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

In this tutorial, we build an advanced end-to-end time-series forecasting workflow with TimesFM 2.5 . We begin by configuring the runtime, installing the required dependencies, detecting available hardware, and generating a realistic multi-store retail dataset with trend, seasonality, pricing, promotions, holidays, temperature effects, and random variation. We then load and compile the TimesFM 2.5 model, examine its forecast configuration, and use it for zero-shot point and probabilistic forecasting. As we progress, we evaluate forecast quality with metrics such as MAE, RMSE, sMAPE, MASE, pinball loss, and prediction-interval coverage, while also testing batched inference, rolling-origin backtesting, context-length sensitivity, covariate integration through XReg, anomaly detection, long-horizon forecasting, throughput tuning, and input robustness. By working through these stages, we develop a practical understanding of how we configure, validate, benchmark, and deploy TimesFM for realistic forecasting tasks.

We configure the Google Colab environment, install TimesFM and its supporting libraries, detect the available CPU or GPU, and initialize reproducible random seeds. We generate a realistic multi-store retail dataset containing trends, weekly and yearly seasonality, pricing effects, promotions, holidays, temperature variations, and random demand noise. We then load the TimesFM 2.5 model, define its baseline forecast configuration, compile it, and create a reusable function for changing model settings in later experiments.

We generate our first zero-shot forecast and visualize the historical data, actual observations, median predictions, and probabilistic quantile bands in a fan chart. We inspect the structure of the point and quantile outputs, define evaluation functions for MAE, RMSE, MAPE, sMAPE, MASE, pinball loss, and prediction-interval coverage, and compare TimesFM against simple forecasting baselines. We also forecast all store series in a single batch, calculate store-level performance metrics, and plot the resulting forecasts and uncertainty intervals across the complete retail dataset.

We perform rolling-origin backtesting across multiple historical cutoffs to evaluate TimesFM under several realistic forecasting periods instead of relying on one holdout window. We compare the model with a seasonal-naive baseline, summarize the average error metrics, calculate the improvement in MASE, and visualize performance across backtest folds. We then run a context-length ablation study to determine how different amounts of historical data influence forecast accuracy, interval coverage, and inference time.

We incorporate numerical, categorical, and static covariates into the forecasting workflow through TimesFM’s XReg functionality. We compare the xreg + timesfm and timesfm + xreg fusion modes against a univariate forecast while using price, temperature, promotion, holiday, weekday, region, and store information. We evaluate the competing approaches, identify the best-performing covariate mode, and visualize how its predictions respond to known future promotion periods.