Mestrado de Matemática Atuarial

2024/2025

Time Series - Problems

Characteristics of Time Series

1. Show that if xs and xt are bivariate normal and γx (s, t) = 0 ∀s, t, then xs and xt are independent.
Recall that bivariate
 normally distributed variableshave jointdensity function:
  −1 x1 − µ1
exp −1/2 x1 − µ1 , x2 − µ2 Σ
x2 − µ2
f (xs , xt ) = p
2π |Σ|
where
σ12

cov(x1 , x2 )
Σ= is the covariance matrix.
cov(x1 , x2 ) σ22

2. Consider time series vt defined as vt = 1/3(wt−1 + wt + wt+1 ), where

wt is white 2 . Show that
noise with variance σw
3/9σw2 s=t


2/9σ 2 |s − t| = 1

w
γv (s, t) = 2
1/9σ w |s − t| = 2




0 |s − t| > 2
Pt
3. Consider random walk time series xt defined as xt = j=1 wj , where
2 . Show that
wt is white noise with variance σw
2.
γx (s, t) = min(s, t)σw

4. Show that the autocovariance function can be written as

γ(s, t) =E [(xs − µs )(xt − µt ) ]= E(xx xt ) - µs µt

5. Consider time series xt defined as xt = st + wt , where

wt is Gaussian white noise with σw2 = 1 and

0 t = 1, ...100
st = (t−100) .
10exp{− 20 }cos(2π(t − 100)/4) t = 101, ..., 200

Simulate and plot 200 observations from the model.

6. For the series of problem 5:

a) Compute the mean function µx (t) for t = 1, ...200

b) Compute the autocovariance function γx (s, t) for s, t = 1, ...200

7. For a moving average process of the form

xt = wt−1 + 2wt + wt+1
2 , determine the autocovariance and auto-
where wt are independent with zero means and variance σw
correlation functions as functions of lag h = s − t.

8. Consider the random walk with drift model

2.
xt = δ + xt−1 + wt with x0 = 0 and wt white noise with variance σw

1
 Pt
a) Show that the model can be written as x(t) = δt + k=1 wk .
b) Define the mean and autocovariance functions of xt
c) Justify why xt is not stationary.
d) Consider the following transformation x0t of xt : x0t = xt − xt−1 .
Prove the stationarity of the transformed series.

9. Let wt for t = 0, ±1, ±2, ... be a white noise process and consider
xt = wt wt−1 .

a) Define the mean and autocovariance functions of xt .

b) Is xt stationary?

10. Simulate a series of n = 500 Gaussian white noise observations.

a) Compute the sample ACF ρ̂(h) to lag 20.

b) Compare ρ̂(h) with ρ(h)
c) Repeat a) using n = 50. How does changing n affect the results?

11. Simulate a series of n = 500 moving average observations as in Problem 2.

a) Compute the sample ACF ρ̂(h) to lag 20.

b) Compare ρ̂(h) with ρ(h)
c) Repeat a) using n = 50. How does changing n affect the results?

12. For the model in Problem 5:

a) Show that the model is not stationary.

b) The sample ACF can be informative. Calculate, plot and comment the sample ACF.

2
 ARIMA Models

1 For an M A(1), xt = wt + θwt−1 , define the values of θ for which ρx (1) attain its maximum and
minimum.Show that |ρx (1)| < 1/2 for any number θ.
2 and let |φ| < 1 be a constant.
2 Let {wt ; t = 0, 1, ...} be a white noise process with variance σw
Consider the process x0 = w0 and xt = φxt−1 + wt , t = 1, 2, ....

a) Show that xt = tj=0 φj wt−j for any t.

P

b) Find E(xt ) and var(xt ).

c) Show that for h ≥ 0 , cov(xt+h , xt ) = φh var(xt ).
d) Is xt stationary?
e) Show that xt is stationary as t → ∞.
p
f) Suppose x0 = w0 1 − φ2 . Is this process stationary?

3 Identify the following ARM A(p, q) models and determine whether they are causal and/or invertible:

a) xt = 0.8xt−1 − 0.15xt−2 + wt − 0.3wt−1

b) xt = xt−1 − 0.5xt−2 + wt − wt−1

4 Consider each of the following models:

(m1) xt = xt−1 + wt − 1.5wt−1

(m2) xt = 0.8xt−1 + wt − 0.5wt−1
(m3) xt = 0.6xt−1 + wt − 1.2wt−1 + 0.2wt−2

a) Verify whether it is causal and/or invertible.

b) Express the model in an MA representation if it exists.
c) Express the model in an AR representation if it exists.

