Sie sind auf Seite 1von 4

ECON0022

Econometrics for Macro and Finance


Problem Set III
Dennis Kristensen

Autumn 2018

1 Practice Problems
1. Consider iterative forecasts over the horizon T + 1; :::; T + h using an AR(1)
model, Yt = + Yt 1 + "t , given data up to time T . We ignore parameter
uncertainty and treat the parameters and as known such that the iterative
forecast takes the form

YT +kjT = + YT +k 1jT ; k = 1; 2; :::; h;

where YT jT = YT . We wish to analyze the behaviour of the forecasts in the


stationary (j j < 1) and non-stationary ( = 1) case.

(a) Show that the …nal h-step ahead forecast, YT +hjT , can be expressed as
n o
YT +hjT = h YT + 1 + + 2 + ::: + h 1 :

[hint: Use backward recursion starting at time T + h and stopping at time


T]
(b) Use the above expression to argue that for long-horizon forecasts (h very
large):

If j j < 1 : YT +hjT = Y;
1
If = 1 : YT +hjT ! 1;

where YT +hjT ! 1 should be interpreted as YT +hjT exploding (increas-


ing in absolute value without any bound) as h ! 1. Interpret the above
results.
(c) Let eT +h be h-step ahead forecast error de…ned as:

eT +h = YT +h YT +hjT :

Expresss eT +h as a function of "T +1 ;P


:::; "T +h . [Hint: Use backward recur-
sion to show that YT +h = YT +hjT + hi=01 i "T +h i ]

1
(d) Show that
n o
2 2 2(h 1)
E [eT +h ] = 0; Var (eT +h ) = " 1+ + :::: :

What is the h-step MSFE for the iterative procedure using an AR(1)
model? What happens to the MSFE as h grows large?
(e) Explain how you could use the results in question (d) to construct a 95%
forecast interval for your h-period ahead point forecast.

2. Suppose we have obtained the following estimates of an ADL model for annual
US unemployment (unemp) using in‡ation (inf l) as additional preditor,

d t = 1:304 + :647unempt
unemp 1 + :184inf lt 1; ^ " = :883: (1)
(:490) (:084) ((:041)

(a) Suppose that unemp2012 = 7:8 and inf l2012 = 1:3. What is your forecast
of unemployment in 2013?
(b) Suppose the errors in the ADL model are normally distributed. Compute
a 95% and 67% forecast interval.
(c) Can you use the above ADL model to forecast employment for 2014?
Explain.
(d) You decide to build a model for in‡ation in addition to the above model
for unemployment. You …nd that an AR(1) model provides an adequate
description of in‡ation,
c lt = 1:277 + :665inf lt
inf 1; ^ " = :653: (2)
(:558) (:107)

Use eqs. (1)-(2) to compute a forecast of in‡ation and unemployment for


2014.
Note: the joint model for inf l and unemp as given in eqs. (1)-(2) is
a particular example of so-called Vector Autoregressive (VAR) models.
These allow for joint modelling of multiple economic time series.

3. Prove the following results:

(a) Let W be a random variable with mean W and variance 2 . Show that
W
for any constant c:
h i
E (W c)2 = 2W + ( W c)2 :

(b) Consider the problem of forecasting YT +1 using data on YT ; YT 1 ; :::. Let


denote fT +1jT denote some forecast of YT +1 based on YT ; YT 1 ; ::: Let
h i
2
M SF E = E YT +1 fT +1jT

2
denote the unconditional mean squared forecast error associated with
fT +1jT . Show that M SF E is minimized w.r.t. fT +1jT at fT +1jT = E [YT +1 ].
Interpret the result.
(c) Consider instead the conditional mean squared forecast error associated
with fT +1jT ,
h i
2
M SF ET = E YT +1 fT +1jT jYT ; YT 1 ; ::: :

Show that M SF ET is minimized w.r.t. fT +1jT at fT +1jT = E [YT +1 jYT ; YT 1 ; :::].


Interpret the result.
(d) Which of the two forecasts, fT +1jT or fT +1jT , would you rather use? Ex-
plain.

4. Let Yt = log (GDPt ) be change in US log-GDP using USMacro_Quarterly.dta


available on Moodle under "Data Sets". The goal is to choose an appropri-
ate AR model for forecasting. After opening the data set in STATA, …rst set
time [tsset time] and then create Yt [gen lgdp = log(realgdp); gen dlgdp
=lgdp - L.lgdp].

(a) As a …rst way of choosing the number of lags for the AR model for Yt ,
compute the AIC and BIC and LR criteria for the given sample [this can
be done with the STATA command varsoc dlgdp - note that STATA call
the BIC criterion SBIC]. Which model(s) does each of the three criteria
recommend?
(b) As a second way of choosing the appropriate number of lags, we can use
the pseudo-forecast evaluation criterion.
To this end, we wish to compute rolling pseudo-forecasts as described in
the slides. To perform this in STATA, you need to use the STATA com-
mands that you …nd in the …le ’rolling forecast US-dgp.txt’ posted
on Moodle/Problem Sets which performs the forecast evaluation using an
AR(3) model.
The …rst part of the code says:
//Pseudo-forecast of dlgdp using AR(3) model for T=150,...,252.
gen f_dlgdp3=.
local start 150
local end 252
forv T=‘start’/‘end’ {
quietly regress dlgdp L(1/3).dlgdp in 1/‘T’
quietly replace f_dlgdp3 = _b[_cons] + _b[L1.dlgdp]*dlgdp[‘T’]
+ _b[L2.dlgdp]*dlgdp[‘T’-1] + _b[L3.dlgdp]*dlgdp[‘T’-2] in ‘T+1’/‘T+1’
The above code runs a loop over T = 150; 151; :::; 252, where for each value
of T STATA does the following:

3
i. Estimate an AR(3), Y^t = ^ T + ^ 1;=t Yt 1 + ^ 2;T Yt 2 + ^ 3;T Yt 3 , using
only observations fY1 ; :::; YT g.
ii. Compute Y^T +1jT = ^ T + ^ 1;T YT + ^ 2;T YT 1 + ^ 3;T YT 2 and save this
value in f_dlgdp3.
To perform steps (i)-(ii), copy the above part of the code from ’rolling
forecast US-dgp.txt’and paste it into the command prompt in STATA
(and press enter...).
If you open the Data Editor, you’ll see that the variable f_dlgdp3 has
been created containing the rolling pseudo-forecasts, Y^T +1jT .
(c) To visualize the performance of the AR(3) as forecast model, plot the
actual observed value of YT +1 together with its pseudo-forecasts. This can
be done with the command ’tsline dlgdp f_dlgdp3 if tin(1984q2,)’.
Does the AR(3) model appear to do a good job?
(d) Next, compute pseudo-RMSFE in STATA by de…ning a new variable which
is the di¤erence between by Y and Y^ , and compute its sample variance
[gen e3 = dlgdp - f_dlgdp3; summarize e3]
(e) Next, re-do questions (b)-(c), this time using an AR(1) model. This can
be done by copy-pasting the second part of the code in the txt …le rolling
forecast US-dgp.txt into the command line in STATA. This generates a
new variable f_dlgdp1 that contains the rolling forecasts using an AR(1)
model. Plot this new forecast vs. actual outcome, and compute the cor-
responding pseudo-RMSFE.
(f) We now have two pseudo-RMSFE’s - one using an AR(1) model, and one
using an AR(3) model. In terms of pseudo-RMSFE, which of the two
models would you recommend? How does this recommendation compare
to the model(s) chosen by AIC, BIC and LR, respectively, in question (a)?

Das könnte Ihnen auch gefallen