, we introduce
Participation Arrival Reversion (PAR), a dynamic way
to determine how long to wait to measure the impact
of the trade.
Once the order hits the market, we start listening.
The higher the participation rate, the greater the temporary impact we expect, and thus the longer we wait
to capture the impact. Our research shows that if the
participation rate is 5%, we should wait five minutes to
capture the temporary impact. If participation is 30%,
then we should wait 30 minutes to capture the temporary
impact. Instead of waiting for 30 minutes after
the trade, our model measures the impact right
after the order starts and does so dynamically
based on participation rate.
There are a number of practical considerations for computing PAR. The most important one is that PAR cannot be more than half
the duration of the order. Otherwise, the temporary impact could be larger than the entire
implementation shortfall.
To ensure we include enough trades close
to the end of the trading day, we have Winsorized the data for all values above a 30%
participation rate. So, for all trades with participation >30%, we measure PAR after 30
minutes.

Fall 2012

Exhibit 2

Instant Impact

Framework of Our Trade Cost Model

The formula for computing PAR is


PAR = M in (participation % minutes*
or 0.5 duration or 30 minutes)

(2)

*For each 1% of participation, we will wait 1


minute to measure the temporary impact.
We also take a different approach to calculating the
permanent impact of a trade. Compared to Almgren, we
f lip temporary versus permanent impacts by measuring
temporary first, then solving for the permanent impact.
Our model calculates the permanent impact as the arrival
price versus the average execution price minus half the
temporary impact. See Exhibit 2.
To express our model in a conceptual formula,
Permanent impact = Market impact 0.5
Temporary impact

(3)

Exhibit 3 summarizes the key differences between


Almgrens and our models.
reexamining all trade
costcomponents

With this new method, we decided to reexamine


each of the three components of trade costs: instant
impact, temporary impact, and permanent impact.

Fall 2012

Instant impact occurs due to trading


through the bidask spread. Our research shows
that more-aggressive implementation strategies,
with higher participation rates, cause traders to
cross deeper into the spread.
We observed that when defining implementation shortfall as the arrival price minus
the average execution price, the arrival price is
often measured as a midpoint between bid and
ask at the moment the order was received. For
a non-aggressive execution strategy however,
traders can sit on their side (ask or bid) and
do not have to come to the midpoint. Thus,
this factor can be negative. In other words, we
believe that instant impact helps to calibrate
arrival price by allowing low-participation trades to stay
closer to their side of the spread and not come to the midpoint. Our research finds that trades with a participation
rate of up to 5% cross, on average, only a quarter of the
spread. At the same time, trades with a participation rate
of 30% and above must fully cross the spread. This factor
is linear and can be expressed as
[0.25 + 3 (max (min (0.3, participation %/100),
0.05) 0.05)] bidask spread %
(4)
In practical trading terms, the instant impact factor
is zero for a trade with a participation rate of 13.3%.
Note that to ensure stability of this factor, it is
important to use a bidask spread calculated over a period
of timenot the spread at the time of order arrival. In
our research, we used the average spread over the last
five trading days.
Temporary Impact

In an earlier section we introduced PAR as a better


way to capture the temporary impact of a trade.
Initially, we thought the participation rate captured
the aggressiveness on its own, making it unnecessary to
look at the duration of the trade. If we use participation
rate only, however, then the temporary impact does not
scale as the trade size grows. In other words, we would
expect a larger temporary impact for a large trade versus
a small one, but participation rate on its own cannot

The Journal of T rading

Exhibit 3

Comparison of Almgren and PAR Models

capture this effect. Thus, it is important to add duration


as a component of the model.
In essence, participation rate versus duration represents a level of aggressiveness that a trader can choose
for an order execution strategy.
Both participation rate and duration are nonlinear
factors, but we would expect duration to be much more
concave. So, if the risk tolerance is high, or risk is not a
factor at all, then the minimal cost would always occur
with the maximum duration and minimum participation rate.
In the end, we found that adding annualized stock
volatility () over the last 30 trading days is an appropriate way to scale this factor. The final formula for the
temporary impact is
Temporary impact
= (participation %/100) 1 (T) 2(5)
where T is duration of the trade.
Permanent Impact

Our research suggests that size/ADV is the only


factor that determines permanent impact. No matter
how quickly or aggressively one trades, the temporary
impact dissipates over time, and the impact that stays
permanently is measured by how many shares changed
hands.
Similar to the temporary impact, we have scaled
the permanent impact by annualized 30-day stock volatility ().

