Time-series analysis — AR, MA, unit roots, cointegration
In this chapter: Autoregressive (AR) models · Moving-average (MA) and ARMA · Unit-root tests, stationarity · Cointegration
Time-series questions on L2 typically present AR/AR(p) output and ask about model adequacy, forecast, unit roots. The art: recognising when to use levels vs first-differences.
Stationarity: mean, variance, covariance with own lag are constant over time. OLS valid only on stationary series. AR(1): Yt = b0 + b1Yt-1 + εt. - |b1| < 1: stationary (mean-reverting). - |b1| = 1: unit root, non-stationary (random walk). - |b1| > 1: explosive, rare. Unit-root test: Augmented Dickey-Fuller (ADF). H0: unit root. Reject H0 → stationary. If two non-stationary series are cointegrated (linear combination is stationary), regression in levels is valid; if not, regression is spurious — use first differences.
Workflow: 1. Plot the series. Trend? Seasonality? Variance change? 2. Test for stationarity (ADF). 3. If non-stationary: difference until stationary, OR test for cointegration with another non-stationary series. 4. Fit AR(p): use partial autocorrelation function (PACF) to choose p. 5. Check residuals: serial correlation? Heteroskedasticity? If autocorrelation persists, increase p. 6. Choose between models using AIC/BIC (lower better). Cointegration test: Engle-Granger. Run regression Y on X (both I(1)); test residuals for unit root. If residuals stationary, cointegrated.
CFA L2 trap: spurious regression. Two unrelated random walks regressed on each other can show high R² and significant coefficients — but it's nonsense. Track record: if stuck on a vignette and series is "asset price" or "GDP," default to non-stationary. Use first differences (returns, growth rates). Seasonality: monthly retail sales has seasonal AR — include lagged 12-month dummy or seasonal AR(12) term.
- CFA Institute Quant curriculum
- Regressing levels of two non-stationary series — spurious regression.
- Ignoring residual autocorrelation — invalidates t-stats.
- Treating high R² as proof of good model in time-series — common with trending data.
Frequently asked
How do I know if a series is stationary?
When can I use levels in regression?
Practice questions
Click each question to reveal the answer and explanation.
Q 1An AR(1) model has b1 = 1.0. The series is:- (a)Stationary
- (b)Has unit root (non-stationary)
- (c)Explosive
- (d)Mean-reverting
- (a)Stationary
- (b)Has unit root (non-stationary)
- (c)Explosive
- (d)Mean-reverting
Q 2Augmented Dickey-Fuller H0:- (a)Series is stationary
- (b)Unit root present
- (c)Cointegration
- (d)Heteroskedasticity
- (a)Series is stationary
- (b)Unit root present
- (c)Cointegration
- (d)Heteroskedasticity
Q 3Two I(1) series regressed in levels yield high R². You should:- (a)Accept results
- (b)Test for cointegration before trusting
- (c)Add more variables
- (d)Use shorter sample
- (a)Accept results
- (b)Test for cointegration before trusting
- (c)Add more variables
- (d)Use shorter sample
Q 4AR(1) Yt = 0.4 + 0.6Yt-1. Long-run mean:- (a)0.4
- (b)0.6
- (c)1.0
- (d)0.67
- (a)0.4
- (b)0.6
- (c)1.0
- (d)0.67
Q 5PACF is used to determine:- (a)Variance
- (b)Order p of AR model
- (c)Heteroskedasticity
- (d)Stationarity
- (a)Variance
- (b)Order p of AR model
- (c)Heteroskedasticity
- (d)Stationarity