Physix FrontierPhysix Frontier · News Briefing Card (Hacker News · Oct 30, 2026)
Test: Timestral zero-shot needs ~2 days of data to beat simple rules
KEY FACTS
- Winder ran zero-shot forecasting tests on Mistral's new Timestral model in October 2026.
- On retail data, Timestral's error was 14% lower than AutoETS and 3% lower than LightGBM.
- Electricity demand forecasting needed about two and a half days of data, while retail needed eight weeks of sales to beat repeating last week.
- When historical data covered less than two months, cross-product trained models outperformed single-product history.
- Faced with never-before-seen seasonal shifts, Timestral's forecasts failed, with simple mechanistic models posting less than one-third its error.
KEY DATA
14%Retail error reduction vs AutoETS
3%Retail error reduction vs LightGBM
about 2 daysMinimum data for electricity demand forecasting
8 weeksMinimum data for retail forecasting
PHYSIX OBSERVATION
Pretrained time-series models do hold an edge when data is plentiful, but cold start and out-of-distribution scenarios remain weak spots. Vendors' claimed zero-shot capability needs a discount: if a user has less than two months of data, traditional models or cross-product transfer are more reliable. This reminds the industry that foundation models are no panacea, and scenario adaptation and mechanistic knowledge remain irreplaceable.
Source: Hacker News report