Forecasting has always been the least glamorous, most useful corner of machine learning. How much stock will this store need next week? What will electricity demand be at 6pm? How many support tickets on Monday? For a decade the answer was the same recipe: take the specific series you care about, build a model for it, train it on its history, tune it by hand, and maintain it forever. One model per series, per problem, per team.

That recipe is quietly collapsing. A new class of time-series foundation models — pretrained once on billions of data points, then pointed at a series they have never seen — now produce forecasts, with zero training on your data, that match or beat the bespoke models teams spent months tuning. Amazon’s Chronos-2 is described plainly by its makers as delivering “state-of-the-art zero-shot forecasting that consistently beats tuned statistical models out of the box.” The generalist that came for language, then for medicine, has arrived at the spreadsheet.

For the rest of us: what “zero-shot forecasting” means

Two ideas, in plain terms.

Forecasting is predicting the next values of a sequence that moves over time — sales, demand, load, traffic. The traditional way to do it well was bespoke: your data, your model, trained and tuned specifically for your series. Accurate, but slow to build and expensive to maintain, and it doesn’t transfer — a model tuned for one store’s sales knows nothing about the next.

Zero-shot means the model does the task with no training on your specific data at all. You hand a foundation model your series — a history it has never seen — and it forecasts immediately, the way a large language model answers a question about a topic you never fine-tuned it on. No per-series training run. No tuning. Point it, and go.

Put those together and the radical part is clear: the entire per-dataset build-train-tune-maintain pipeline — the thing forecasting teams are — becomes a single call to a model somebody else pretrained.

One model, no training, versus one model per series

The shift isn’t subtle when you draw it. The specialist approach is a loop you run for every new series: collect its history, pick an architecture, train, validate, tune, deploy, monitor, retrain. The generalist approach is one pretrained model you call — the same model for demand, load, and tickets — that has never seen your data and forecasts anyway. The remarkable, repeatedly-measured result is that the second approach doesn’t just save effort; it frequently wins on accuracy against models built specifically for the series.

The category, not one model

This isn’t one lab’s trick, which is what makes it a real shift rather than a headline. It’s a competitive field:

  • Google TimesFM — a decoder-only model pretrained on roughly 100 billion real-world time points, with strong zero-shot performance across domains. It’s the one that opened the category.
  • Amazon Chronos-2 — shipped in five sizes from 9 million to 710 million parameters, handling univariate, multivariate, and covariate-informed forecasting, at 300+ forecasts per second on a single GPU. It currently leads the shared benchmark.
  • Salesforce Moirai-2 — trained on 27 billion observations across nine domains, handling any frequency, any number of variables, any horizon.

They are measured against each other on GIFT-Eval, an open benchmark of 97 tasks across 55 datasets built specifically to test zero-shot and universal forecasting. The leaderboard reshuffles month to month — Chronos-2 recently overtook TimesFM-2.5 — but the direction is one-way: the frontier of forecasting is now a general, pretrained model, not a bespoke one.

Notice the parameter counts. These aren’t trillion-parameter giants — Chronos-2’s largest is 710 million, TimesFM around 200 million. A model small enough to run hundreds of forecasts a second on one GPU is beating pipelines that took a team a quarter to build. The moat was never the size of the model.

Why this keeps happening

If this feels familiar, it should. It’s the same pattern I wrote about in medicine: the general-purpose model trained on everything beating the tool built for the specific job. Richard Sutton called it the bitter lesson — general methods that scale with data and compute tend, eventually, to overtake hand-crafted, domain-specific ones. Chess, then Go, then speech, then language, then medicine, and now the demand forecast in a retail planning system.

The lesson for anyone who owns a forecasting pipeline isn’t “fire the team.” It’s that the model was never the defensible part. Once a pretrained generalist matches your tuned model out of the box, your value moves up the stack: the quality of your data, the covariates only you have, the judgment about which forecast matters and what decision it drives. The commodity is the prediction. The moat is everything around it.

The honest caveat

Zero-shot foundation models are not magic, and the data-honesty rule cuts both ways. “Beats tuned models out of the box” is a benchmark statement about aggregate performance across many datasets — not a promise about your series. Domains with strong, idiosyncratic structure, rich covariates, or unusual regimes can still favour a tuned model, and the only way to know is to test the foundation model against your current pipeline on your own data before you retire anything. The category has clearly crossed the threshold where it belongs in every forecasting team’s evaluation. It has not reached the point where you take the win on faith.

But the burden of proof has flipped. For a decade you had to justify not building a bespoke model. Now you have to justify building one.

References