Forecast Accuracy Review
A forecast is only worth what it adds over the free alternative: shipping last period's number. Every review must answer "how many points does this process add over naive?" before any model discussion.
Required data
Per-SKU demand history at the planning bucket (usually monthly): sku, period, qty. If evaluating an existing forecast, also the forecast values with their creation dates (to avoid hindsight leakage). 18+ periods per SKU for a meaningful backtest; flag SKUs with less.
Workflow
- Profile the demand first. Per SKU compute mean, CV and zero-period share; classify smooth / erratic / intermittent / lumpy (defaults: CV 0.5 and 1.0 boundaries, intermittency at >25% zero periods - state them, adjust to natural breaks). Accuracy expectations differ by class; never report one blended number alone.
- Set the benchmarks. Naive (last period) always; seasonal naive when 2+ full seasons exist. These are non-negotiable controls.
- Backtest rolling-origin. One-step-ahead forecasts for each of the last 6+ periods, expanding window, using only data before each origin. A single train/test split is one lucky draw - do not accept it.
- Score with honest metrics:
- WMAPE = sum(|error|) / sum(actual) - the volume-weighted headline
- Bias = sum(error) / sum(actual) - direction; a fine WMAPE with persistent bias is quietly building excess stock or stockouts
- MAPE only as a footnote, and always disclose how many zero-actual periods it dropped
- Deliver the FVA verdict. FVA = WMAPE(naive) - WMAPE(candidate), per segment and overall. Negative FVA means the process destroys value - say it plainly.
- Validate. Recompute WMAPE for one model directly from the raw backtest rows and confirm it matches the table before presenting.
Pitfalls to check explicitly
- MAPE with zeros: undefined on zero-actual periods; silently dropping them fakes precision on intermittent SKUs.
- MAPE asymmetry rewards under-forecasting (errors capped at 100% below, unbounded above).
- Aggregation mix: a good total can hide terrible A-item accuracy; always show the value-weighted cut.
- Lumpy segments: if WMAPE > ~100%, the honest recommendation is an inventory-policy answer (buffers, MTO), not a better model.
- Hindsight leakage: forecasts must predate actuals; check timestamps when auditing an existing process.
Output format
- Scoreboard table: model x (WMAPE, bias, MAPE-footnote), sorted by WMAPE
- FVA statement: "the process adds/destroys X points vs naive" - overall and per segment
- Segment table (pattern x best approach)
- Two or three recommendation sentences tied to segments, not globals
Worked example with five baseline models and charts: https://github.com/gulmezeren2-byte/forecast-accuracy-lab
Source: industrial-engineering-ai-skills by Eren Gulmez (MIT). The full method pack - entry skill, role agents, data-hygiene rules and artifact templates - lives there.
