I'm having a problem calculating the best linear predictor of a time series. I'm using the book Brockwell-Davis 2016 - Introduction to Time Series and Forecasts. First let me write down one notational convention, one proposition and one prerequisite problem:

  1. The best linear predictor in terms of $1,X_n,...,X_1$ is denoted by $P_nX_{n+h}$ and clearly has the form $P_nX_{n+h}=a_0+a_1X_n+...+a_nX_1$
  2. Let $(X_t, t\in\mathbb{Z})$ be a timeseries with $\text{Var}(X_t)<\infty$ for all $t\in \mathbb{Z}$ and $X^n := (X_{t_1},...,X_{t_n})$ a collection of random variables of the time series at $n$ different times. Then the best linear predictor of $X_t$ is given $P_nX_{n+h}$ as above in point 1. The coefficients $a_0,...,a_n$ are determined by the linear equations \begin{align} \mathbb{E}(X_t-P_nX_{n+h})&=0\\ \mathbb{E}(X_{t_j}(X_t-P_nX_{n+h})) &= 0, \quad \text{for all} \quad j=1,...,n. \tag1 \end{align}

The needed problem: Show that the process $X_t=A\cos(\omega t)+B\sin(\omega t), \ t=0,\pm1,...$ where $A$ and $B$ are uncorrelated random variables with mean 0 and variance 1 and $\omega$ a fixed frequency in $[0,\pi]$, is stationary and find its mean and autocovariance function.

Solving the above I came up with the answers $\mu_X=0$ and $\gamma_X(h)=\cos(\omega h)$ which are verified correct. Now the task is the following:

Let $\{X_t\}$ be the process defined in the problem above. Find $P_1X_2.$


According to the above we have $P_1X_2 = a_0+a_1X_1 = a_0+a_1A\cos(\omega)+a_1B\sin(\omega)$. The first equation in $(1)$ gives

\begin{align} \mathbb{E}[X_2-P_1X_2] &= \mathbb{E}[X_2-a_0-a_1A\cos(2\omega)-a_1B\sin(2\omega)]\\ &=\mathbb{E}[X_2]-a_0=0+a_0 = 0 \Longleftrightarrow a_0=0. \end{align}

The second equation in $(1)$ gives

\begin{align} \mathbb{E}[X_1(X_2-P_1X_2)] &= \mathbb{E}[X_1X_2] - \mathbb{E}[(X_1(a_0 + a_1A\cos(2\omega)+a_1B\sin(2\omega)))] \\ &= \mathbb{E}[2A^2\cos^3(\omega)-A^2\cos(\omega)+2B^2\cos(\omega)-2B^2\cos^3(\omega)]\\ &-a_1\mathbb{E}[2B^2\cos^3(\omega)-A^2\cos(\omega)+2B^2\cos(\omega)-2B^2cos(\omega)]\\ &= \mathbb{E}[(2B^2-A^2)\cos(\omega)]-a_1\mathbb{E}[(2B^2-A^2)\cos(\omega)]\\ &= \cos(\omega) - a_1\cos(\omega) = 0\Longleftrightarrow a_1 = 1. \end{align}

after plugging in $X_1$ and using that $\mathbb{E}[A^2]=\text{Var}[A]=1 =\mathbb{E}[B^2]=\text{Var}[B]$ as well as some trigonometric identities. Thus the best linear predictor of $X_2$ based on $1, X_1$ is $P_1X_2 = a_0 + a_1X_1 = X_1.$ Am I doing this correctly?


1 Answer 1


So the question is asking: Let $X_t$ be the process defined in the problem above. Find $P_1X_2$.

$P_nX_{n+h}$ is the form of the linear predictor, I don't quite understand why you are assuming that they are asking you to predict $X_2$, as you are assuming that $n+h = t$.

They themselves state that $t = 0$ in the question statement above in the equation:

$X_t=A\cos(\omega t)+B\sin(\omega t), \ t=0,\pm1,...$

Let me know if this helps. I believe you may be incorrectly evaluating $X_t$ as $X_2$.

I'm not an expert and could very well be wrong however.

  • $\begingroup$ Because $P_nX_{n+h}$ is the best linear predictor of $X_{n+h}$ in terms of $1, X_1,...,X_n$. So in my case $n=1$ and $h=2$ so we get $P_1X_2$ which is the best linear predictor of $X_2$ given $1, X_1$. $\endgroup$
    – Parseval
    Commented May 24, 2021 at 22:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.