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

This article gives you a full tutorial on using the TimesFM 2.5 model for advanced time series forecasting. The workflow covers everything from initial setup and zero-shot prediction to strict backtesting and adding external factors like price and promotions. It ends with practical steps for scalable deployment in places like Google Colab.

Key takeaways

• The tutorial offers a complete workflow for TimesFM 2.5, from environment setup and synthetic data generation to model compilation and forecasting.
• It highlights strict evaluation through backtesting and a set of accuracy metrics for both point and probabilistic forecasts.
• Advanced features like adding external covariates (XReg) and finding anomalies through residual analysis are shown.
• The process includes performance benchmarking, sensitivity checks, and configuration tuning for the best hardware use.
• The whole pipeline is built for practical, scalable deployment, using GPU acceleration in Google Colab.

Model Setup, Configuration, and Initial Zero-Shot Forecasting

The setup starts by configuring the Colab runtime and installing TimesFM 2.5 with its PyTorch dependencies. The code finds available hardware, defaulting to CUDA if a GPU is there, and sets reproducible random seeds.

A synthetic multi-store retail dataset is created, covering 1200 days for six stores. Each series has a base level, linear trend, weekly and yearly seasonality, price elasticity, promotional lifts, holiday effects, temperature influences, and random noise.

Loading and Compiling the Model

The TimesFM 2.5 model is loaded from its pretrained checkpoint. A baseline forecast configuration is made, setting key parameters like max_context=1024, max_horizon=256, and per_core_batch_size=16. The model is compiled with this configuration.

For zero-shot forecasting, a target store’s sales series is chosen. The model runs inference on the historical context, making both point forecasts (mean predictions) and probabilistic forecasts across nine quantiles (0.1 to 0.9).

Comprehensive Forecast Evaluation and Backtesting

Forecast quality is checked using a set of metrics. These include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), symmetric Mean Absolute Percentage Error (sMAPE), Mean Absolute Scaled Error (MASE), pinball loss, and prediction-interval coverage. These metrics check both point forecast accuracy and the trustworthiness of probabilistic forecasts. To strictly test model performance across many time windows, the workflow does rolling-origin backtesting.

Performance Benchmarking and Sensitivity Analysis

The process tests batched inference for speed and checks context-length sensitivity to see how input size changes forecast accuracy. The model’s throughput is also benchmarked. Its parameters get tuned for the best performance on the available hardware, like changing the per_core_batch_size configuration. This full evaluation, mixing metric checks, backtesting, and system benchmarking, builds a practical sense of how to validate and deploy TimesFM for real forecasting jobs.

Advanced Feature Integration: Covariates and Anomaly Detection

The workflow adds external covariates using TimesFM 2.5’s XReg feature. This lets the model include factors like price and promotions directly in its forecasts, going beyond simple past sales data. The tutorial makes a synthetic retail dataset that has these exact variables to test this ability.

Anomaly Detection and Input Robustness

Anomaly detection happens by studying the forecast residuals—the gaps between predicted and real values—to spot outliers and odd patterns in the time series. The model’s input strength is also tested by checking its performance with noisy or irregular data inputs, seeing how stable it is under poor conditions.

Long-Horizon Forecasting

The tutorial runs long-horizon forecasting tests to check the model’s ability for extended future predictions. This involves setting the model with a specific horizon and checking its accuracy over longer future periods, a main part of the end-to-end validation process.

Practical Deployment and Scalable Workflow

The tutorial builds a reusable function, recompile(**overrides), which quickly recompiles the TimesFM 2.5 model with different configuration changes by blending new settings with a baseline configuration dictionary.

The whole process is structured as an end-to-end workflow. It starts by configuring the runtime environment, installing dependencies, and generating a realistic multi-store retail dataset with trend, seasonality, pricing, promotions, holidays, and temperature effects. The workflow then loads and compiles the model, performs forecasting and evaluation, and tests features like batched inference, backtesting, and covariate addition.

For scalable deployment, the workflow is shown inside the Google Colab environment, using its available computing power. The code finds and uses a GPU if present, printing its name and available memory to confirm the hardware acceleration is active.

By moving through these stages—data generation, model configuration, validation, and benchmarking—the tutorial builds a practical sense of how to set up, validate, benchmark, and deploy TimesFM 2.5 for real forecasting jobs in a scalable way.