// the find
Nixtla/neuralforecast
Scalable and user friendly neural :brain: forecasting algorithms.
NeuralForecast is a PyTorch-based library from Nixtla that wraps 30+ neural time series models—NBEATS, NHITS, TFT, iTransformer, DeepAR, and more—behind a single sklearn-style fit/predict API. It's aimed at practitioners who want production-grade neural forecasting without stitching together paper implementations from scratch. Works best when paired with the rest of the Nixtla stack (StatsForecast, MLForecast, HierarchicalForecast).
The model breadth is genuinely useful: you get canonical implementations of NBEATS, NHITS, and TFT that are actually tested against benchmarks, not just ported from a paper appendix. The unified interface across univariate, multivariate, probabilistic, and hierarchical forecasting means you can swap models without touching your data pipeline. Built-in Ray/Optuna integration for hyperparameter search is production-ready rather than bolted on. The backward compatibility tests (stored checkpoints in tests/backward_comp/) show they care about not breaking saved models across versions.
Training speed on CPU is painful—these models expect a GPU and the docs don't surface that clearly enough for newcomers who try it on a laptop. The transformer models (Informer, Autoformer, FedFormer) are included more for completeness than because they reliably beat simpler baselines; the literature has largely moved on but the repo hasn't pruned them. TimeLLM requires a separate pretrained LLM download with no clear size or licensing guidance in the main README. The nbdev-based development workflow (notebooks as source of truth) makes reading the actual model code in neuralforecast/models/ less painful than contributing to it—PRs that touch core logic require running notebook pipelines most contributors won't have set up.