Skip to main content

Model training

The metrics on this page measure the quality of the base statistical model that generates the AI forecast. They do not reflect manual adjustments made by Sales, Marketing, or Consensus; to evaluate those adjustments, see Forecast accuracy and Forecast Value Added (FVA).

Usage context

These metrics are computed internally during the model training and algorithm selection process. Their primary audience is technical (data science / analytics team). Business users can safely disregard them.

Metrics catalogue

KPIFull nameWhat it measuresUnitGood when
MASEMean Absolute Scaled ErrorAbsolute error normalised against a naive reference model (random walk)Dimensionless< 1 (better than naive)
WAPEWeighted Absolute Percentage ErrorDemand-weighted MAPE; robust to low-volume SKUs%Minimum
wQLWeighted Quantile LossHow well the model estimates demand distribution quantiles (confidence intervals)Dimensionless / normalisedMinimum
MSEMean Squared ErrorSame as ECM in the accuracy section; penalises large errorsUnits²Minimum
MSE and ECM

The MSE (Mean Squared Error) in this training section is the same indicator as the ECM (Error Cuadrático Medio) documented in Forecast accuracy. The formula is identical: (1/n) · Σ(Dᵢ − Fᵢ)². The difference is purely one of naming convention.

Description of each metric

MASE

MASE scales the model's MAE against the MAE of a simple reference model (typically a random walk or seasonal lag-1 difference). MASE < 1 means the model outperforms the naive benchmark; MASE > 1 means it performs worse.

TODO: exact formula (pending confirmation from the analytics team).

WAPE

WAPE weights the percentage error by each SKU's demand, so high-volume SKUs carry more influence. This makes it more stable than MAPE when the portfolio includes very low-demand or intermittent items.

TODO: exact formula (pending confirmation from the analytics team).

wQL (weighted Quantile Loss)

wQL evaluates the quality of the model's probabilistic estimates: how well calibrated the quantiles are (for example, the 0.9 quantile should cover 90 % of real observations). It is especially relevant for safety stock calculations and fill rate targets.

TODO: exact formula (pending confirmation from the analytics team).

MSE

See ECM in Forecast accuracy. Formula: (1/n) · Σ(Dᵢ − Fᵢ)².

Relationship to model selection

AInventory evaluates several time-series algorithms and selects the model with the best aggregate performance on these metrics over the validation set (held-out historical data). The selected model produces the Baseline (AI) visible in the forecast module.