Time series foundation models learn from millions of series before they see your data, the way ChatGPT learned from millions of texts. What they are, how they differ from statistical models, machine learning and a forecasting agent, and how we think forecasting will evolve with them.
ChatGPT never read your emails or your documents, and it still understands them when you hand them over: it learned from a huge amount of text of every kind, and that’s enough to get started with yours. Time series foundation models do the same with demand. They learned from millions of series of sales, energy, traffic and weather, and they forecast your product’s sales without ever having been trained on your data.
Until recently, forecasting demand meant training models on each company’s history. These models change that logic, so it’s worth understanding what they are, where they help and where they don’t yet.
Foundation model
A forecasting model trained in advance on millions of time series of every kind, which forecasts a new series without being trained on it.
Three ways to forecast
The real difference between the three families is what data each one learns from.
Statistical models
- Learn from: the history of a single series, its level, its trend and its seasonality.
- Examples: exponential smoothing (ETS), ARIMA, Theta and, for intermittent demand, Croston and its variants.
- Pros: they’re robust and easy to explain, and fitting one is cheap.
- Cons: they only know what that one series tells them. A product with six months of sales has little to teach, and what the product next to it learned doesn’t help. At scale the time adds up too: you fit one per series, and the automatic versions try several configurations on each one to keep the best.
Machine learning
- Learns from: your whole catalog at once, plus variables such as price, promotions, calendar and holidays.
- Examples: LightGBM, XGBoost, CatBoost, and neural networks built for time series such as N-BEATS, N-HiTS or TFT.
- Pros: it learns patterns that many products share, like a promotion lifting a whole category or December behaving differently.
- Cons: it has to be trained on each company’s data, with the right variables and tuned parameters.
Foundation models
- Learn from: millions of series from outside your catalog (sales from all kinds of businesses, energy, traffic, weather, server metrics), once, and before they ever see your data.
- How they forecast: they read your product’s history as context and return a forecast without training anything on your data. This is called zero-shot, and it’s the part that looks most like ChatGPT.
- Examples: Chronos, TimesFM, Moirai, TimeGPT, Granite, TiRex and Toto, which was trained mostly on server metrics.
- They give a reasonable forecast from day one, even for a product with very little history, with nothing to train or tune. But serving them is hard. They have hundreds of millions of parameters, and running them in production across thousands of products takes GPUs and a technical team to keep them running.
What about a forecasting agent?
Both come from the world of AI, so they’re easy to mix up. A foundation model returns numbers; a forecasting agent runs the whole process: it gathers the data, decides which models to run, checks whether the result makes sense, adjusts it with what it knows about the business and explains it.
| Foundation model | Forecasting agent | |
|---|---|---|
| What it takes | Sales history | Data, instructions and business context |
| What it returns | The forecast and its ranges | Decisions: which model to use, what to adjust, what to watch |
| What it’s built on | A model trained on time series | A language model, like ChatGPT, that uses tools |
They complement each other: to an agent, a foundation model is one more tool, next to statistical and machine learning models.
What they can bring to demand forecasting
Pros:
- Products with little history: a launch or a new SKU has nothing of its own to learn from, and a foundation model brings what it learned from millions of series.
- A fast start: there’s nothing to train or tune to get a first forecast.
- Ranges, not just a number: most of them return quantiles, which is what you need to size safety stock.
Their limits:
- They don’t know your promotions, prices or stockouts, unless the model accepts those variables and you pass them in.
- They learned from other companies’ series, so they’ve never seen what’s particular about your business.
- They have hundreds of millions of parameters, and running them over tens of thousands of SKUs takes compute.
How we think forecasting will evolve under this new paradigm
This part is opinion: what we see coming, based on what’s already published.
- Foundation models join the game. They’re not here to replace anyone: they’re one more tool, most useful where history is short.
- Many more are coming. Several came out in the last year, and the newest, like Chronos 2, already accept external variables (price, promotions, calendar), which is what they most lack for demand.
- Fine-tuning on your data. Start from the pretrained model and adjust it a little with a company’s history. On public benchmarks, fine-tuned versions already beat the ones used out of the box.
- Statistical and machine learning models are here to stay, and they keep contributing. They’re cheap, explainable and good at what foundation models can’t see yet, like your promotions or your stockouts.
- The trick is still the same: have several tools and keep, for each product, whatever works best. What will matter most is how they’re combined and what’s done with the result. That’s where the agents from the previous section come in.
Our forecasting engine already combines around 1,000 model variants per series, trained on your own data, without you writing a single line of code. Starting today, foundation models compete as candidates against the rest of the models in Supplai. Everything runs on our infrastructure, so you don't have to worry about serving them, compute, or updating servers when new models come out. If you want to ride the next wave of forecasting and try these and the rest of our advanced models, book a pilot and let's talk :) trysupplai.com