Time series

Assignments for homework and seminars

Final score for the course = 0.4 · exam score + 0.3 · control work score +
+ 0.3 · seminars and homework activity score.
The control work and exam will have tasks that very similar to those listed below.

1. Introduction to time series analysis.

AR, MA, ARMA, ARIMA processes.

1. Find on the Rosstat website monthly data on weighted average wage and inflation in the
regions of the Russian Federation, select one region, take the longest available series for
it, find or calculate the following time series:
• weighted average nominal wage,
• inflation rate,
• price level (when end of the first month= 1),
• real wage in prices of the beginning of the period,
• real wage in prices of the end of the period,
• logarithm of real wages,
• growth rate of real wages (in logarithms),
• seasonally adjusted growth rate of real wages in logarithms.

Draw the graphs.

Based on seasonal average growth rates, make a forecasts of growth rates in logarithms
for the next 12 months.
Make the real wages forecast for the next 12 months, cumulatively adding seasonal average
growth rates.
If the inflation forecast is 5% per year (evenly without seasonality), make a forecast of
nominal wages.
You may look thought the xls files and do the same.
Suggest your own way to make such a forecast.
• Present the results of the work as a structured report in Word (what did you do, your
results and brief comments).
• Be ready to present the report at the seminar.

2. • Consider the time series sample Yt = (5, 1, 1, −3, 2, 9, 6, 2, 5, 2)

Calculate mean, variance, autocovariance and autocorrelation of order 1 and 2. Test if
autocorrelations are different from zero.

1
 • Do the same for the time series sample Yt = (6, 6, 1, 6, 0, 6, 6, 4, 3, 2)

3. For the following time series (from the task 1):

• logarithm of real wages,
• growth rate of real wages (in logarithms),
• seasonally adjusted growth rate of real wages in logarithms

Calculate autocorrelation, autocovariance and partial autocorrelation functions, test au-

tocorrelation in these series, discuss the results.

4. • Generate simulations of ARMA processes with different parameters, draw the graphs,
compare and discuss the results, linking the type of series with the processes parameters.

5. • Simulate MA, AR, ARMA processes with different parameters (with orders no more
than 2-3), draw the graphs, calculate theoretical true autocorrelation and autocovariance
functions (for AR and MA processes), calculate their sample values with different series
lengths and test autocorrelation, compare and discuss the results.

6. • Calculate the expected value, variance, autocorrelation function for the following process
MA(3) yt = 1.1.εt−1 + 0.38εt−2 + 0.04εt−3 + εt , εt ∼ N (0, 1)
Check the invertibility of the process. In the case of invertibility calculate the first three
coefficients of the corresponding autoregressive process.
• Calculate the expected value, variance, autocorrelation function for the following AR(2)
process yt = 0.3yt−1 + 0.4yt−2 + εt , εt ∼ N (0, 1)
Check the invertibility of the process. Calculate the first three coefficients of the corre-
sponding moving average model.
• Examine the following process ARMA (2,1) for common roots.
yt = −0.5yt−1 + 0.14yt−2 + 0.7εt−1 + εt , εt ∼ N (0, 1)
Calculate the expected value and variance of the reduced equation.

2. Specification selection and estimation of ARMA models.

1. For some time series the following sample autocorrelation coefficients were calculated:
d1 = 0.7 , cor
cor d3 = −0.2 and sample mean X̄ = 1.7 .
d2 = 0.4, cor
Find estimates of the models coefficients for
1. AR(1);
2. AR(2);
3. AR(3).

2. From the file GDP per capita from 1800 select one of the remaining 9 countries (except
US) or the world as a whole.

2
 • Analyze the stationarity of the GDP per capita indicator (in logarithms) and its first
differences, choose and estimate the AR(p) model, justify the choice meaningfully and
using all necessary tests.
• Present the results of the work as a structured report in Word (what had you do, your
results and brief comments)
• Be ready to demonstrate and present the report at the seminar.

3. From the file GDP per capita from 1800, select one of the remaining 9 countries (ex-
cept US) or the world as a whole. From the "DJA" file, select the Dow Jones Industrial
Average or the Dow Jones Inflation-Adjusted Industrial Average.
• According to the specified algorithm, build an ARMA model, justify the choice mean-
ingfully and using all necessary tests.
• Present the results of the work as a structured report in Word (what did you do, your
results and brief comments).
• Be ready to demonstrate and present the report at the seminar.

3. Forecasting AR(p), MA(k) and ARMA(p, k) processes.

1. For some time series, the MA(2) model was estimated: yt = −0.6εt−1 + 0.3εt−2 + εt , εt ∼
N (0, σ 2 ), σ 2 = 0.5
Let’s assume you have data:

time t t-1 t-2

y -0.5 1.2 -0.9
ε -0.2 0.3 -0.4

Compute the mean forecasts yb for periods t+1, t+2, t+3, as well as 95% confidence in-
tervals. What is the long-term forecast and its confidence interval?

