TimesFM (Google Research)
Google Research's foundation model for time series: you hand it a history (daily sales, hourly traffic, weekly demand) and it predicts what comes next without you training anything. That is zero-shot forecasting: the model learned the recurring shapes of curves on a very large corpus of public series, and recognises them in yours. It returns a median plus nine quantiles (0.1 to 0.9), so a worst-to-best range, and accepts external variables known in advance (holidays, prices, weather) as covariates. Three ways in. The repo: `pip install timesfm[torch]` (or `timesfm[mlx]` on Apple silicon Macs) with the weights on Hugging Face. BigQuery ML, where version 3.0 finished rolling out in September 2026. And Google Sheets through Connected Sheets, which goes through BigQuery ML. Mind the license, which changed with version 3.0 in August 2026: the code stays Apache 2.0, the weights up to 2.5 too, but downloaded 3.0 weights sit under a non-commercial license; commercial use of 3.0 goes through authorised Google Cloud services. Seen in [Instagram reel DfKIC1blIot](https://www.instagram.com/reel/DfKIC1blIot/) (Ghilès Derbal), which shows the old method alongside: a homemade `train.py` on an A100, three hours per epoch, for a single metric.
Strengths
- Forecasting without training: a history in, a continuation out, the same afternoon
- Median plus nine quantiles, so a readable worst-to-best range
- Reachable without code from BigQuery ML and Google Sheets (Connected Sheets), or in Python from the repo
Limitations
- Downloaded 3.0 weights are under a non-commercial license; the commercial route goes through Google Cloud, so through a bill
- It extends learned shapes: a break specific to your business (launch, competitor, crisis) is in none of them
- Without a clean, regular history (no gaps, a constant step), the forecast is well-presented noise
Best for
- A PM or manager who wants a sales or traffic forecast with its error margin, without a data team
- A team with dozens of series to forecast that cannot train one model per series
- A dev who wants to compare a foundation model with their in-house model on their own series, in a day