You are on page 1of 2

Time series analysis and applications

Homework

Please return (in PDF file) by May 24, 2020

Notes:
- Stationary will mean weakly stationary.
- Time series names in blue refer to data provided in the book. To
access those time series (and related R functions) you need to
install the “astsa” package (install.packages("astsa"))
once and load it whenever you need it (library(“astsa”)).
- Questions (Q1, Q2, …) are graded as equivalent.

Q1.Consider the random walk with drift (δ ≠ 0model x t=δ + x t − 1+ wt , for t=1,2 , ⋯,
with x 0=0 , where w t is (zero-mean) white noise with variance σ 2w.
t
(a) Show that the model can be written as x t=δt + ∑ w k , t=0,1,2 , ⋯.
k=1
(b) Find the mean and autocovariance functions of x t .
(c) Argue that x t is not stationary.
t −1
(d) Show that its normalized autocovariance satisfies ρ x ( t − 1, t )=
Comment on this result.
√ t
.

(e) Suggest a transformation to make the series stationary. Prove it.


(f) What if δ =0?
(g) Could you view the given time series model as something different from
random walk? Comment on the properties of such an alternative model.

Q2.The variogram of a time series x t is defined as the function


1
V x ( h ) = E [ ( x t+h − x t ) 2 ]
2
of the lag h=0 , ±1 , ±2 , ⋯.
(a) Show that, if x t is stationary with autocovariance function γ x, then
V x ( h ) =γ x ( 0 ) − γ x ( h ).
(b) Express the result of (a) in terms of the normalized autocovariance, ρ x.
Comment on the values taken by the variogram for small and large lags.
(c) Compute the variogram of the process of Question 1. Comment on its
dependence on h and compare with (b).
(d) Plot the variogram of (c) with δ =0 and δ =0.5, for h=0,1 , ⋯ , 20.
Comment on the appearance of the plot.

Q3.The glacial varve time series, x t , exhibits some nonstationarity that can be
improved by transforming to logarithms and some additional nonstationarity
that can be corrected by differencing the logarithms.

1
(a) Compare the sample variance of x t over the first half and the second half
of the data. If you see a significant difference, this is called
heteroscedasticity.
(b) Check whether the transformation y t =log x tstabilizes somehow the
variance over the series.
(c) Plot the series x t , y t . Plot the histograms of x t and y t . Do they look close
to normal (Gaussian)?
(d) Plot and compare the sample normalized autocovariances of x t and y t .
(e) Compute the difference ut = y t − y t −1, examine its time plot and sample
normalized autocovariance, and argue that differencing the logged data
produces a reasonably stationary series. Can you think of a practical
interpretation for ut (assuming x t is close to x t −1 )?

Q4.Consider the time series x t=ϕ xt −1 +w t, with ϕ=0.9 and w t (zero-mean) white
noise with variance σ 2w =1.
(a) Is this model causal? Is it invertible? Justify your answers.
(b) Generate a realization of x t of length n=100. Using an estimation
procedure of your choice, estimate the parameters ϕ and σ 2w. Use the
estimated model to predict the future 4 values of the series.
(c) Repeat for a realization of length n=300. Compare with (b).
(d) Derive an expression for the normalized autocovariance function of x t .
Plot this function along with its sample estimate.
(e) Plot the normalized partial autocovariance function of x t . Compare with
(d).

Q5.The sunspotz time series is the biyearly (two measurements per year) sequence
of numbers of sunspots from June 1749 to December 1978, with n=459
samples in total. Perform a spectral analysis of this time series to identify its
predominant periods (say, there are two of them), using:
(a) Raw periodogram
(b) Smoothed periodogram
(c) Periodogram with tapering
(d) Parametric spectral estimation based on an AR model of appropriate order.
Compare the results of the various methods.

You might also like