Study guides / CFA L2: CFA Program Level II / Time-series analysis
Time-series analysis
CFA L2: CFA Program Level II · quick review
Modeling a variable from its own past: linear and log-linear trends, autoregressive models and their forecasts, covariance stationarity, random walks and unit roots, seasonality, ARCH, and regressions that combine two time series.
The rules
- Trend models: a linear trend adds a constant amount each period. A log-linear trend grows at a constant rate, so use it when the series compounds (sales, prices) and linear-trend residuals curve away. To forecast from a log-linear trend, compute b0 + b1 × t, then raise e to that power. If trend residuals are serially correlated (Durbin–Watson works here), the trend model is inadequate; try an autoregressive model.
- Covariance stationarity: a constant and finite mean, a constant and finite variance, and a constant and finite covariance between values a given number of periods apart. AR models need it; on a nonstationary series, estimates and tests are invalid and regressions can be spurious.
- AR(p) models: x(t) is regressed on its own p previous values. Forecast one period ahead from the latest observations; for two periods ahead, feed the one-period forecast back in (the chain rule of forecasting). Uncertainty grows with each step.
- Checking an AR model: Durbin–Watson isn't valid when the regressors are lagged values of the dependent variable. Instead, test each residual autocorrelation with t = autocorrelation ÷ (1 ÷ √T), where T is the number of observations. If none is significant, the model fits; if some are, add lags.
- Mean reversion: an AR(1) series reverts to b0 ÷ (1 − b1), which requires |b1| < 1. Above that level the model forecasts a fall, below it a rise, and at it no change. Every covariance-stationary series has a finite mean-reverting level.
- Comparing models: in-sample fit can mislead, so compare out-of-sample root mean squared error (RMSE); the model with the lower RMSE forecasts better. Coefficients are often unstable across periods because of regime changes (policy, regulation, technology). A longer sample adds data but risks mixing regimes; the sample should come from one stationary regime.
- Random walk: x(t) = x(t − 1) + ε(t), or with drift if an intercept is added. It has b1 = 1, a unit root: its mean-reverting level is undefined and its variance grows without limit, so it isn't covariance stationary. Many asset prices and exchange rates behave this way. First-differencing, y(t) = x(t) − x(t − 1), makes it stationary, and the differenced series can be modeled with an AR model.
- Unit root test (Dickey–Fuller): an ordinary t-test of b1 = 1 is invalid when there's a unit root. Instead regress x(t) − x(t − 1) on x(t − 1); the slope is g1 = b1 − 1. Test H0: g1 = 0 (unit root) against g1 < 0 using Dickey–Fuller critical values, which are larger in absolute value than standard t values. Failing to reject means difference the series before modeling.
- Seasonality: a significant residual autocorrelation at the seasonal lag (lag 4 for quarterly data, lag 12 for monthly) reveals it. Add that lag to the model; a forecast then needs the value from one year earlier.
- ARCH: error variance depends on past squared errors. Test ARCH(1) by regressing squared residuals on their own first lag; a significant a1 means ARCH. Standard errors are then unreliable, so use generalized least squares; the fitted ARCH model also forecasts next period's variance.
- Regressing one time series on another: valid if both are covariance stationary. If only one has a unit root, it's invalid. If both have unit roots, it's valid only if they're cointegrated (they share a long-run relation). Test with Engle–Granger: run the regression, then apply a Dickey–Fuller test to its residuals using Engle–Granger critical values; rejecting a unit root in the residuals means cointegration.
Formulas
- Linear trend: y(t) = b0 + b1 × t
- Log-linear trend: ln y(t) = b0 + b1 × t, so y(t) = e^(b0 + b1 × t)
- AR(1): x(t) = b0 + b1 × x(t − 1) + ε(t)
- One-period-ahead forecast: x(t + 1) = b0 + b1 × x(t)
- Two-period-ahead forecast: x(t + 2) = b0 + b1 × [forecast of x(t + 1)]
- Mean-reverting level (AR(1)): b0 ÷ (1 − b1)
- Residual autocorrelation test: t = residual autocorrelation ÷ (1 ÷ √T)
- RMSE: √[Σ(actual − forecast)² ÷ number of forecasts]
- Random walk (with drift): x(t) = x(t − 1) + ε(t); with drift x(t) = b0 + x(t − 1) + ε(t)
- First difference: y(t) = x(t) − x(t − 1)
- Dickey–Fuller regression: x(t) − x(t − 1) = b0 + g1 × x(t − 1) + ε(t), g1 = b1 − 1; H0: g1 = 0
- Seasonal AR (quarterly): x(t) = b0 + b1 × x(t − 1) + b2 × x(t − 4) + ε(t)
- ARCH(1): e(t)² = a0 + a1 × e(t − 1)² + u(t); next-period variance forecast = a0 + a1 × e(t)²
Common traps
- A log-linear trend forecasts ln y; forgetting to exponentiate is a classic error.
- Durbin–Watson is invalid for AR models; test residual autocorrelations with t-statistics instead.
- A unit root can't be tested with a standard t-test of b1 = 1, and a random walk has no mean-reverting level.
- Regressing one unit-root series on another is valid only if the two are cointegrated.
How it's tested
Vignettes give AR or trend output and residual autocorrelation tables; you forecast one or two periods ahead, find the mean-reverting level, judge stationarity or seasonality, choose between models on RMSE, decide whether a regression of two series is valid, or pick the right model for a described series.
Practise it
The full lesson, worked examples and practice questions are in Tickmark, where every answer choice is explained.
Open the full lesson