Scribd
Upload
59 views3 pages

Exercises2 - TSA

This document provides exercises on time series analysis and forecasting using autoregressive (AR) and autoregressive moving average (ARMA) models. It asks the reader to derive point forecasts, forecast errors, and other statistical properties for univariate time series data generated from AR(1), AR(2), and ARMA(1,1) processes. The exercises also involve applying a estimated ARMA model to multi-step ahead forecasting and interval forecasting.

Uploaded by

Luca Leimbeck Del Duca
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
59 views3 pages

Exercises2 - TSA

This document provides exercises on time series analysis and forecasting using autoregressive (AR) and autoregressive moving average (ARMA) models. It asks the reader to derive point forecasts, forecast errors, and other statistical properties for univariate time series data generated from AR(1), AR(2), and ARMA(1,1) processes. The exercises also involve applying a estimated ARMA model to multi-step ahead forecasting and interval forecasting.

Uploaded by

Luca Leimbeck Del Duca
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

(International) Bachelor Econometrics and Operations Research

FEB23001(X) Tijdreeksanalyse / Time Series Analysis

Exercises - Basic concepts (Chapter 3)

1. Consider the AR(2) model for a time series variable yt ,

yt = φ1 yt−1 + φ2 yt−2 + εt , t = 1, . . . , T, (1)

where the parameters φ1 and φ2 are such that both roots of the AR(2)-polynomial
φ2 (L) = 1 − φ1 L − φ2 L2 are outside the unit circle, where L is the usual lag operator
(defined as Lk yt ≡ yt−k for all integers k). Furthermore, the shocks εt are independent
and normally distributed with constant variance σε2 , that is εt ∼ N (0, σε2 ).

a. Derive the optimal 1-step, 2-steps and 3-steps ahead point forecasts of yt in case
of the standard quadratic loss function (that is, derive expressions for ŷT +k|T =
E[yT +k |YT ] for k = 1, 2 and 3, where YT denotes the available history of the
time series up to time T ). [You may assume that the parameters φ1 and φ2 are
known.]
b. Express the corresponding forecast errors

eT +k|T ≡ yT +k − ŷT +k|T ,

for k = 1, 2 and 3 in terms of the shocks εt in (1).


c. What is the covariance between the errors for subsequent 1-step, 2-steps and 3-
steps ahead forecasts made at time T and T +1? That is, what is E[eT +j|T eT +j+1|T +1 ]
for j = 1, 2, and 3?

2. The time series yt is generated according to an AR(1) process:

yt = φ1 yt−1 + εt , (8)

where |φ1 | < 1, and εt is a white noise series with E[εt ] = 0, E[ε2t ] = σε2 for all t, and
E[εt εs ] = 0 for all s 6= t.

Suppose that a researcher uses a ‘subset’ AR(2) model1

yt = φ∗2 yt−2 + ηt , (9)

1
A subset AR(p) model includes only a subset of the lags 1, 2, . . . , p. Put differently, the coefficients
of some lags are equal to zero.

Page 1 of 3
where she sets the parameter φ∗2 equal to the square of φ1 in equation (8), that is,
φ∗2 = φ21 . Furthermore, the researcher assumes that ηt is a white noise sequence with
(among others) mean equal to zero. (Note that this assumption may be incorrect!)

The researcher uses the subset AR(2) model in equation (9) to compute one-step
ahead point forecasts for yt at t = T + 1, T + 2, T + 3, . . . using a quadratic loss
function (that is, she chooses the point forecast in order to minimize the expected
mean squared prediction error). The resulting point forecast for T + j + 1 (j =
0, 1, 2, 3, . . . ) is denoted as ŷT +j+1|T +j , with corresponding forecast error eT +j+1|T +j =
yT +j+1 − ŷT +j−1|T +j .

Answer the following questions:

a. Show that the one-step ahead point forecasts obtained from the subset AR(2)
model in equation (9) are (unconditionally) unbiased, that is E[eT +j+1|T +j ] = 0
for all j > 0.
b. Derive an expression for the unconditional variance of the forecast errors eT +j+1|T +j
in terms of the parameters σε2 and φ1 .
c. Derive an expression for the k-th order autocorrelation ρk,e of the one-step
ahead forecast errors (that is, ρk,e is the correlation between eT +j+1|T +j and
eT +j+1−k|T +j−k ) for k = 1, 2, 3, . . . . Interpret your results.

3. Consider the ARMA(1,1) model for a time series variable yt ,

yt = φ1 yt−1 + εt + θ1 εt−1 , t = 1, . . . , T, (11)

where the parameters are such that the model is stationary and invertible, that is
|φ1 | < 1 and |θ1 | < 1. Furthermore, the shocks εt are independent and normally
distributed with constant variance σ 2 , that is εt ∼ N (0, σ 2 ).

a. Derive the optimal 1-step, 2-steps and 3-steps ahead point forecasts of yt in case
of the standard quadratic loss function (that is, derive expressions for ŷT +k|T =
E[yT +k |YT ] for k = 1, 2 and 3, where YT denotes the available history of the
time series up to time T ).
b. Express the corresponding forecast errors

eT +k|T ≡ yT +k − ŷT +k|T ,

for k = 1, 2 and 3 in terms of the shocks εt in (11).

Page 2 of 3
c. Suppose that the parameters in (11) have been estimated using observations yt
for t = 1, . . . , T , with the following results:

φ̂1 = 0.78, θ̂1 = 0.32, and σ̂ 2 = 0.51.

Furthermore, assume that yT −1 = 9.79, yT = 7.24 and the corresponding resid-


uals ε̂T −1 = 0.51 and ε̂T = −0.56.
c.1 What is the implied 2-step ahead point forecast ŷT +2|T ?
c.2 What is the variance of the corresponding 2-step ahead forecast error eT +2|T ?
c.3 Give a 90% “equal-tail probability” interval forecast for yT +2 at time T (i.e.
2-steps ahead). That is, derive the lower bound L bT +2|T and upper bound
U
bT +2|T such that

P [L
bT +2|T < yT +2 < UbT +2|T |YT ] = 0.90, and
bT +2|T |YT ] = P [yT +2 > U
P [yT +2 < L bT +2|T |YT ] = 0.05.

Page 3 of 3

You might also like