T rade Cost: H andicapping on PAR

Permanent impact = (size/ADV) (6)


We found that the size/ADV factor is sublinear.
When we compare regression results from buy-side
parent orders to child orders received by brokers, we find
on average a 2% boost to R 2 in the parent order level
one from the size/ADV factor. We think that improved
predictive power of the size/ADV factor for orders is
caused by its nonlinearity.
A number of articles were written to prove the
linearity of the permanent impact, including the arbitrage explanation by Huberman and Stanzl [2004].
For example, if the broker receives two child orders of
500,000 shares for the same stock from different buyside firms (Scenario 1) or one child order of 1,000,000
shares from one client (Scenario 2), the prevailing
market view is that permanent impact is the same in
both scenarios.
For child orders (slices that brokers receive from
the buy side), assuming linearity of the size/ADV factor
makes sense. On the parent order level, however, two
orders of 500,000 shares from two buy-side firms are not
equal to one order of 1,000,000 shares from one buy-side
firm. In the case of two orders, we face two investments
decisions and, in the end, holdings of the same stock by
two different firms. In our opinion, this scenario carries
more permanent impact compared with a decision of
one firm to hold 1,000,000 shares. In addition, two child
orders of 500,000 shares will compete with each other
in the market, creating a larger temporary impact and
thus a larger trade cost. We mostly care about the trade
cost as a whole, as we explore in a later section.

Fall 2012

Based on the aforementioned explanation, we


expect size/ADV to be slightly convex (sublinear). When
we look at the entire trade cost model later in this article,
we will revisit whether the arbitrage argument holds up
for the parametric form we suggest.
data set

We used trade data from more than 250 buy-side


firms from around the world that executed millions of
orders of varying sizes, market caps, and level of aggression across a wide range of brokers.
Because we used buy-side data, we have both parent
orders (portfolio managers to buy-side trader) and child
orders (buy-side trader to broker). All broker models,
including Almgrens, are based on child orders.
We have applied the following filters to the data:
Size 1,000 shares
Market cap micro
Order duration 2 min and 1 day
30-day volatility 8 and 100
Size/ADV 1% and 100%
Participation % 1% and 100%
Order life price momentum < abs 3%
Realized impact (arrival versus average execution
price) abs 500bp
Country = United States
Dates = June 1, 2011, to May 31, 2012
We also removed orders executed during the first
two minutes and last five minutes of the day to reduce
the noise when calibrating the model. Subsequently, we
tested that our model works well for these orders. R 2 for
orders at market open and close is only slightly lower
(about 0.5%) than for the rest of the orders.
After applying all the filters, we have about 25,000
parent orders and 40,000 child orders that we can use to
calibrate our U.S. model.
our trade cost model

The new trade cost model combines the three


impacts we have outlined above:
Trade cost = Instant impact + Temporary impact
+ Permanent impact

Fall 2012

(7)

The parametric form for this model is


(participation %) bidask spread %
+ * * (participation %/100) 1
(T) 2 + (size/ADV) (8)
Because instant impact (crossing the spread) and
temporary impact (PAR) happen on the child order
level, although the size/ADV factor should be measured
on the parent order level because of nonlinearity, we use
dual optimization to calibrate the trade cost.
We first calibrate the bidask spread (instant cost)
component and the temporary cost from the child orders.
We then use the calibrated coefficients for instant and
temporary costs in the parent orderlevel regression to
learn the permanent cost component from size/ADV
of parent orders.
Below is a dual objective function we use:
minimize (permanent realized cost PC) 2
+ (temporary realized cost (IC + TC)) 2 (9)
A dual-optimization method is preferable to analyzing temporary and permanent costs separately and
robustly. Optimizations are performed using the LevenbergMarquardt method (Levenberg [1944], Marquardt
[1963]).
The coefficients and confidence intervals we have
learned are
(participation %)
= 0.25 + 3 (max (min (0.3, participation %/
100), 0.05) 0.05)
= 0.023 0.0014
1 = 0.76 0.06
2 = 0.19 0.04
= 0.030 0.0017
= 0.81 0.08
Coefficient = 0.81 0.08 above suggests the
nonlinear nature of size/ADV factor. As we discussed
in the permanent impact section of this article, let us
address a concern of potential arbitrage.
If instant and temporary factors depended on participation rate alone and if we assumed a participation
rate close to zero, we would face potential arbitrage from
the nonlinear size/ADV factor. Because our temporary
impact is also dependent on the duration factor, a participation rate close to zero would cause duration to be very
high. The duration factor in temporary impact would
The Journal of T rading

Exhibit 4

Regression Results with Dual Optimization

compensate for size/ADV sublinearity and ensure the


nonarbitrage nature of the entire model.
It is worth noting that small trades dominate our
data set. To calibrate the model correctly, we created a
matrix of buckets in terms of size/ADV and participation %. We then recorded the count of trades in each
bucket. In the calibration scheme, each trade gets the
weight of (1/bucket count), where bucket count is the
data count of the bucket to which the trade belongs.
This adjustment effectively gives an equal weight for
each bucket and allows us to calibrate the model with
no bias toward the smaller trades.
Exhibit 4 shows the models observed predictive
power (R 2 ).
As shown, the models predictive power grows as
the size/ADV and participation rate of an order grow.
This relationship makes sense because large and aggressive orders have more impact and thus are easier to measure in the noise of momentum created by other market
participants.
Interestingly, at a participation rate greater than
35% we mostly get R 2 > 17%. In an earlier section, we
stated that sourcing liquidity constitutes about 17% and
this is the only part that we can estimate in a pretrade
model. Do we have a contradiction?
Not necessarily. When the participation rate is greater
than 35%, the order itself is a major source of momentum,
and our model captures this momentum in addition to the
liquidity sourcing component.
We believe that our models high predictive power
comes from
a. PAR, which captures the signal (temporary impact)
right away and does so dynamically based on the
participation rate.
T rade Cost: H andicapping on PAR

b. Order level data for size/ADV, given the nonlinearity of permanent impact.
c. Precise calibration of the spread in the instant impact
by allowing it to scale with participation rate from
a negative quarter to a positive half vis--vis the
midpoint.
d. Two-step optimization, which unleashes predictive power hidden in both parent orders and child
orders.
simplified Methodology

