Google open-sources TimesFM

Most forecasting work starts with a long column of dated numbers: yesterday’s sales, last week’s power use, last month’s shipments. The hard part is guessing the next values without being fooled by holidays, trends, or sudden spikes. (research.google) For years, companies usually solved that by building a custom model for each dataset. A retailer might train one model for store demand, while a utility trained another for electricity load, because older methods did not travel well from one series to another. (research.google) Google’s bet with TimesFM was that forecasting could work more like language models. Instead of training from scratch on one company’s data, it pre-trained a single model on 100 billion real-world time points so it could make zero-shot forecasts on new data it had never seen before. (research.google) That “zero-shot” phrase just means you hand the model your past numbers and ask for the future, without first tuning it on your own history. In the original Google Research write-up, TimesFM had about 200 million parameters and came close to top supervised methods on several public benchmarks even though those rivals were trained directly on each benchmark. (research.google) Now Google has pushed that idea into two places people can actually use. The TimesFM code is public on GitHub, where Google Research says the latest open version is TimesFM 2.5, and Google Cloud has also turned TimesFM into a built-in forecasting model inside BigQuery. (github.com ) (cloud.google.com) The BigQuery piece is what changes the day-to-day workflow for data teams. Google’s documentation says you can call `AI.FORECAST` with the built-in TimesFM model and generate forecasts without creating or training your own model, which removes a lot of the usual model-management work. (cloud.google.com) Google says the BigQuery version is available in all BigQuery-supported regions. Its tutorial shows the model forecasting hourly bike-share trips with a 720-step horizon and a 95% prediction interval directly from a Structured Query Language query. (cloud.google.com 1) (cloud.google.com 2) The open-source version gives teams a different path. The GitHub repository lets developers run inference with public checkpoints outside Google Cloud, while the same repository also warns that the open version is “not an officially supported Google product,” which is an important distinction for production users. (github.com) Google has been widening the model since the first paper. The repository notes that TimesFM 2.5 added back covariate support through a feature called XReg, which means the model can use side information alongside the raw history instead of only reading one bare sequence of past values. (github.com) So the real shift is not that forecasting suddenly became easy in every case. It is that a team with sales, energy, traffic, or inventory data can now try a pre-trained model in minutes inside a database they already use, and only build something custom if that first pass is not good enough. (cloud.google.com 1) (cloud.google.com 2)