Does white noise always have zero mean?

This entry adopts that common convention; some usage permits a constant mean and centers it first.

Must white noise be Gaussian?

No; the two-point construction demonstrates otherwise.

Can white noise contain runs?

Yes. Finite sequences can display runs by chance, and weak white noise also need not be independent.

AI Glossary / Time series and signal processing / Intermediate

White noise

Definition

White noise is a time-series process with a constant finite variance and no correlation between observations at different times. This entry uses the common zero-mean, weak-white-noise definition: E(e_t) = 0, Var(e_t) = sigma squared, and Cov(e_t, e_s) = 0 when t differs from s. Independence and normality are additional assumptions, not automatic consequences.

What it means

White noise provides a reference for variation that has no linear temporal pattern. It is central to forecasting because a useful fitted model should leave little obvious autocorrelation in its residuals. That does not mean every residual must be small, or that a model with apparently uncorrelated errors is automatically good.
The distinction between a mathematical process and a finite observed sequence matters. A genuinely white-noise process can produce apparent runs and nonzero sample correlations by chance. Conversely, a short sequence with a quiet autocorrelation plot does not prove that the underlying process satisfies every white-noise assumption.

How it works

Under the stated definition, all nonzero-lag population autocorrelations are zero, while the lag-zero autocorrelation is one when variance is positive. For discrete-time white noise, its covariance structure corresponds to a flat power spectrum. Gaussian white noise adds joint Gaussianity; independent, identically distributed zero-mean observations with finite variance are another familiar sufficient construction.
In practice, inspect the time plot, mean, variability, and sample autocorrelation function. Consider a joint residual-correlation test with appropriate allowance for fitted model parameters. Check squared residuals or other diagnostics when changing volatility is plausible. Zero ordinary autocorrelation does not rule out nonlinear dependence, and lack of statistical significance can reflect limited data rather than the absence of structure.

Three practical examples

A demand forecaster checks whether residuals still have a weekly correlation pattern. A visible pattern suggests the model missed recurring information rather than leaving only white noise.
A simulator generates independent errors taking values minus three and plus three with equal probability. These errors provide a simple non-Gaussian white-noise input under the stated assumptions.
An AI monitoring service compares its alert residuals across time and separately checks whether large errors cluster. Small ordinary correlations alone would not establish stable error risk.

Think of it this way

Think of repeatedly drawing a signed token from a bag, replacing it, and mixing before every draw. The sequence can contain streaks, even though earlier draws give no advantage in predicting the next one under this independent construction. White noise describes a process, not a requirement that every short stretch look perfectly balanced.

Worked example

Construct an original independent process in which each e_t is -3 or +3 with probability one half. Its mean is (-3)/2 + 3/2 = 0. Its second moment is 9/2 + 9/2 = 9, so its variance is 9. For different times, independence gives E(e_t e_s) = E(e_t)E(e_s) = 0, hence zero covariance. This construction is white noise but is not Gaussian: its values have only two possible outcomes.
For the average of four observations, the variance is (9 + 9 + 9 + 9)/16 = 2.25 because the cross-covariances vanish. Its standard deviation is 1.5. The particular outcome [+3, +3, +3, -3] has average 1.5 and probability (1/2)^4 = 1/16, or 6.25%, for that exact ordered sequence.
The nonzero observed average does not contradict the process's zero expected mean. These are exact theoretical calculations for the specified construction, not a statistical test of a real dataset or a claim that ordinary business errors follow this distribution.

Illustrative case study

Fictional case: ClearRoute evaluates an AI model that predicts daily delivery volume. The analyst finds little residual autocorrelation and initially describes the forecast errors as fully random. A second plot shows that the magnitude of errors rises sharply during disruption periods.
The team revises its conclusion: the first diagnostic found little evidence of linear dependence at the inspected lags, but uncertainty still needs investigation. It examines data quality, changing variance, and performance on later periods before changing planning buffers. This fictional example illustrates careful interpretation; it supplies no measured forecasting improvement or proof of a particular stochastic model.

Common misconceptions and mistakes

Uncorrelated does not generally mean independent. White noise need not be normally distributed, and the weak definition does not exclude every nonlinear pattern. A stationary process can have autocorrelation and therefore fail to be white noise.
Do not confuse a random walk with white noise: a random walk accumulates shocks, while its increments may be white noise. Do not insist that sample autocorrelations equal zero exactly. Approximate individual ACF bounds are not simultaneous guarantees over many lags. Finally, white-looking residuals do not establish unbiased decisions, accurate prediction intervals, or strong out-of-sample performance by themselves.

Business use

Use white noise as a diagnostic benchmark and an explicit modeling assumption. It helps teams ask whether a forecasting system has left exploitable linear structure unexplained. It also provides controlled disturbances for simulation and testing.
A business report should name the diagnostic, sample period, and assumptions rather than announce that randomness has been proven. Evaluate prediction accuracy, residual bias, interval coverage, and operational consequences separately. An expensive model can produce apparently uncorrelated residuals while still being worse than a simple baseline for the decision at hand.

Check your understanding

What is the variance in the worked construction?

Answer: Nine, from the zero mean and second moment of nine.

Why is the four-observation average less variable?

Answer: Averaging uncorrelated equal-variance observations divides variance by four.

Would a small sample correlation prove independence?

Answer: No. It neither proves a zero population correlation nor establishes the stronger independence condition.

Related terms

Autocorrelation
Stationarity
Random walk
Ljung-Box test
Gaussian noise
residual diagnostics
heteroscedasticity

Learn next

Compare independent Gaussian and two-point white-noise simulations. Examine how their sample ACFs fluctuate, then contrast them with an autoregressive process. Study nonlinear dependence and residual-variance diagnostics before treating an absence of ordinary autocorrelation as a comprehensive model check.

Sources

Supplied Glossary of Forecasting Terms: https://robjhyndman.com/mwh3/FG4.pdf . Its brief entry was read; this article makes the weak-process definition explicit. Ryan Tibshirani, Berkeley time-series lecture notes: https://www.stat.berkeley.edu/~ryantibs/timeseries-f23/lectures/characteristics.pdf . Definition and independence distinction consulted. Hyndman and Athanasopoulos, graphics chapter: https://otexts.com/fpp3/graphics.html . Sample ACF guidance consulted. The calculations and case are original.

Previous: Whiskers Next: White pages schema