Google Cloud Previews TimesFM 3.0 in BigQuery: Multivariate Time-Series Forecasting in SQL

What happened

Google Cloud announced that TimesFM 3.0 is now available as a preview in BigQuery, its data analytics service. Preview means a trial stage before official release. Google Cloud says the model was developed by Google Research and has 330 million parameters. It is a foundation model (a large, pre-trained model that can be used as is) meant to let users run multivariate time-series forecasting directly in BigQuery with SQL, the language used to query databases. A time series is data arranged in chronological order, such as daily sales or ride counts.

Google Cloud says BigQuery's forecasting function, AI.FORECAST, was first powered by TimesFM 2.0 and later upgraded to TimesFM 2.5, so users can generate forecasts in SQL without training or managing a model themselves. However, up to 2.5 these models supported only univariate forecasting, that is, predicting the future from the history of a single time series.

Google Cloud points out that most real-world forecasting problems are multivariate: related targets and external factors such as weather, promotions and holidays jointly influence future outcomes.

How the versions differ

In the table below, a "covariate" is a factor that is not the forecast target but affects the forecast, such as weather, promotions or holidays. Some covariates have only past records, while others have future values known in advance.

Google Cloud also says TimesFM 3.0 ranks first on three public time-series leaderboards, GIFT-Eval, FEV-Bench and TIME, and achieves the best zero-shot accuracy on covariate and multivariate forecasting tasks. Zero-shot means the model makes predictions without first being trained on the user's own data. This ranking is Google Cloud's own claim.

How it works in practice

Google Cloud uses the public New York Citi Bike dataset as an example, forecasting daily ride counts over the next seven days for commuters (subscriber_trips) and casual riders (casual_trips).

According to Google Cloud, BigQuery takes a single input table that contains both the rows of historical observations and the rows for the future forecast horizon. In the future rows, future covariates known in advance are filled in, while the target columns to be forecast and covariates with only past values are left empty (set to NULL).

• Target columns (target_cols): the values to forecast; several related targets can be forecast together.
• Past covariates (past_covariate_cols): factors with only historical records that cannot be known in advance, such as average ride duration.
• Future covariates (future_covariate_cols): factors known in advance, such as weekends, holidays or planned promotions.

Google Cloud explains that AI.FORECAST returns a forecast value for each target column, along with the upper and lower bounds of the prediction interval, meaning the range in which the actual value may fall. Google Cloud also says users can make forecasting requests in natural language in BigQuery Conversational Analytics, which automatically generates and runs the corresponding AI.FORECAST query and returns a summary, interactive charts with prediction intervals, and key insights.

What it means for users

For data teams using BigQuery, Google Cloud says existing univariate forecasting and anomaly detection (finding unusual values in data) workflows can switch to the new model simply by specifying model => 'TimesFM 3.0' in the query. Google Cloud also mentions that AI.EVALUATE can likewise be used for univariate time-series evaluation.

Google Cloud stresses that native multivariate support lets BigQuery handle complex enterprise data without moving data, building custom feature-engineering pipelines (processes that prepare raw data into columns a model can use) or hosting models yourself. The feature is still in preview.

FAQ

Browse the latest news in this topic

• Tech news