Google's New Forecasting Model Crushes Competition with Zero-Shot Generalizations

Google has launched TimesFM-3, a powerful time-series forecasting model that outperforms its predecessors and competitors. The model is trained on over a trillion real-world and synthetic data time points and can handle multiple time series with zero-shot generalizations. This means it can forecast related time series in parallel and include historical data like past foot traffic.

In benchmarks, TimesFM-3 outperforms other models like Chronos-2 from Amazon and Moirai 2.0 from Salesforce, often by a significant margin. The team also notes that the field is developing rapidly, as seen in the recent launch of Datadog's Toto 2.0 model.

The architecture of TimesFM-3 involves two alternating attention layers: one looks backward within a single series and keeps things strictly causal, while the other looks sideways across all series at a given moment. This allows the model to capture complex relationships between time series.