5 Verify the following expressions for an ARM A(1, 1):

2
2 1+2θφ+θ , γ(1) = σ 2 (1+θφ)(φ+θ)
a) γ(0) = σw 1−φ2 w 1−φ2
2 (1+θφ)(φ+θ) φh−1
b) for h ≥ 1, γ(h) = σw 1−φ2
(1+θφ)(φ+θ) h−1
c) for h ≥ 1, ρ(h) = 1+2θφ+θ2
φ

6 Plot the ACFs of the ARM A(1, 1), ARM A(1, 0) and ARM A(0, 1) on the same graph for Φ = 0.6
and θ = 0.9. Comment on the diagnostic capabilities of the ACF on this case.

7 Generate n = 100 observations from each of the three models discussed in Problem 6. Compute the
sample ACF and PACF for each model.

8 Fit an AR(2) model to the cardiovascular series (cmort series) using least squares and using Yule-
Walker.

9 Generate n = 500 observations from the ARMA model given by

xt = 0.9xt−1 + wt − 0.9wt−1
with wt ∼ iidN (0, 1). Plot the simulated data, compute the sample ACF and PACF of the simulated
data.

10 Generate 10 realizations of length n = 200 each of an ARM A(1, 1) process with φ = 0.9, θ = 0.5
and σ 2 = 1. Find the MLEs of the three parameters in each case.

3
11 Consider the stationary series generated by
xt = α + φxt−1 + wt + θwt−1
where
E(xt ) = µ, |θ| < 1, |φ| < 1 and the wt are iid random variables with zero mean and variance σ 2 .

a) Determine the mean as a function of α for the above model.

b) Consider yt = xt − µ. Show that the process is weakly stationary.
c) Find the autocovariance and ACF of the process xt .

12 Let yt = β0 + β1 t + ... + βq tq + xt , βq 6= 0 where xt is stationary. First, show that ∇k xt is stationary

for any k = 1, 2, ... and then show that ∇k yt is not stationary for k < q, but is stationary for k ≥ q.

13 Consider the IM A(1, 1): xt = xt−1 + wt − λwt−1Pwhere |λ| < 1 for t = 1, 2, .. and x0 = 0.Verify
that the model can be inverted and written as xt = ∞
j=1 (1 − λ)λ
j−1 x
t−j + wt .

14 Analyse the diagnostics of the AR(2) fit of the GNP growth rate series (gnp series).

15 Fit an ARIM A(p, d, q) to the growth rate of oil crude prices series (oil series). Comment the results.
Solution: ARMA(1,1) for diff(log(oil))

16 Fit an ARIM A(p, d, q) to the global temperature data (globtemp series). Comment the results.
Solution: ARIMA(1,1,1)

17 Fit an ARIM A(p, d, q) to the sulfur dioxide series (so2 series). Comment the results.
Solution: ARIMA(0,1,1)

18 Let x and y be jointly distributed random variables with density function f (x, y) and
M SE = E[(y − g(x))2 ].

a) Show that MSE is minimized by the choice g(x) = E(y|x).

Hint : M SE = EE[((y − g(x))2 |x)].
b) Apply the above result to the model y = x2 + z where x and z are independent normal variables
with zero-mean and variance one. Show that M SE = 1.

19 Let xt represent the cardiovascular mortality series (cmort series).

a) Fit an AR(2) to xt using linear regression.

b) Based on the model defined in a), find the forecasts xnn+m for m = 1, 2, 3, 4 and the correspon-
ding 95% prediction interval.
c) Fit and AR(2) to xt using the sarima command. Relate the estimated E(xt ) and the estimated
intercept defined in a).
d) Use the sarima.f or command to define the forecasts xnn+m for m = 1, 2, 3, 4 and the corres-
ponding 95% prediction interval.

20 Using the model defined in ex. 14, forecast the global temperature for the next 10 years. Define limits.
Comment.

21 Using the model defined in ex. 15, forecast the sulfur dioxide series (so2 series) into four periods
ahead and calculate 95% prediction intervals for each forecast.

22 Fit a seasonal ARIMA model to the chicken series (chicken series). Use the estimated model to
forecast the next 12 months.
Solution: SARIM A(2, 1, 0) × (0, 1, 1)12

23 Fit a seasonal ARIMA model to the unemployment series (unemp series). Use the estimated model to
forecast the next 12 months.
Solution: SARIM A(2, 1, 0) × (0, 1, 1)12

4
24 Fit a seasonal ARIMA model to the U.S. live birth series (birth series). Use the estimated model to
forecast the next 12 months.
Solution: SARIM A(1, 1, 1) × (0, 1, 1)12