Chronos-2, Amazon's time-series foundation model, requires fine-tuning when zero-shot performance falls short. The model struggles with unfamiliar data patterns, systematic errors, or misaligned objectives. This second installment builds on a building electricity-demand case study, demonstrating how targeted adaptation can significantly improve forecasting accuracy.

Low-Rank Adaptation (LoRA) enables efficient fine-tuning without updating all 120M parameters. Instead of modifying full weight matrices, LoRA learns small adapter matrices that nudge model behavior in specific directions. This approach reduces GPU memory usage, creates smaller checkpoints, and minimizes overfitting risk on limited datasets.

The article walks through five fine-tuning scenarios using hourly electricity data from eight commercial buildings. Single-building adaptation focuses on individual assets, while portfolio fine-tuning pools fleet-wide history. Covariate-informed approaches leverage known-future signals like temperature. Researchers also test combining covariates with fleet data, plus held-out transfer for unseen assets.

Each scenario uses the same base model with identical LoRA configuration, varying only training data. The 168-hour forecast horizon demonstrates practical applicability for real-world energy management. This hands-on approach provides developers with a replicable template for adapting time-series foundation models to their specific domains.