Ai Ml Timeseries

This is a production-focused forecasting skill that covers the full pipeline from EDA to deployment. It defaults to LightGBM for most tabular cases, brings in transformers (TimesFM, Chronos) for long horizons, and handles the annoying parts like temporal validation splits, point-in-time feature engineering, and horizon-wise evaluation. The guidance is opinionated in useful ways: always start with naive baselines, never use random splits, avoid MAPE when targets hit zero. Good for anyone moving past tutorials into actual forecast systems where leakage bugs and retraining cadence matter. The backtest patterns and fallback strategies show real operational experience.