chapter seven

7 Deterministic forecasting with TimesFM

 

This chapter covers

  • Exploring the architecture of TimesFM
  • Zero-shot forecasting with TimesFM
  • Predicting with exogenous features

All foundation models we’ve explored so far are probabilistic forecasting models, which output a future probability distribution for each step in the forecast horizon. This allows us to derive arbitrary quantiles and quantify the uncertainty of the outcome as prediction intervals.

Although this approach provides a more complete view of the future, a model’s output requires more processing steps. Also, we may be interested in only the point forecast, not the intervals. This approach is especially useful when a definitive forecast, rather than a range of possible values, is necessary for planning purposes.

TimesFM comes into play in this scenario. This model is a deterministic foundation forecasting model developed by Google Research [1]. Because it returns point forecasts, it can’t be used for anomaly detection, which relies on uncertainty quantification to label anomalies. Still, TimesFM is ideal for forecasting values only, not ranges of values.

7.1 Examining TimesFM

7.1.1 Architecture of TimesFM

7.1.2 Pretraining TimesFM

7.2 Forecasting with TimesFM

7.2.1 Zero-shot forecasting with TimesFM

7.2.2 Cross-validation with TimesFM

7.2.3 Forecasting with exogenous features

7.3 Fine-tuning TimesFM and anomaly detection

7.4 Next steps

Summary