Definition
Forecast residual autocorrelation measures association between model residuals separated by a time lag. For one-step demand predictions, residuals are observed demand minus its fitted prediction. A recurring pattern can indicate that the model has left usable time structure unexplained; it is a diagnostic rather than a complete forecast-quality score.
Why It Matters
- A buyer can see acceptable average error while the model repeatedly misses the same day of the week. That timing matters for replenishment.
- A Commerce Intelligence OS needs to connect forecast diagnostics with the ordering decision affected by those errors.
How It Works
- Keep a regularly spaced demand series and its aligned one-step fitted values. Record missing periods and availability constraints before interpreting residuals.
- Inspect a time plot and autocorrelation plot at relevant lags. For daily retail data, a weekly pattern is a useful investigation target rather than automatic proof of missing seasonality.
- Use a suitable joint test such as Ljung-Box with documented lags and model degrees-of-freedom adjustment. An isolated spike is weaker evidence when many lags have been inspected.
- Test a revised model on future-held-out periods. Uncorrelated residuals alone do not establish the best forecast, and overlapping multi-step forecast errors need separate interpretation.
Ecommerce Example
Context: Illustrative example: an accessory forecast leaves positive errors on successive Saturdays and negative errors on adjacent weekdays.
Recommended move: Investigate the weekly calendar, campaign schedule, and stock availability before testing a revised demand model.
Why it matters: A buyer can judge whether the revision improves the actual replenishment horizon. This scenario is hypothetical and reports no iKawn model result.
iKawn Framework
Align
The iKawn framework links demand observations and forecast versions.
Diagnose
Expose repeated errors at commercially relevant time lags.
Investigate
Connect unexplained timing with merchandising and availability records.
Validate
Require held-out planning evidence before changing the forecast workflow.
Concise Summary
Residual autocorrelation helps locate unexplained time patterns. Investigate their cause and validate any model change beyond the diagnostic itself.