Settings

Models & accuracy

Control how forecasts are generated and when SKUs are flagged for review.

Forecast granularity

Time bucket the models forecast at.

Forecast horizon

Number of buckets to forecast ahead.

Model selection

How the champion model is chosen per SKU.

Prediction interval

Confidence level for the forecast band.

Recursive optimisation

Iteratively re-tune models against history each cycle. (Not yet wired to the forecasting service.)

Enabled model families

Toggle which model families participate in champion selection. Seasonal naive is always on — it's the FVA benchmark and champion-selection guardrail.

StatsForecast (Nixtla)

AutoARIMA, AutoETS, Holt-Winters and Theta.

Machine learning

XGBoost and LightGBM with engineered features.

Ensemble

Weighted blend of statistical and ML champions.

Exception thresholds

Accuracy and bias limits that flag a SKU for planner review.

Watch WAPE (%)

Above this, a SKU is flagged to watch.

At-risk WAPE (%)

Above this, a SKU is flagged at risk.

Bias tolerance (±%)

Sustained bias beyond this triggers an alert.