Models

Model · kindAvg MAPE · lower is better

1. 01SARIMAX + weathersarimaxA seasonal ARIMA over a seven-day season, given the weather and the calendar’s day types as regressors.Heavy · not in the browser33.4σ 16.1 · 25 d
2. 02Ensemble v2ensemble-v2Blends three simple forecasts: the same-weekday average at half weight, and an exponentially weighted average and a seven-day mean at a quarter each.Light · runs in the browser in v235.3σ 19.1 · 25 d
3. 03Holt-Wintersholt-wintersExponential smoothing with a damped trend and a weekly season. Falls back to the last week’s mean when the history is short or the fit fails.Heavy · not in the browser35.8σ 19.8 · 25 d
4. 04XGBoostxgboostGradient-boosted trees over the same features as the random forest: each tree corrects the errors of the ones before it.Heavy · not in the browser37.2σ 21.3 · 25 d
5. 05Random Forestrandom-forestMany decision trees over recent sales, the weekday, the weather and the calendar, their forecasts averaged.Heavy · not in the browser37.7σ 19.7 · 25 d
6. 06ThetathetaSplits the series into a long-run trend and short-run smoothing, and averages the two lines. Cheap, steady, and hard to beat on a flat week.Light · runs in the browser in v239.3σ 21.2 · 25 d
7. 07Same weekdaysame-weekdayForecasts each dish as its average on the same weekday over the last four weeks. The baseline every other model has to beat.Light · runs in the browser in v240.0σ 21.0 · 25 d