A new time series forecasting model called Timer-M1 just topped two of the three benchmarks researchers use to judge this category of AI.
The model comes from a team that built it around what they call primitives: small recurring temporal and relational patterns, like trends or seasonal cycles, that show up across wildly different kinds of data, whether that is retail sales, sensor readings or financial series. Instead of training only on real-world data, the researchers generated synthetic series that embed those shared primitives, then combined them with real series into multivariate samples. They also split each sample into target variables, past-only covariates and known-future covariates, so the model learns to use whatever outside information is actually available at forecast time. Tested across three large benchmarks called FEV, TIME and GIFT-Eval, it ranked first on FEV and TIME and second on GIFT-Eval against other recent forecasting foundation models.
This matters because most forecasting tools still specialize: one for retail demand, another for energy load, another for traffic. A model that generalizes across domains without retraining could cut the cost of deploying forecasting into new contexts, from inventory planning to grid management. That is the same zero-shot promise that has driven interest in foundation models generally, just applied to numbers instead of text.
Still, this is one paper claiming top results on benchmarks it was partly built to win, and a second-place finish on GIFT-Eval is a reminder that no single model has actually solved general forecasting yet.