In March 2024 we tested the first forecasting foundation models on a real daily sales series. Our verdict was clear: not there yet. A classic method tracked the daily data more closely than all three new models.
This week Mistral released Timestral, its first family of forecasting models. Google's TimesFM is now in its third version. Both promise what the 2024 models could not yet deliver: hand them a sales history and get a usable forecast, with no training per retailer.
So we ran the test again. This time on grocery data, at store and product level, and against the kind of classic methods many planning systems still rely on.
The answer has changed. The new models beat every classic method we ran, on both datasets, without any training on the data. A small model came within half a percent of the largest one at about a fifth of the computing time. And one of them, TimesFM 3.0, got clearly better again when we gave it planned discounts and promotions.
Key findings
Accuracy: Timestral and TimesFM made 4.5 to 6.3% less forecast error than ETS, a classic statistical method, on Walmart and FreshRetailNet grocery data, with no training on that data.
Speed: Timestral 28M was within 0.4% of the 209M model's accuracy at one fifth of the time, and faster than ETS.
Promotions: with planned discounts and promotions, TimesFM 3.0 made 7.8% less error than ETS on fresh food. Timestral showed no clear gain from that data.
Stock-outs: scoring only days without empty shelves did not change the ranking.
Licences: Timestral is free only below $20 million monthly revenue; TimesFM 3.0 is non-commercial; TimesFM 2.5 is Apache 2.0.
How we tested Timestral and TimesFM against classic forecasting
The test is simple. We stand on a date in the past, ask each model to forecast the next days, and compare with what was really sold. We do this for three periods in a row, so one lucky week cannot decide the result. Every model gets exactly the same products, stores and dates.
We used two open grocery datasets:
Walmart M5: daily food sales in 10 US stores. 300 store-product series, forecast 28 days ahead.
FreshRetailNet-50K: daily sales of fresh products in Chinese stores, with stock-out flags. 1,000 store-product series, forecast 7 days ahead.
We compared two groups:
New foundation models: Timestral in three sizes (5, 28 and 209 million parameters) and Google's TimesFM 2.5 and 3.0. Used as they come, with no training on this data.
Classic methods: ETS, a statistical method found in many planning tools, and a per-product regression on calendar, price and promotion data. As a floor: "same as last week".
How we score a forecast
A store does not order on one number. It orders on a range: how much will we probably sell, and how much on a busy day? So every model gave a range, from a low value (only a 10% chance of selling less) to a high one (only a 10% chance of selling more).
We scored that range with the standard quantile loss. A simple example:
A model says tomorrow's sales will be between 4 and 12 units.
The store sells 15, above the top of the range.
The model gets a heavy penalty: it said this was unlikely.
Had the store sold 8, the penalty would have been small.
We add up the penalties over all products and days. In the charts, every model is compared with "same as last week": 20% means 20% less error than that simple rule.
Result 1: AI foundation models beat classic forecasting methods on grocery data
Every green bar is longer than every grey bar, on both datasets. That held in each of the three test periods, not just on average.
Against ETS, Timestral 28M made 5.9% less error on Walmart and 5.5% less on fresh food. TimesFM 3.0 made 6.1% and 4.5% less.
With planned discounts and promotions, TimesFM 3.0 made 7.8% less error than ETS on fresh food, the best result in the test.
The per-product regression did worse than ETS, even with price and promotion data. More inputs did not make up for a weaker method.
Among the new models, the scores were close. No single model won everywhere.
One caveat: TimesFM 2.5 was trained on data that includes the Walmart set, so its Walmart bar is faded. We do not count it as a fair result.
Result 2: Timestral 28M is almost as accurate as the largest model, and five times faster
In 2024, speed was one of our objections: a single foundation-model forecast was about 100 times slower than a classic one. That has flipped.
Timestral 28M had only 0.4% more error than the 209M model, at about one fifth of the time.
The smallest model, 5M, gave up about 1% of accuracy and was 17 to 23 times faster than 209M.
ETS, the classic method, was both less accurate and slower than Timestral 28M: 48 against 7 milliseconds per forecast on Walmart, 21 against 1 on fresh food.
