Autocorrelation measures the linear relationship between a time series and lagged versions of itself. At lag k, the question is whether observations tend to move with or against observations k periods earlier after the series has been centered and scaled.
Lag defines the relationship being measured
Lag 1 compares the current observation with the immediately preceding observation. Lag 5 compares it with the observation five periods earlier. The meaning of that lag depends on sampling frequency: five daily observations are not the same horizon as five one-minute observations.
The autocorrelation coefficient is bounded
A population autocorrelation at lag k can be written as the lag-k autocovariance divided by the variance when those quantities are defined. Like an ordinary correlation coefficient, its value lies between −1 and +1.
Positive values indicate a tendency for deviations to move in the same direction at that lag; negative values indicate an inverse relationship. A value near zero means weak linear serial relationship at that lag, not proof of complete independence.
Sample autocorrelation is an estimate
An autocorrelation function calculated from historical data is noisy, especially at long lags or with a short sample. Looking across many lags also creates multiple-comparison problems: some apparently unusual values can arise by chance.
Confidence bands and formal tests can help, but their assumptions and the number of lags examined should be part of the interpretation.
Returns and volatility proxies can behave differently
A return series may show little linear autocorrelation while squared or absolute returns display persistence associated with volatility clustering. That does not contradict the return result; the transformations answer different dependence questions.
Volatility Modelling and GARCH Models cover conditional variance directly.
Autocorrelation in residuals can reveal an incomplete model
After fitting a time-series model, researchers often inspect residual autocorrelation. Remaining serial structure can indicate that the model has not captured all of the dependence in the observed series.
That finding is a diagnostic, not an instruction to keep adding parameters until every historical pattern disappears.
Dependence is not the same as economic profitability
A statistically measurable lag relationship can be too small, unstable or costly to trade. Bid-ask effects, stale prices, non-synchronous trading and other microstructure features can also create short-horizon serial patterns.
Any trading interpretation still requires realistic costs, execution and out-of-sample validation through pages such as Strategy Validation.
Stationarity affects interpretation
Standard autocorrelation analysis assumes a relationship that is meaningful across the sample. If the mean, variance or dependence structure changes materially, a single full-sample coefficient can hide different regimes.
Stationarity explains that modeling assumption, while Market Regime Detection covers changing-state models.
Do not infer direction from volatility autocorrelation
Persistence in squared or absolute returns suggests that large magnitudes can cluster. It does not say whether the next return will be positive or negative.
Use autocorrelation as one diagnostic in a wider workflow
- Define the variable and sampling interval.
- Inspect the series and relevant transformations.
- Choose economically meaningful lags before interpreting the plot.
- Account for estimation noise and multiple lags.
- Check stability across time.
- Validate any decision rule on later data.
Autocorrelation is useful because it makes serial dependence measurable. Its limit is equally important: correlation through time does not by itself identify cause, stability or a trade that survives real-world frictions.