Time series foundation models can be few-shot learners

Forecasting is one of those unglamorous but essential jobs in tech — predicting inventory, energy loads, traffic — and until recently it meant building a bespoke model for every single use case. Google's TimesFM tried to fix that with zero-shot forecasting, letting one pre-trained model handle new tasks without extra training. It worked well enough, but there was an obvious gap: sometimes you have a few relevant examples lying around, like traffic data from a nearby highway, and a smart model should be able to use them without dragging you back into full supervised fine-tuning.

That's what TimesFM-ICF, unveiled at ICML 2025, is built to solve. Google's researchers didn't bolt on some external retrieval trick. Instead, they retrained TimesFM itself to recognize and use in-context examples during continued pre-training. The tricky part was keeping the model from scrambling separate data streams into one confused signal — imagine feeding it sunglasses sales from one store and umbrella sales from another, and having it treat that as a single, nonsensical trend. Their fix was a learnable separator token, essentially a digital paragraph break, so the model knows where one example ends and another begins.

The architecture itself stayed familiar: a patched decoder that chops time series into 32-point chunks, runs them through stacked transformer layers with causal self-attention, and reconstructs 128-point forecasts through a shared MLP. What changed was the training data and objective, teaching the model to treat context examples as genuinely useful signal rather than noise to ignore.

On 23 datasets the model had never seen, TimesFM-ICF beat the base model by 6.8 percent and matched TimesFM-FT, the supervised fine-tuned version that Google considers its strongest baseline. That's the real headline here — not the accuracy bump alone, but the fact that it arrived without the usual fine-tuning grind of curating training splits and running a separate job per dataset. Give the model more examples and it gets more accurate, at the cost of slightly longer inference, which is a fair trade for most businesses.

What's next sounds almost mundane compared to the technical lift: figuring out which in-context examples actually help. Right now the system samples from a dataset's own history, which works, but smarter example selection could squeeze out even more accuracy. Still, the practical implication is straightforward. A single forecasting model that adapts on the fly to a new product launch or a new region, without spinning up a fresh ML pipeline, is exactly the kind of unglamorous efficiency gain that actually changes how teams work day to day.