Definition
Mean absolute scaled error (MASE) divides mean absolute forecast error on evaluation data by the mean absolute naive forecast error calculated on training data. For seasonal demand, the baseline uses observations one seasonal cycle apart. This produces a unit-free score for comparing series with different scales.
Why It Matters
- A high-volume staple and a low-volume accessory can have very different unit errors. A common reference helps a buyer interpret both.
- Within a Commerce Intelligence OS, a comparison should retain the assortment and planning decision behind the score.
How It Works
- Freeze training history, evaluation dates, forecast horizon, and seasonal period. Calculate the denominator only from training observations.
- Divide evaluation MAE by that denominator. Flag a zero denominator as undefined; a constant training series cannot provide this scale.
- A score below one means error is below the training baseline scale. It does not establish that the model beat a naive forecast on the same future test period; evaluate that competitor directly.
- Review SKU scores alongside shortage exposure and inventory decisions. Document any portfolio weighting instead of silently allowing the largest category to determine the result.
Ecommerce Example
Context: Illustrative example: a homeware SKU has evaluation MAE of six units and training naive MAE of eight units.
Recommended move: Its MASE is 0.75. The buyer also compares the proposed model with a naive forecast issued for the same evaluation dates.
Why it matters: This hypothetical calculation supports a documented forecast review. It is not an iKawn customer performance result.
iKawn Framework
Define
The iKawn framework connects each score to a buying horizon and SKU population.
Scale
Keep baseline history and forecast errors in the decision record.
Compare
Show category-level exceptions as well as portfolio results.
Review
Use forecast evidence together with commercial stock constraints.
Concise Summary
MASE scales forecast error against training-history naive error. Keep the baseline explicit and assess actual planning outcomes separately.