Nixtla/hierarchicalforecast

Probabilistic Hierarchical forecasting 👑 with statistical and econometric methods.

What it solves

It addresses the problem of inconsistency in time series forecasts when data is organized in hierarchies (such as geographical regions, product categories, or temporal scales like weeks and months). When forecasts are made independently for each level, the sums of the lower-level forecasts often do not match the totals of the higher-level forecasts. This library provides methods to reconcile these predictions so they are coherent across all levels.

How it works

The library takes "base forecasts" (initial predictions) and applies reconciliation algorithms to adjust them. It supports several approaches:

  • Classic Methods: BottomUp (summing lower levels to create upper levels) and TopDown (distributing top-level forecasts downward).
  • Alternative Methods: MiddleOut (anchoring predictions at a middle level), MinTrace (minimizing total forecast variance), and ERM (using L1 regularization).
  • Probabilistic Methods: Tools like Normality, Bootstrap, PERMBU, and Conformal prediction to provide distribution-free or assumption-based prediction intervals.
  • Temporal Reconciliation: Most of these methods can also be applied to temporal hierarchies (e.g., reconciling daily forecasts with monthly totals).

Who it’s for

Data scientists, econometricians, and ML engineers working with hierarchical time series data in industry or academia who need consistent, coherent predictions across different aggregation levels.

Highlights

  • Supports both cross-sectional and temporal hierarchies.
  • Includes a wide range of statistical and econometric reconciliation methods.
  • Provides probabilistic coherent prediction methods for uncertainty estimation.
  • Integrates with the Nixtla ecosystem (e.g., statsforecast and datasetsforecast).

Related

  • Project
    deel-ai/puncc

    A Python library for predictive uncertainty calibration and conformalization, providing rigorous uncertainty estimations for any machine learning model across various tasks.

  • Project
    skforecast/skforecast

    A Python library for time series forecasting that integrates scikit-learn compatible models, statistical methods, and foundation models into a unified forecasting workflow.

  • Project
    sktime/pytorch-forecasting

    A PyTorch-based library for time series forecasting that provides a high-level API and state-of-the-art deep learning architectures like Temporal Fusion Transformers and N-BEATS.

  • Project
    sktime/sktime

    A Python library providing a unified interface for time series analysis, including forecasting, classification, and anomaly detection, compatible with the scikit-learn ecosystem.

  • Project
    TorchSpatiotemporal/tsl

    A PyTorch-based library for neural spatiotemporal forecasting that provides a unified framework for data preprocessing, model prototyping, and Graph Neural Network implementation.