2. For some time series, the ARMA(1,2) model was estimated: yt = 0.6yt−1 + 0.6εt−1 −
0.7εt−2 + εt , εt ∼ N (0, σ 2 ), σ 2 = 0.5
Let’s assume you have data:

time t t-1 t-2

y 0.98 1.2 -0.9
ε -0.2 0.3 -0.4

Compute the mean forecasts yb for periods t+1, t+2, t+3, as well as 95% confidence in-
tervals (to do this you need to find corresponding M A(∞) process parameters). What is
the long-run forecast?

3. For some time series, the ARMA(2,1) model was estimated: yt = 0.6yt−1 − 0.2yt−2 −
0.3εt−1 + 0.4 + εt , εt ∼ N (0, σ 2 ), σ 2 = 0.6
Let’s assume you have data:

3
 time t t-1 t-2
y 1.01 1.2 -0.9
ε -0.2 0.3 -0.4

Compute the mean forecasts yb for periods t+1, t+2, t+3, as well as 95% confidence in-
tervals (to do this you need to find corresponding M A(∞) process parameters). What is
the long-run forecast?

4. • From DJA file, select the Dow Jones Industrial Average or Dow Jones Industrial
Average, adjusted for inflation or data on real GDP per capita from the file GDP per
capita from 1800.
• Build 2-3 AR, MA, ARMA models for these variables or their first differences, justifying
the model choice meaningfully and using all the necessary tests.
• Based on the estimated models, calculate forecasts and their confidence intervals for 1
and 2 period without using R. To build confidence intervals for the forecasts based on AR
and ARMA models, you can use the parameters of the estimated MA process.
• Build interval forecasts of growth rates of variables and initial variables for 10-100 steps
using R.
• Identify the differences between alternative specifications of AR, MA and ARMA models
in terms of the resulting forecasts (you can plot the graphs in Excel or otherwise)
• For one of the specifications, build one imitation forecast and show it on the graph.
• The results of the work should be presented as a structured report in Word
Be ready to present it at the seminar.

4. Modeling volatility. ARCH and GARCH models.

1. Generate a series of length 10,000 observations according to the ARCH(4) model using
the equation:
εt = σt ϵt ; ϵt ∼ N (0, 1); σt2 = 1 + 0.3ε2t−1 + 0.25ε2t−2 + 0.15ε2t−3 + 0.1ε2t−4
Estimate the ARCH(4) model for the generated series. Compare the estimates with the
true model parameters.
Estimate the unconditional variance and the mean of the series, the coefficient of asym-
metry. Compare with their true values.
Repeat for different lengths of the generated series. Make some conclusions.

Do the same for the GARCH(1,1) model. Select the model parameters by yourself.

2. The coefficient of the ARCH(1) process is α1 = 0.8 , the unconditional variance is 2. Write
down the process equation.

3. Compute the unconditional variance of the following GARCH process:

σt2 = 0.4 + 0.1σt−1
2
+ 0.7ε2t−1

4. Prove that the following GARCH process cannot have finite unconditional variance: σt2 =

4
 2
1 + 0.6σt−1 + 0.6ε2t−1 (Hint: you can use proof by contradiction.)

5. The GARCH(1,1) process is given by the equation σt2 = 1 + 0.8σt−1

2
+ 0.1ε2t−1 , it is known
that σT = 9, εT = −2. What are the volatility forecasts for the next 3 periods? What is
2

the long-term volatility forecast?

6. Consider the following model with autoregressive conditional heteroskedasticity:

2
σt2 = α0 + βσt−1 + α(|εt−1 | − γεt−1 )2 .
Provide the parameter values at which this model reduces to the regular GARCH model.
It is argued that in this model there is an asymmetry of “shock’s” εt impact on the con-
ditional variance. What does it mean? What parameter values should be if the impact is
symmetrical?

7. • Data: take price of one of Russian shares from Russtockdaily file or from DJA
file select Dow Jones Industrial Average or Dow Jones Industrial Average, adjusted for
inflation.
• For the selected series, select an ARMA model and compare various specifications of
volatility models (GARCH, IGARCH, EGARCH, TGARCH, GARCH-M, total number
of parameters no more than 4-6) based on: absence of autocorrelation in residuals and
squared residuals, information criteria, number of parameters.
• Select the two best volatility model specifications.
• Based on the estimated models, construct forecasts of the initial indicator and forecasts
of volatility (standard deviation of the models errors)
• Construct one imitation forecasts of the series and its’ volatility. When constructing
imitation forecasts, carefully read Introduction to the rugarch package (there may
be difficulties).
• Additionally, based on many imitation forecasts, calculate their average, 95% confidence
intervals.
• Present the results of the work as a structured report in Word. Be ready to present the
report at the seminar.

