Prévision zero-shot (modèle de fondation de séries temporelles)
Predicting the continuation of a series of numbers over time (sales, traffic, consumption) with a model that has never seen that series. 'Zero-shot' means zero training examples on your data: the model was trained once, by its publisher, on billions of points from very varied public series, and it recognises in yours shapes it already knows (season, trend, weekly cycle, end-of-month spike). It is the same reversal that large language models brought to text, applied to curves: a generalist model, usable as is, instead of one model trained per series. Before, forecasting a curve followed the method shown in [Instagram reel DfKIC1blIot](https://www.instagram.com/reel/DfKIC1blIot/): pick an architecture, write a `train.py`, rent a GPU, wait hours per epoch, and start over for each new series or each change of metric. A time-series foundation model skips that step. It usually returns a median and several quantiles, so an uncertainty range, and can take into account variables known in advance (public holidays, prices, weather). Google's TimesFM is the most reachable for a non-developer, because it is wired into BigQuery ML and Google Sheets; Chronos (Amazon) and Moirai (Salesforce) are of the same kind.
Strengths
- No training, no GPU to rent: the series in, the forecast and its range out
- One model for dozens of different series, where one model per series was needed
Limitations
- Extends learned shapes: blind to a break specific to your business, which it still predicts with confidence
- Requires a clean, regular series; gaps or a variable step yield a forecast that looks serious and is worth nothing
Best for
- Understanding why a 'no training' forecasting tool can work, and where it stops