In an earlier run on a small cloud machine with two processor cores, Timestral 28M forecast all 50,000 FreshRetailNet store-product series in under two minutes.
Result 3: promotion data helps TimesFM 3.0, and stock-outs do not change the ranking
The second objection in 2024 was context. A model that only sees past sales does not know a promotion is coming. Both new models can now take that information.
Promotions. On fresh food, giving TimesFM 3.0 the planned discounts, promotion days and holidays cut its error by a further 3.5%. That gain was clear in every check we ran. Timestral, in this first release, showed no clear gain from the same data.
Empty shelves. On 43% of the fresh-food store-days, a product was out of stock for at least an hour. On those days, sales understate what customers wanted. So we also scored only the days without a stock-out. The ranking did not change: the new models still made 5 to 7% less error than ETS, and TimesFM 3.0 with promotion data 9% less.
One point still needs care. The models' high value should be exceeded on about one day in ten. On fresh food it was exceeded on 11 to 14% of days. The ranges are useful, but a little too narrow to set safety stock on directly.
What this means for demand planners and grocers
A good forecast is becoming a commodity. Two years ago, the new models could not yet beat a classic method. Today a ready-made model does better than the classic methods, out of the box, in milliseconds.
The value moves to the decision. A better forecast does not by itself mean fewer empty shelves or less waste. That depends on what you do with the range: how much to order, for which store, from which depot, on which truck. Those decisions affect each other, and that is where most margin is won or lost.
Plan on the range, not the single number. For slow movers, the middle value sits well below the average, because most days sell little and a few days sell a lot. Ordering on the middle value alone means regular gaps on the shelf.
Three questions to ask
Does the advantage hold on our own data? Run the same test on your own products, promotions and seasons before switching anything.
Does a better forecast give better decisions? Measure service level, waste and stock, not only forecast error.
Can we use it commercially? Check the licence of the exact model before building on it. More on that below.
Frequently asked questions
What is Timestral?
Timestral is a family of time series foundation models from Mistral AI, released on 15 October 2026 in four sizes (5M, 28M, 123M and 209M parameters). You give it a sales history and it returns a forecast range, without training on your data.
Is Timestral better than TimesFM for retail demand forecasting?
On our grocery tests they were close. With sales history alone, neither won everywhere. With promotion data, TimesFM 3.0 was the most accurate on fresh food, but its licence does not allow commercial use without Google's permission.
Do foundation models beat classic forecasting methods like ETS?
In our test, yes: every foundation model beat ETS and a per-product regression on both datasets and in every test period. Against ETS, the gain was 4.5 to 6.3% with sales history alone, and 7.8% for TimesFM 3.0 with promotion data on fresh food.
Can a grocer use Timestral commercially?
Only if the company's monthly revenue is below $20 million. Larger retailers need a commercial licence from Mistral or its hosted service.
Which model is the fastest?
Timestral 5M and 28M. Timestral 28M took about 1 to 7 milliseconds per forecast on a laptop processor, faster than ETS.
Before you build on it
Licences differ a lot. We read them in full on 16 October 2026; this is our reading, not legal advice.
Timestral (Modified MIT): free, unless your company had more than $20 million revenue in the previous month. Above that, a commercial licence from Mistral or its hosted service.
TimesFM 3.0 (non-commercial): research and testing only. Production use needs permission from Google.
TimesFM 2.5 (Apache 2.0): free for commercial use.
Limits of this test. Two public datasets and three test periods; results on a European retailer may differ. The Walmart sample is food only and balanced across slow and fast sellers. We compared with standard implementations of classic methods, not with any vendor's planning system. Planned prices and promotions were assumed to be known in advance. TimesFM 2.5 has seen the Walmart data in training, and Timestral does not publish its full training data, so we cannot rule out that a model saw part of a test set.
Method. Every model used the same series and dates; all forecasts were checked against the source data; differences were tested by resampling the series 10,000 times. Details and code are available on request.
Davio Larnout is CEO and co-founder of Superlinear.
