How do autoregressive models work in time series analysis, and why does stationarity matter so much?

Time series analysis uses past values of a variable to forecast its future values. Autoregressive (AR) models are the workhorse, but they require stationarity — a property that many financial time series violate.

The AR(1) Model:

x_t = b0 + b1 * x_(t-1) + e_t

The current value depends on the previous value plus an error term. The key parameter is b1:

• If |b1| < 1: the series is covariance stationary (mean-reverting)
• If b1 = 1: the series has a unit root (random walk, non-stationary)
• If |b1| > 1: the series is explosive (non-stationary)

Why Stationarity Matters: If a series is non-stationary (has a unit root), standard regression results are spurious — you get high R-squared and significant t-statistics even when the variables have no real relationship. This is because both series are trending (possibly in the same direction) by coincidence.

Loading diagram...

Characteristics of a Covariance Stationary Series:

1. Constant mean over time
2. Constant variance over time
3. Covariance between any two observations depends only on the lag distance, not the time period

Practical Example: Suppose you model quarterly GDP growth as an AR(1):

GDP_growth_t = 0.8% + 0.65 x GDP_growth_(t-1) + e_t

Since |0.65| < 1, the series is stationary. The long-run mean = b0 / (1 - b1) = 0.8% / (1 - 0.65) = 0.8% / 0.35 = 2.29% quarterly growth rate. Forecasts will revert toward this mean over time.

Contrast with a stock price series where b1 is approximately 1.0 — the series wanders without reverting. Forecasting future stock prices with an AR model would be meaningless because the series is non-stationary. You'd need to model returns (first differences of log prices) instead.

Common Exam Traps:

• A high R-squared alone does NOT validate a time series model — check for autocorrelated residuals and unit roots
• Always test residuals with the Durbin-Watson statistic for remaining autocorrelation
• If residuals are autocorrelated, the model is misspecified (try adding more lags)

Practice time series diagnostics in our CFA Level II question bank.