Skip to content
Machine Learning

Stationarity and Autocorrelation

Account required to view full content

Almost every model in the first two modules quietly assumed the data points were independent and identically distributed. Financial data are neither. A price today is the price yesterday plus a small change, volatility comes in clusters, and the statistical properties of a series drift over time. Stationarity is the property that makes a time series learnable at all, and autocorrelation is the tool that measures the memory inside it. This lesson defines both, shows why a non-stationary series wrecks an otherwise sound model, works through the sample autocorrelation of a return series and the unit root of a random walk by hand, and covers the tests (ADF and KPSS) and transforms you use to make a series stationary. It is the foundation for the rest of the module: ARIMA, GARCH, and cointegration all build directly on these ideas.