forecast-accuracy-review

原文涵盖的主题

• Forecast Accuracy Review
• Required data
• Workflow
• Pitfalls to check explicitly
• Output format

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. 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. 1. 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. 2. Set the benchmarks. Naive (last period) always; seasonal naive when 2+ full seasons exist. These are non-negotiable controls. 3. 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. 4. Score with honest metrics: WMAPE = sum(|error|) / sum(actual) - the volume-w

这里只展示节选，请以完整原文和当前要求为准。