Your forecast model's residuals fail a Ljung-Box test at lag 10 — what does that mean?

## What the test does When a forecasting model has done its job, the residuals — actual minus fitted, at each time point — should look like white noise: no remaining correlation with their own past, because any such correlation is predictable signal the model failed to use. Eyeballing the residual correlogram lag by lag works, but it invites the multiplicity problem: with twenty lags plotted, one or two chance crossings of the significance bands are expected under the null, so a per-lag reading gives no clean verdict. A **portmanteau test** solves this by pooling. The Ljung-Box statistic is ``` Q = n(n + 2) * sum over k = 1..h of [ r_k^2 / (n - k) ] ``` where `n` is the number of observations, `r_k` is the sample autocorrelation of the residuals at lag k, and `h` is the number of lags tested. Each squared autocorrelation is weighted by `1/(n-k)`, which upweights longer lags to compensate for their being estimated from fewer overlapping pairs — that small-sample correction is what distinguishes Ljung-Box from the earlier Box-Pierce statistic, which simply summed `n * r_k^2`. Under the null, `Q` follows approximately a chi-squared distribution. Since `Q` grows as the residual autocorrelations get larger in either direction, the test is one-sided in the statistic: **large Q, small p-value, reject**. ## The null, stated precisely The null hypothesis is: *the autocorrelations of the residual series at lags 1 through h are all zero.* It is a joint statement about the whole block of lags, not about any single one. So: - **Small p-value (reject).** At least one of those lags carries real autocorrelation. The model has missed structure. The test does not say which lag — that is what the residual ACF plot is for. - **Large p-value (fail to reject).** No detectable linear autocorrelation in the first h lags. This is *not* proof the model is correct. It is the absence of one specific kind of evidence against it, and on a short series the test has little power to find modest autocorrelation at all. The standard misstatement to avoid: a large p-value does not "accept" the null or prove the residuals are white noise. It says the data give no strong reason to reject that description. ## Degrees of freedom This is where candidates most often slip. Applied to a *raw* series, `Q` has h degrees of freedom. Applied to the **residuals of a fitted model**, the residual autocorrelations have already been shrunk by the fitting procedure — the estimation deliberately drove some of them toward zero — so `Q` is smaller than it would be for a genuinely independent series. The correction is to subtract the number of estimated lag parameters from the degrees of freedom. Using the uncorrected h on fitted residuals produces p-values that are too large, biasing you toward concluding the model is fine when it is not. ## Choosing h There is no single right answer, but two conventions dominate. For non-seasonal data, a moderate h in the region of 10 is common. For seasonal data with period m, testing to about `2m` is the usual choice, so that at least the first two seasonal lags fall inside the window — a model that mishandles seasonality will show its failure at lag m, and a test truncated at lag 5 on monthly data would never see it. The tradeoff: too small an h can miss structure that lives at longer lags; too large an h dilutes real autocorrelation at a few lags across many null ones, costing power. Do not tune h until the test passes — that is a form of selective reporting. ## What the test does not cover Ljung-Box tests **linear autocorrelation of the residual level**, and nothing else. Residuals can pass it comfortably and still be badly behaved: - **Changing variance.** Calm periods followed by volatile ones leave the residual level uncorrelated while the residual *magnitudes* cluster. A common probe is to run the same test on squared residuals, where volatility clustering shows up as autocorrelation. - **Non-normality.** Fat tails or skew affect interval forecasts and are invisible to this test. - **Nonlinear dependence.** A relationship the residuals carry in a non-linear form contributes little linear autocorrelation. - **Bias.** A residual mean far from zero is a systematic offset the test says nothing about. ## Durbin-Watson, the narrower relative Durbin-Watson is the other statistic you will be asked about here. It is computed on residuals as ``` d = sum (e_t - e_{t-1})^2 / sum e_t^2 ``` and ranges from 0 to 4, with the approximation `d ≈ 2(1 - r_1)`. A value **near 2 is the no-serial-correlation reading**; values near 0 indicate strong *positive* first-order autocorrelation, and values near 4 strong *negative* first-order autocorrelation. Its limitation is right there in the algebra: it looks only at **lag 1**. Residuals with a clean seasonal spike at lag 12 and nothing at lag 1 will produce a Durbin-Watson comfortably near 2 while being obviously autocorrelated. It is also not valid in the usual form when a lagged value of the dependent variable appears among the regressors, a situation that arises constantly in time-series work. That is precisely why a portmanteau test over a block of lags is the better default for residual checking. ## What to do after a rejection Rejection is a prompt, not a prescription. Plot the residual ACF and PACF and locate the offending lags. A lag-1 spike says short-memory structure remains. A spike at the seasonal period says the seasonal component is mishandled. Large positive autocorrelation across many short lags usually means the series was never adequately leveled and the model is fitting a moving mean. A single marginal crossing in a long lag window, with a p-value hovering near the threshold, may be the multiplicity problem reasserting itself and deserves less alarm than a clear, interpretable spike.