End-to-End Forecasting with TimesFM 2.5: Backtesting, Covariates, Anomaly Detection, and Scalable Colab Deployment

MarkTechPost · 1 month ago · 30

Google's TimesFM 2.5 model is demonstrated in an end-to-end time-series forecasting tutorial using a synthetic multi-store retail dataset with 1,200 days of data across 6 stores. The tutorial evaluates TimesFM's performance using metrics including MAE, RMSE, sMAPE, MASE, and pinball loss, with a 56-day forecast horizon and rolling-origin backtesting across 6 folds. Results show TimesFM beats seasonal naive baselines and enables batch inference across multiple series while supporting probabilistic quantile forecasts, covariate integration, anomaly detection, and uncertainty quantification through prediction intervals.