Definition
Dependence means observations are statistically related, so the joint distribution does not factor into independent pieces. Autocorrelation is a correlation between a series and a lagged version of itself. Dependence changes effective information and can alter standard errors, bootstrap design and convergence rates.
How it is calculated
Lag-k autocorrelation rho_k is the correlation between X_t and X_{t-k}.
| Symbol | Meaning | Unit |
|---|---|---|
| rho | lag autocorrelation | dimensionless |
| Corr | correlation operator | dimensionless |
| X | time-series observation | series units |
| k | lag | periods |
Interpretation
If yesterday changes what today tends to look like, observations are not fresh independent draws. A positive autocorrelation can make a naive standard error too small because the data contain fewer independent pieces of information than their row count suggests. Zero correlation at one lag does not prove independence.
Assumptions and conventions
-
The lag and sampling frequency are defined.
-
The series has finite positive variance and is covariance stationary so correlation depends only on lag.
-
Correlation is used as a descriptor and does not exhaust all forms of dependence.
Autocorrelation is dimensionless; k is a period count.
Limitations and common misreadings
-
Dependence is broader than linear autocorrelation.
-
More rows do not necessarily mean the same number of independent observations.
References
- Peter J. Brockwell, Richard A. Davis. (2016). Introduction to Time Series and Forecasting. Springer Texts in Statistics. Chapter 2, Stationary Processes. doi:10.1007/978-3-319-29854-2. Source

