Figures from run e42c6f452fe9 · the latest run · the restaurant
Model · rank 05 of 09
LSTM
lstmHeavy · not in the browser
A small neural network that reads the last four weeks of a dish’s sales and learns what comes next, retrained for every day it forecasts. Sales only.
Avg MAPE
37.3
rank 05 of 09
σ
17.8
lowest: sarimax
OK days
25 / 30
eligible (≥ 80 %)
Day wins
4
of 25 scored days
- lstm
- day winner
- MAPE % per origin day · gaps are closed days (ND)
0102030405060708090100110
Sun 7 DecSun 21 DecSun 4 Jan
Hyperparameters
- lookback
- 28
- hidden
- 16
- layers
- 1
- epochs
- 60
- learning_rate
- 0.01
- optimizer
- adam
- loss
- mse
- scale
- max
- seed
- 42
- Min history
- 35 days
MAPE %lower is better
- 0–20
- 20–25
- 25–30
- 30–40
- 40+
- ND
- fail
- short
- day winner
Order · run e42c6f452fe99 Oct 2026
- Dataset
- synthetic-year · seed 42 · 7dea…
- Window
- 30 origin days · h = 1
- Roster
- 9 models · 3 light, 6 heavy
- Engine
- py 3.12.15 · uv.lock 5edc…
- Result hash
- 936d 4fbf fe39