Trend models
- Linear trend: yt = b0 + b1t. Grows by a constant amount each period.
- Log-linear trend: ln yt = b0 + b1t. Grows at a constant rate; suits series such as sales or prices.
If residuals from a trend model are serially correlated (Durbin-Watson far from 2), the trend model is inadequate and an autoregressive model may fit better.
Autoregressive models
xt = b0 + b1xt−1 + b2xt−2 + … + bpxt−p + εt
AR models require covariance stationarity: constant mean, constant variance, constant autocovariance at each lag.
Check the model: after estimating, test whether residual autocorrelations are significant (t = autocorrelation ÷ (1/√T)). If any are, add lags.
Mean reversion: an AR(1) series reverts to b0 ÷ (1 − b1). If xt is above that level, the model forecasts a decline. Watch it below:
Multi-period forecasts use the chain rule: forecast xt+1, then use it to forecast xt+2. Uncertainty grows with the horizon.
Unit roots and random walks
If b1 = 1, the series is a random walk: xt = xt−1 + εt (with drift if b0 ≠ 0). It has no mean-reverting level and infinite variance over time; it is not stationary.
Dickey-Fuller test: regress xt − xt−1 on xt−1; test whether the coefficient (b1 − 1) is zero, using special critical values. Failing to reject means a unit root. Fix: model the first differences.
Seasonality
A significant residual autocorrelation at the seasonal lag (4 for quarterly, 12 for monthly) signals seasonality. Add that lag to the model.
ARCH
Autoregressive conditional heteroskedasticity: error variance depends on past squared errors. Test by regressing squared residuals on lagged squared residuals; a significant coefficient means ARCH. Then standard errors are wrong, but the ARCH model can forecast variance. Use GARCH-type models in that case.
Regressions with two time series
| Dependent | Independent | OK to regress? |
|---|---|---|
| Stationary | Stationary | Yes |
| Unit root | Stationary (or vice versa) | No |
| Unit root | Unit root, not cointegrated | No (spurious) |
| Unit root | Unit root, cointegrated | Yes, captures the long-run relationship |
Test cointegration with the Engle-Granger method: regress one on the other, then test the residuals for a unit root.
Exam focus
Choose trend or AR models, compute mean-reverting levels and forecasts, diagnose unit roots, seasonality and ARCH, and decide when a regression between two time series is valid.