Author: Eiko

Tags: probability theory, time series, stochastic process, stationary, stationarity

Time: 2025-01-08 10:00:09 - 2025-01-08 10:00:09 (UTC)

Basic Concepts

Let \(\{X_t, t\in T\}\) be a stochastic process / time series. We have some concepts of covariance matrices to understand the dependence between them.

  • When \(\mathrm{Var}(X_t)<\infty\), we can define auto-covariance function

    \[\gamma_X(r,s) := \mathrm{Cov}(X_r, X_s) = \mathbb{E}[(X_r - \mathbb{E}X_r)(X_s - \mathbb{E}X_s)], \quad r,s\in T.\]

  • \(X_t\) is called stationary or weakly stationary / covariance stationary / second order stationary, if

    • The covariance is time-homogeneous, i.e. \(\gamma_X(r,s) = \gamma_X(r+h, s+h)\)

    • The expectation is also time-homogeneous, i.e. \(\mathbb{E}X_t = \mu\) for all \(t\in T\).

    • Additionally, all \(X_t\) are in \(L^2\).

  • When we are in the stationary setting, the covariance \(\gamma_X(r,s)\) can be reduced to a single parameter function \(\gamma_X(r-s)\), so

    \[\gamma_X(h) = \mathrm{Cov}(X_{t+h}, X_t).\]

    And the auto-correlation is then

    \[\begin{align*} \rho_X(h) &:= \frac{\mathrm{Cov}(X_{t+h}, X_t)}{\sqrt{\mathrm{Var}(X_{t+h})\mathrm{Var}(X_t)}} \\ &= \frac{\gamma_X(h)}{\sqrt{\gamma_X(0)\gamma_X(0)}} \\ &= \frac{\gamma_X(h)}{\gamma_X(0)}. \end{align*}\]

  • \(X_t\) is said to be strictly stationary if the joint distribution of \((X_{t_1}, \ldots, X_{t_k})\) is the same as \((X_{t_1+h}, \ldots, X_{t_k+h})\) for all \(t_1, \ldots, t_k, h\in T\).