5. Nonlinearity in mean: structural breaks.

1. Generate one series from two ARMA processes with different parameters “glued" with
each other at the point of the structural break.
Perform the procedure for searching for point of structural break.
If process parameters slightly differ in two parts of the series, is it possible to determine
the break point?
How does the length of the series and error standard deviation affect the result? Perform
some experiments and answer these questions, present the results with the appropriate
graphs and all necessary information.

2. For US inflation series (growth rate of the price level in USA, file DJA) and AR(1),
AR(4), AR(5) ARMA (1,1) and ARMA(2,1) models, determine structural break point (if
any). Present your results, draw conclusions.

5
 3. Generate one series from two AR processes with a difference in one parameter “glued"
which each other at point of structural break.
Perform full and partial structural break testing procedure. Perform series of experiments
and draw conclusions, document the results, providing appropriate graphs and all neces-
sary information.

4. Generate one series “glued" from several AR processes with different parameters.
Perform multiple structural breaks testing procedure.
Perform series of experiments and draw conclusions, document the results, providing ap-
propriate graphs and all necessary information.

5. Generate one series “glued" from several AR processes with different parameters.
Perform the multiple structural break testing procedure using the Bai-Perron test.
Perform series of experiments and draw conclusions, document the results, providing ap-
propriate graphs and all necessary information.

6. Take data on price of one share (or index) from the files DJA, Rustockdaily, WorldInd-
Daily.
• Estimate ARIMA model. It is better to use AR representation to further search for
structural breaks.
• Examine the selected time series for the presence of structural breaks using Sup-F, Ave-
F, Exp-F tests, CUSUM, MOSUM tests, RE, ME tests, test for the presence of partial
structural breaks.
• Investigate possibility of multiple structural breaks using the Sup-F test and the Bai-
Perron test.
• When looking for structural breaks, adjust the series for changes in volatility if neces-
sary.
• Construct forecasts, one based on ordinary ARMA-GARCH and another on ARMA-
GARCH taking into account structural breaks (if they were found). Discuss the difference
in the forecasts obtained.
• Present the results with appropriate graphs and all necessary information.

6. Multivariate time series: vector autoregression (VAR)

model
1. Let’s assume we estimated a VAR(1) model for two variables and got the following esti-
mates:
        
yt,1 0.2 0.4 yt−1,1 εt,1 1 −0.5
= + ; Cε =
yt,2 0.2 0 yt−1,2 εt,2 −0.5 4.25
Find the structural form of the model (coefficient matrices and diagonal covariance ma-
trix of the SVAR model residuals) if it is known that it is recursive (the second variable

6
 without any lag enters the equation for the first, but not vice versa).

2. • Take data from files DJIA_1919 (monthly data on Dow Jones index, industrial pro-
duction index (IPI), and consumer price level (CPI)) or MSCIRUS_2020 (daily data,
description in the file).
• Select one variable of interest.
• Study some factors influencing this variable using CCF and Granger causality test. But
do not analyze functionally related variables (for example, American stock market indices
DJIA, SP500, Wilshire).
• Build optimal VAR model. Use step-by-step inclusion of variables; justify feasibility of
including variables using F-test. If possible, place restrictions on significance of model
parameters and test for auto- and crosscorrelation of residuals (Hosking, LiMcLeod).
• Construct interval forecast of the variable of interest based on the optimal VAR model
and compare result with interval forecast based on ARIMA.
• Construct IRFs and FEVD for the selected variable and give interpretation.
• Present the results of the work in the form of a structured report in Word.

7. Analysis of non-stationary series.

Cointegration and vector error correction model (VECM)
1. • Generate several independent random walk processes and check whether the statement
is true: “Regressions with independent random walk processes are likely to have a high
coefficient of determination R2 due to nonstationarity. In more than half of the cases, the
coefficient of determination exceeds 20%, and in about 5% of cases it exceeds 70%.”
• Estimate the dependence of random walk processes on the time trend.
• For comparison, estimate similar regressions for independent normally distributed pro-
cesses such as white noise, and draw conclusions.

2. • From files with weighted average interest rates of Russian banks, take 3 or 4 series of
any interest rates.
• Another option is to take the GDP per capita of 3–4 countries from the files on the GDP
of countries.
• Check the series for cointegration using the Engle-Granger method. Take different series
as the first variable and compare the results.
• Using Johansen’s methodology, test series for cointegration, find the number of cointe-
gration relations, estimate VECM (p-1) in two ways, build a forecast, calculate VAR(p)
parameters for the original series, build a forecast with confidence intervals.
• Present all the results and try to interpret them.