Horizon: the Sweet 1.3 stock forecast, product by product
Horizon combines the statistical model and a day-by-day forecast, then keeps for each product the one that predicts best. Measured on eight shops.
Horizon is the stock forecasting AI of Sweet 1.3: for each product, it pits the statistical model against a day-by-day forecast, replays the last four weeks, and keeps the one that predicted best. It takes into account what the calendar already announces: public holidays, sales periods, seasons.
Two models, one arbiter
Every product has its own rhythm. A steady best-seller is forecast very well by a statistical model; a seasonal or trending product is better read day by day. Horizon does not choose once and for all:
- it computes both forecasts;
- it replays them over the product's last 28 days;
- it weights them by accuracy, with the statistical model as the baseline;
- when they genuinely diverge, Sweet Instinct decides.
Restocking, the stock-out date and pre-orders are then calculated on the selected forecast.
The results
On our benchmark, eight shops, four dates, up to 250 products per shop:
| Statistical model | Day-by-day forecast alone | Horizon (Sweet 1.3) | |
|---|---|---|---|
| Range error (lower = better) | 0.0906 | 0.0914 | 0.0905 |
| Mean error (MASE) | 1.439 | 2.047 | 1.424 |
Horizon has the best accuracy of the three, and on difficult products it cuts the error by more than half:
| Shop | Day-by-day forecast alone | Horizon |
|---|---|---|
| Children's brand | 4.14 | 1.77 |
| Outdoor | 3.57 | 1.64 |
| Menswear | 2.77 | 1.43 |
Tuned shop by shop
Each shop has its own Horizon settings, in its configuration. On the tennis and padel shop, the setting chosen on July and August lowers stock cost by 2.4% when replayed on May and June, against the statistical model alone.
Frequently asked questions
Why combine two models rather than pick one?
Because no model wins everywhere. By pitting them against each other product by product, Horizon keeps the best of both, and stays on the statistical model when the day-by-day forecast adds nothing.
Does Horizon take sales periods and public holidays into account?
Yes: public holidays, day of the week, sales periods and the product's seasonality.
What happens if the day-by-day forecast is unavailable?
Horizon carries on with the statistical model. Your recommendations arrive as usual.