Time Series Foundation Model
A Time Series Foundation Model is a large AI model trained on vast amounts of measurement sequences that can subsequently be used for many different forecasts. It aims to do for time series what language models do for text: pre-trained once, then immediately applicable to new data.
Many data in economics and technology are measurement sequences: values recorded at regular intervals. A city’s hourly electricity consumption is one such sequence, as is a supermarket’s daily revenue or a stock’s price per minute. Experts call such data sequences time series. A Time Series Foundation Model is a very large computational model that has learned from billions of such measurements how number sequences typically develop further. One can then show it a completely new sequence it has never seen during training, and it predicts the next values. The term foundation model describes exactly this purpose: a base that many users can reuse for very different tasks.
What changes for forecasting as a result
Until now, forecasting almost always worked like this: a separate model was built for each data sequence. A supermarket with 30,000 items might therefore need up to 30,000 small models in the extreme case, each individually adjusted and monitored. This costs the time of experts, who are not available in sufficient numbers at many companies. A foundation model ideally replaces this collection with a single system.
This is particularly valuable in so-called zero-shot use. This means: the model is applied to new data without any additional training. A newly opened store has no sales history from which a classical model could learn. A foundation model, however, has seen patterns from thousands of other stores and still delivers a usable estimate.
Behind this lies an economic argument. Forecasts determine inventory levels, staffing plans, and energy procurement. If a retail group predicts its demand a few percent more accurately, it saves millions in storage costs and discarded goods. That is why Google, Amazon, IBM, and Salesforce are working on their own models of this kind.
How measurements become learning material
Technically, these models are closely related to language models like ChatGPT. Both use the transformer architecture, a type of network particularly good at recognizing distant relationships within a sequence. In language, the sequence is a sentence; in time series, it is a chain of measurements. A language model guesses the next word, a time series model the next numerical value.
For this to work, the sequence is broken down. The models usually group several consecutive measurements into a chunk, technically called a patch. These chunks are for the model what word fragments are for a language model. In addition, the numbers are normalized, that is, brought to a comparable scale. Only this way can the same model work with electricity consumption in megawatts and with sales figures in units.
Training takes place on mixed datasets from traffic, energy, weather, finance, and industrial sensors, partly supplemented by artificially generated sequences. It is important to note the limits of this approach: the model recognizes patterns, it does not understand causes. An unexpected strike, a war, or a change in the law does not appear in any measurement sequence. It is precisely at such breaking points that even the best forecasting models fail.
Where such models are already computing
They are most visible in cloud services. Google offers its model TimesFM via its own platform, Amazon uses Chronos in its forecasting tools, and IBM released the Granite Time Series series. Several of these models are freely available and can be downloaded. With just a few hundred million parameters, they are significantly smaller than language models and therefore also run on modest hardware.
In everyday life, you encounter the results indirectly. When a weather app shows an hourly forecast, an electricity provider buys tomorrow’s consumption, or an online shop estimates delivery times, time series forecasts are behind it. Predictive maintenance of machines also belongs here: sensors report temperature and vibration, and the model warns before a bearing breaks.
In financial news, these models often appear with exaggerated expectations. A common misconception is that they can reliably predict stock prices. However, prices contain an extreme amount of randomness, and studies show only minor advantages over simple methods in this area. The models are strong where regularities prevail, such as daily and annual rhythms in energy consumption.