neuralforecast: a user-friendly collection of state-of-the-art neural time-series forecasting models

What it solves

NeuralForecast addresses the difficulty of implementing state-of-the-art neural forecasting models. Many existing implementations are hard to use, computationally expensive, and often fail to outperform traditional statistical methods. This library provides a user-friendly, efficient, and robust collection of neural networks designed specifically for time-series forecasting.

How it works

It provides a unified interface (using a familiar .fit and .predict syntax similar to scikit-learn) to a wide array of neural forecasting architectures. These range from classic RNNs (like LSTM and GRU) and CNNs (TCN) to modern Transformer-based models (such as PatchTST and iTransformer) and specialized architectures like N-BEATS and NHITS. The library also supports exogenous variables, static covariates, and probabilistic forecasting through quantile losses and parametric distributions.

Who it’s for

Data scientists and machine learning engineers who need to perform high-accuracy time-series forecasting using deep learning models without the need to build these complex architectures from scratch.

Highlights

• Extensive Model Library: Includes over 30 state-of-the-art models, including official implementations of NHITS and NBEATSx.
• Usability: Uses a scikit-learn style API for easy integration into existing workflows.
• Advanced Forecasting: Supports probabilistic forecasting and the inclusion of exogenous regressors.
• Optimization: Integrates with Ray and Optuna for distributed automatic hyperparameter tuning.
• Transfer Learning: Ability to predict with minimal historical data using transfer learning techniques.