White Noise and Innovation Processes

White noise is a sequence of random values where each one has the same average (usually zero), the same spread, and no relationship whatsoever to any of the values before or after it. Knowing the last ten values tells you nothing about the next one. It's the statistical equivalent of static: no pattern to extract, no signal left to model.

This matters because it's the target every good time-series forecast is aiming for in its errors. If a model's residuals — the difference between its predictions and what actually happened — still show a pattern, like today's error predicting tomorrow's error, that's a sign the model left usable information on the table and could be improved. If the residuals look like white noise, the model has extracted everything predictable and what's left really is irreducible randomness. In this context, the white noise term is often called the "innovation" — the genuinely new, unpredictable information arriving at each time step, as opposed to whatever was already forecastable from the past.

A worked example

Fitting a model to a stock's daily returns and checking the residuals: if the correlation between today's residual and yesterday's residual is close to zero across many lags, the residuals look like white noise and the model has captured the predictable structure available. If instead residuals are still noticeably correlated with their own recent past, the model is under-specified — there's a systematic pattern in the errors themselves that a better model should absorb.