pytorch-forecasting

Why we included this project

This library puts a solid set of deep learning forecasters, among them Temporal Fusion Transformer, DeepAR, N-BEATS and N-HiTS, behind a single high-level API built on PyTorch Lightning. The TimeSeriesDataSet class handles the fiddly parts of forecasting data, such as group IDs, static and time-varying covariates, encoder and decoder lengths, and train/validation splits, so a raw table can become a trained model without hand-rolling data pipelines. Quantile-based losses produce genuine uncertainty intervals rather than just point forecasts, which is what planning and inventory decisions usually need. Since the models plug into Lightning's trainer, GPU scaling, checkpointing, early stopping, and hyperparameter tuning with Optuna come almost for free. That makes it a reasonable base for teams that want one codebase for experiments and production forecasting instead of stitching together ad-hoc scripts.