Automatic model optimization

Auto is the default. It picks the best forecast model per variant by comparing how well each model would have predicted the recent past — and refreshes that choice whenever new data shows up.

What Auto does

For every variant, Auto:

Looks at the order-history shape — length, trend, seasonality, variance.
Runs every applicable model against the history.
Scores each model by how well it would have predicted the recent past (the held-out validation window).
Picks the winner and uses it for that variant’s forecast.

The result is one model per variant, refreshed as new data arrives.

When it runs

Auto re-checks a variant whenever its inventory or orders change in a way that could shift the optimal model — typically within minutes of the event. New sales, received inventory, and PO updates all count.

It also runs:

Once after the initial data sync on install.
On demand, when you re-trigger it manually from Settings → ABC Analysis.

You don’t need to think about cadence — Auto follows the data.

What Auto doesn’t do

It can’t see future events. Use growth amplifiers to tell the forecast about campaigns or launches that don’t appear in history.
It doesn’t expose per-variant tuning knobs. Smoothing factors, window lengths, seasonality lengths — all internal. See Forecast models.
It doesn’t replace your judgement on outliers. If you know a one-off spike (a viral moment, a press mention) won’t repeat, an amplifier or a manual exclusion is the right tool.

Triggering manually

The only manual entry point: Settings → ABC Analysis → re-run model optimization. Useful after a long stretch where many variants have accumulated meaningful new history at once.