If we look at the trade cost components delineation that current models use, we cannot help but note
that the market is a complex combination of cause and
effect reactions triggered by numerous participants. The
separation between permanent and temporary cost thus
might be theoretical and indistinguishable.
If we simplify the goal and try to solve for trade
cost as a whole, then we can do it with single optimization of the same three factors on the parent order
level:
bidask spread % +
(participation %/100) 1 (T) 2
+ (size/ADV) (10)
In essence, our simplified parametric form has two
logical components: order size and execution strategy.
Order size is defined by the amount of shares that an
investor has to trade relative to ADV. Execution strategy
is defined by the traders choice of participation rate,
which drives duration and depth of crossing into the
bidask spread.

Fall 2012

Exhibit 5

Regression Results with Single-Step Optimization (Simplified Methodology)

We should mention that there are advantages and


disadvantages of single-step optimization versus twostep optimization. The two-step optimization leads to
more-robust identification of parameters, because the
permanent and temporary components have a high correlation and the two-step method identifies them separately rather than jointly. The two-step methodology
would lead to lower R 2 numbers, however, because we
lose the models fitting ability by decomposing the trade
cost and modeling its components from parent and child
orders.
So, in our new simplified approach, we use singlestep optimization and produce an easier to grasp and
more straightforward model with slightly higher R 2, as
shown in Exhibit 5.
Conclusion

Trade cost models estimate a signal, which


is fairly weak because it is surrounded by the sea of
momentum, or noise. We introduce PAR (participation arrival reversion) to capture the signal as close to
the source as possible.
Our model enjoys a balance between permanent
and temporary costs, which we believe ref lects stock
market behavior. Permanent cost is dominant for large
size/ADV and temporary cost takes over for small trades
with a high participation rate.

Fall 2012

We also ask whether it is important to measure


trade cost as a whole as opposed to delineating between
temporary and permanant impacts. We realize that for a
portfolio manager with a large position in a given security,
understanding the temporary versus permanent decomposition is equally important. It seems to us that in most
cases, whether for portfolio construction or optimal
trading, market participants are interested in just finding
the trade cost.
In parallel, we cast doubt on whether delineation
between permanent and temporary costs is realistic to
measure, because there are numerous market participants
and their decisions and behavior affect each other.
It might be more straightforward to think about
trade cost as a combination of order size/ADV and execution strategy, defined as the participation rate, which
drives duration and the depth of crossing into the bidask
spread.
As the result, we introduce a new cost model with
high predictive power (R 2 ).
endnote
We are thankful to Bruno Dupire for overseeing the
quantitative side of this project, as well as for his original
thinking and challenging the status quo. We also thank Don
Carey for coining the term PAR and suggesting the title
for this article, Alex Clode for his feedback throughout the
project, and Todd Levinson for his editorial remarks.

The Journal of T rading

References
Almgren, R., and C. Neil. Optimal Execution of Portfolio
Transactions. Journal of Risk, Vol. 3, No. 2 (2000), pp. 5-39.
. Optimal Execution with Nonlinear Impact Functions
and Trading-Enhanced Risk. Applied Mathematical Finance,
Vol. 10, No. 1 (2003), pp. 1-18.
Almgren, R., T. Chee, H. Emmanuel, and L. Hong. Equity
Market Impact. Risk ( July 2005), pp. 57-62.
Huberman, G., and W. Stanzl. Price Manipulation and
Quasi-Arbitrage. Econometrica, Vol. 72, No. 4 ( July 2004),
pp. 1247-1275.
Levenberg, K. A Method for the Solution of Certain NonLinear Problems in Least Squares. Quarterly of Applied Mathematics, 2 (1944), pp. 164-168.
Marquardt, D. An Algorithm for Least-Squares Estimation
of Nonlinear Parameters. SIAM Journal on Applied Mathematics, Vol. 11, No. 2 (1963), pp. 431-441.
Mehta, N. A New Way to Judge the Buyside. Traders Magazine, Vol. 22, No. 292 (March 2009), p. 48.
Perold, A.F. The Implementation Shortfall: Paper vs.
Reality. The Journal of Portfolio Management, 14 (Spring 1988),
pp. 4-9.
To order reprints of this article, please contact Dewey Palmieri
at dpalmieri@ iijournals.com or 212-224-3675.

T rade Cost: H andicapping on PAR

Fall 2012

Das könnte Ihnen auch gefallen