Google has open-sourced TimesFM, a foundation model designed specifically for time series forecasting.
The tool is capable of predicting a wide range of time-dependent data, including stock prices, energy demand, weather patterns and retail sales.
TimesFM was pre-trained on a vast dataset of real-world time series information, drawn from multiple domains rather than a single industry.
A key feature of the model is that it works "out of the box" without requiring any fine-tuning.
Users simply supply historical data and the model returns predictions, skipping the lengthy training process traditionally required for forecasting tools.
From internal tool to open release
TimesFM was originally built for Google's own internal use, developed to handle forecasting tasks across the company's operations.
The technology giant decided to release it publicly after concluding that its performance was strong enough to benefit the wider research and business community.
The model is now intended for use by researchers, developers and businesses looking to build forecasting capabilities without starting from scratch.
A GPT moment for forecasting
Google has compared the significance of TimesFM in time series forecasting to the role GPT played as a foundation model for language.
Just as GPT learned patterns across enormous volumes of text and could then generalise to new tasks, TimesFM has learned patterns across vast amounts of temporal data and can apply them to series it has never seen before.
That approach marks a departure from traditional forecasting methods, which typically require a custom model to be trained from scratch for each new dataset and each new domain.
Under the older approach, a model built to forecast energy demand would be of little use for predicting retail sales without substantial retraining.
TimesFM's foundation model design aims to remove that barrier, allowing a single pre-trained system to generalise across sectors as varied as finance, weather and consumer demand.
Availability
The model is currently available for access on GitHub, where developers can download it and apply it to their own datasets.
Its release adds to a growing body of foundation models being made public by large technology companies, following a similar pattern to the open release of large language models in recent years.
For businesses, the appeal lies in the prospect of skipping the cost and complexity of building bespoke forecasting systems, instead plugging existing historical data directly into a pre-trained model.
Whether TimesFM lives up to the comparison with GPT will depend on how it performs against established forecasting benchmarks as researchers and companies begin testing it more widely.
