Models & methods
Zero-shot forecasting with foundation models
Zero-shot means you get a forecast without training on your series first. A foundation model uses what it learned from many other series, then reads yours. Fine-tuning is the later step, once you have enough of your own data to make it worth the spend.
Updated Aug 15, 2026·8 min read
What zero-shot actually means
Zero-shot means you get a forecast without a training job on that series first. A foundation model uses weights it learned from many other series, then reads yours at inference. Your history is input. It is not “the model never sees data.” It is “the model was not optimized on your labels.”
Pretrain once · infer many
When it shines
Short history. Many related series. A new region that looks like old ones. The first path on a Tuesday afternoon when the alternative is a blank cell. Cold start, with a modern engine.
It can also win on series you thought were “too small to model” because the prior carries season and spike shapes you did not have time to hand-code.
Where it stumbles
A long, quirky series with a promo calendar the pretraining corpus under-represents. A measurement that is actually a delayed accounting date. Anything where a simple seasonal naive is already excellent — zero-shot can lose to last week, and that is allowed. You find out on the folds, not on a keynote slide.
When to fine-tune
When zero-shot is close but systematically off, you have enough of your own signal, and a backtest says the extra job is worth the spend. Fine-tuning can overfit a short workspace as easily as training from scratch. One pass, a holdout, a baseline. “We fine-tuned” is not a metric.
FAQ
- What is zero-shot forecasting?
- Producing a forecast for a series the model was not trained on, using a pretrained foundation model. You pass history (and sometimes covariates) at inference time. No local training job required.
- Is zero-shot as accurate as a trained model?
- On short or related series, it can win. On a long, quirky series with a private promo calendar, a locally trained model often catches up. Put both in the same backtest and stop guessing.
- Does zero-shot use my data to train other customers' models?
- It should not. Inference uses your series as input. Training other customers' models on your workspace is a policy question — on Predict.ai, workspace data stays in the workspace.
Keep going
Guide
Forecasting new products and locations
A cold start is a forecast with almost no history of its own. You borrow shape from similar items or places, use leading signals, keep the horizon short, and replace the borrowed prior as soon as real sales show up.
Guide
Forecasting models compared
Classical models are interpretable and cheap. Machine learning eats more drivers. Foundation models arrive pretrained and can go zero-shot. None of them win every series — which is why serious stacks keep a tournament, not a favorite.
Guide
How much history do you need to forecast?
Enough to see the pattern you are betting on. A weekly seasonal business wants more than one year; an hourly series can learn a week of shape faster. Quality of the target beats another dusty decade of the wrong grain.
Use case
Digital traffic forecasting
Forecast traffic, transactions, and concurrency before campaigns, launches, and seasonal peaks.
Use case
Retail demand forecasting
Anticipate demand by product, store, channel, and region before buying or allocation decisions are locked.