multivariate time series forecasting foundation model

A multivariate time series forecasting foundation model is an AI model that ingests several time-varying data streams at once—sales, visitor counts, weather, and so on—and calculates how they interact with one another to produce future predictions.

It's similar to a weather forecaster who doesn't predict tomorrow's weather from temperature alone, but also looks at humidity, wind, and seasonal events. Take an ice cream shop as an example: a model that only looks at past sales records can find a day-of-week pattern like "sales are strong every weekend," but if you also feed in the upcoming promotion schedule, it can predict something like "sales will be higher than usual this Tuesday because there's a promotion."

The word "foundation" here means the model doesn't need to be retrained from scratch for a single retailer or a single industry. Because it has already been trained on broad, diverse data in advance, it can be applied to time series forecasting in different fields—retail, finance, healthcare, and more—without major modification.

In articles, you'll see phrasing like "Google Research unveiled TimesFM-3, a multivariate time series forecasting foundation model." A common misconception is assuming that simply feeding in more related data automatically improves accuracy. In reality, the benefit only shows up when you add signals that are genuinely related to the outcome—like a promotion schedule tied to sales—not just any additional data.