MerQuR+Statistics Glossary

Statistics Glossary → White noise

White noiseconcept

A time series with a constant mean (usually zero), constant variance and values at different times that are uncorrelated with each other; pure randomness with no predictable structure.

In white noise the autocorrelation at every lag is zero and the spectrum is flat across all frequencies; the name comes from white light, which contains all frequencies equally. Past values therefore carry no linear information about future values. The residuals of a well-built time series model are expected to be white noise; if autocorrelation remains in the residuals, the model has missed part of the structure in the series.

Testing uses autocorrelation tests such as Ljung-Box or periodogram-based tests (e.g. the departure of the cumulative periodogram from a straight line). Uncorrelated is not the same as independent: values can be uncorrelated and still dependent, as in financial returns whose variance clusters over time. Only for Gaussian white noise does uncorrelatedness imply independence. White noise does not have to be normally distributed either.

An everyday example

The numbers from successive throws of a fair die, taken as deviations from their mean, are an example of white noise: a 6 on the last throw does not make the next one tend to be high or low. If the residuals left by a model of a sales series behave like this, the model has captured all the predictable structure in the series.

In MerQur

In the Time Series analysis, 'Spectral extras: White noise test' tests whether the series is white noise using the periodogram.

Analyses where this concept appears

Time Series

Related concepts

Ljung-Box (portmanteau) testPeriodogram and spectral density