The AI company Mistral released Timestral on Oct. 15, its first forecasting models. Instead of predicting the next word in a sentence, they predict the next values in a time series, a stream of numbers recorded over time, such as a factory sensor's readings or a computer system's monitoring data. Mistral says forecasts like these drive decisions in energy, supply chains, risk and maintenance.
Traditional forecasting methods usually learn from one series at a time, using only that series' own history. Timestral was trained in advance on a large and varied mix of data, most of it made up by computer, and Mistral says it can forecast data it has never seen without first being trained on it. It can also weigh related readings together, such as a pump's vibration, temperature and pressure, and use things known ahead of time, such as a production schedule or a weather forecast. It gives a range of likely values rather than a single guess, so a team can see how uncertain the forecast is.
The models come in four sizes, from 5 million to 209 million parameters, the numbers a model tunes as it learns. Mistral says none of them needs a graphics chip, the costly kind most AI runs on, to run: on an ordinary server processor, forecasting a day ahead from three weeks of readings taken every 15 minutes takes about a hundredth of a second with the smallest model and about a seventh of a second with the largest.
Mistral says the largest model ranks first on three public tests used to compare forecasting models, among models that haven't been trained on the test data, while using two-thirds as many parameters as the previous leader. The margins are small: on one measure in its own table, its error is less than half a percent lower than the next model's. The files are free to download, but the license covers only people whose company or employer took in $20 million or less the month before. Bigger companies must ask Mistral for a commercial license, which it may refuse, or use the models through Mistral's own online services. Mistral has also published code so people can run the models on their own data. Companies can adapt the models to their own data through Mistral's training service, Forge, and, Mistral says, run them on their own hardware.
Mistral ran the tests itself. It says it has sent its results to the people who run two of the tests' public rankings, that a technical report with the full results will follow, and it has published code that lets others rerun its tests. Mistral itself notes that a company's real data can behave quite differently from public tests, and it has not said what its training service costs.