Improved
- Every confidence range is now guaranteed to contain its own forecast, enforced across all four model backends and the ensemble. The guarantee also covers the full quantile fan when you request one, so the forecast, the range and the fan always describe a single consistent picture — whatever the underlying model produced.
- The
advanced-quantizedmodel now publishes the median of its predictive distribution as the point forecast. The median is the value that minimises the error measures forecasts are actually judged on, and it is drawn from the same distribution as the confidence range — so the point and the range around it are two readings of one forecast rather than two separate estimates. - Model selection is more robust on intermittent series — the sparse, spiky demand patterns where many periods are zero. Percentage-based error measures are mathematically undefined against a zero actual, so on this kind of data the selector now ranks candidates on MASE, which is scale-free and defined everywhere, rather than on a measure only some candidates can be scored with. The result is that every candidate model competes on equal footing on exactly the data where the choice is hardest.
- Realized-accuracy tracking now recognises when a series changes scale between runs — a feed switched from units to thousands, or from a total to a per-store figure. Runs measured against history on a different scale are held out of the scorecard, so
model: autorouting and auto-calibration read accuracy figures that describe forecast quality rather than a change in the data itself.
No request or response changes are needed to pick any of this up, and requests unaffected by the above return exactly what they did before.