Definition
- Temporal autocorrelation (also called serial correlation) means that values in a time series are correlated with their own past values.
- In other words, the value at time $t$ depends (partly) on values at times $t-1, t-2, \dots$.
Mathematical Form
The autocorrelation function (ACF) at lag $k$:
$\rho_k = \frac{\text{Cov}(X_t, X_{t-k})}{\sigma^2}$
- $X_t$ = value at time $t$
- $\sigma^2$ = variance of the series
- $\rho_k$ = correlation between series and its lag-$k$ version
Example
- Stock prices: Today’s price is often close to yesterday’s.
- Weather: Today’s temperature is correlated with yesterday’s.
- Website traffic: Traffic on Monday looks similar to previous Mondays.
Why It Matters
- Violates i.i.d. assumption
- Many ML models assume data points are independent.
- Temporal autocorrelation breaks that — past strongly influences future.
- Forecasting
- ARIMA, SARIMA, and other time-series models explicitly model autocorrelation.
- Autocorrelation plots (ACF, PACF) help identify model orders (AR, MA terms).
- Statistical Testing
- Durbin-Watson test checks autocorrelation in regression residuals.
- If residuals are autocorrelated → model is misspecified.
Positive vs. Negative Autocorrelation
- Positive autocorrelation → if value is high now, it’s likely high next step (momentum).
- Negative autocorrelation → if value is high now, it’s likely low next step (mean-reversion).
Visualization
- ACF plot → shows correlation at different lags.
- Example: Strong spikes at lag 1, 2, 7 → suggests weekly seasonality.
Summary
Temporal autocorrelation = correlation of a time series with its past values.
- Core property of time-series data.
- Measured with autocorrelation function (ACF).
- Critical for forecasting models (ARIMA, SARIMA).
- Must be checked in residuals to ensure valid regression/forecasting.
