| Model Type | Equation | | ----------------------- | ---------------------------------------- | | Linear trend | $X_t= \beta_1+\beta_2t+W_t$ | | Internal regressor (AR) | $X_t = \phi_1+X_{t-1}+\phi_2X_{t-2}+W_t$ | | External regressor | $X_t = \beta_1X_{t-1}+\beta_2Y_s+W_t$ | ## Cross-Variance If we decide to include external regressors, then we can look at the cross-covariance, i.e. $\mathrm{cov}(X_t, Y_s)$. Also cross-covariance needs to fulfill the [[Stationarity#Weak Stationarity|Weak Stationarity]] condition. It must be only a function of the gap $h$ between $s,t$ and not a function of absolute time. $ \begin{align} \gamma_{XY}(t,s)&=\mathrm{cov}(X_t, Y_s) \\[4pt] \gamma_{XY}(h)&=\mathrm{cov}(X_t, Y_{t+h}) \end{align} $ To estimate cross-covariance, we again average over all available lagged pairs along both time series. $ \hat\gamma_{XY}(h)=\frac{1}{n}\sum_{t=1}^{n-h} (x_{t+h}- \bar x)(y_t - \bar y) $ Cross-correlation factor (CCF): $ \rho_{XY}(h)=\frac{\gamma_{XY}(h)}{\sigma_X\sigma_Y} $ ## Procedure for Time Series Model Fitting *Step 1: Transformations to make time series stationary* - Log-transform - Remove trends & seasonality - Differentiate successively *Step 2: Check if transformed series is white-noise* - If yes, then we are done. - If not look at ACF and decide if remainder is [[Autoregressive Model|Autoregressive Model]] or a [[Moving Average Model|Moving Average Model]] (this is dependent on how quickly ACF drops). *Step 3:* $\text{AR}(p)$ - Compute [[Partial Autocorrelation Function|PACF]] to get order (i.e. number of useful lags). - Estimate $\phi_k$ coefficients and $\sigma_W^2$ of white-noise r.v. $W_i$ via [[Yule-Walker Equations]]. - Compute residuals and test them for white-noise. *Step 4:* $\text{MA}(q)$: - Compute ACF to get order(i.e. number of useful lags). - Estimate $\theta_k$ coefficients and $\sigma_W^2$ of white-noise r.v. $W_i$ via [[Maximum Likelihood Estimation]]. - Compute residuals and test them for white-noise. *Step 5:* $\text{ARMA}(p,q)$: - Fit an AR(p) model, and compute residuals. - Fit an MA(q) model on the residuals (or original data). - Fit an ARMA(p, q) using the $p,q$ parameters that have been obtained in the previous steps. - Compute residuals and test them for white-noise.