How the models compete
For each product several forecasting models are trained on your own history, tested on sales they did not see, and the one whose mistakes cost you least wins.
On this page
No single model fits everything: a stable product, a seasonal one and one that sells now and then need different tools. StockAI does not choose by intuition; it makes the candidates compete and keeps whichever would have done best for your wallet.
The candidates
- StockAI has nine models available: statistical ones, machine-learning ones and a global one that learns from your whole catalogue at once.
- In a normal update from My sales, five of them compete, including the global one — the one that gives a new or short-history product a usable forecast.
- Each product only gets the models that suit its series type (stable, seasonal, intermittent, volatile or short). That routing can only shrink the list, never add a model that was not selected.
On screen the models are called "Model 1", "Model 2"…: the algorithm's name changes no purchasing decision.
How they are tested
Each model is evaluated walking forward: it is trained on history up to a cut-off and measured against the weeks that follow, which it did not see. That is repeated over several successive cut-offs, so a good result does not hinge on one lucky date.
A product needs at least 20 periods of history to enter this competition. Those that fall short are left out of the forecast.
The cheapest wins, not the least wrong
Running out costs more than holding extra: you lose the sale and sometimes the customer. So errors are weighted: a shortage counts three times as much as a surplus. The winner is the model with the lowest cost measured that way, which may not be the one with the lowest percentage error. You see it in the "Metrics" tab of Forecasts, sorted by "Cost".
References and the combined model
- References: "last value", "season" and "average". They are the bar to beat — what ordering the same as last time would do — and never win.
- Combined model: a blend of the models that competed. It is shown for comparison, but no purchase is computed from it.
The final forecast uses all your history
The model being graded stops at the cut-off, because it is scored on what comes after. The model that forecasts, once chosen, is retrained on all your history, including the most recent weeks.