HierarchicalForecast: A Reference Framework for Hierarchical Forecasting in Python (2207.03517v6)

Abstract: Large collections of time series data are commonly organized into structures with different levels of aggregation; examples include product and geographical groupings. It is often important to ensure that the forecasts are coherent so that the predicted values at disaggregate levels add up to the aggregate forecast. The growing interest of the Machine Learning community in hierarchical forecasting systems indicates that we are in a propitious moment to ensure that scientific endeavors are grounded on sound baselines. For this reason, we put forward the HierarchicalForecast library, which contains preprocessed publicly available datasets, evaluation metrics, and a compiled set of statistical baseline models. Our Python-based reference framework aims to bridge the gap between statistical and econometric modeling, and Machine Learning forecasting research. Code and documentation are available in https://github.com/Nixtla/hierarchicalforecast.

• The paper introduces a reference framework that integrates diverse forecasting methods to ensure coherent predictions across hierarchical time series levels.
• It deploys state-of-the-art techniques like BottomUp, TopDown, and advanced probabilistic methods to boost forecast accuracy and reliability.
• The framework supports reproducible research with minimal dependencies and robust evaluation tools, including sCRPS metrics and visualization aids.

HierarchicalForecast: A Comprehensive Framework for Hierarchical Forecasting in Python

The paper "HierarchicalForecast: A Reference Framework for Hierarchical Forecasting in Python" introduces a Python-based open-source library, designed to address the growing needs for coherent hierarchical time series forecasting in both academia and industry. This framework endeavors to unify statistical, econometric, and machine learning approaches within a consistent, robust set of baselines, thereby facilitating advancements in hierarchical forecasting methodologies.

Key Features and Contributions

The HierarchicalForecast library addresses several identified challenges in the field:

1. Coherent Forecasting: Hierarchical time series data often entail multiple levels of aggregation, necessitating coherent forecasts that respect these structures. The library assists by providing reconciliation measures to ensure coherence across hierarchical levels.
2. Comprehensive Method Set: The framework includes a wide array of state-of-the-art methods such as BottomUp, TopDown, MiddleOut, MinTrace, and ERM for point forecasting, along with advanced probabilistic forecasting techniques like PERMBU, NORMALITY, and BOOTSTRAP.
3. Minimal Dependencies: Leveraging efficient Python libraries like NumPy, Pandas, and sklearn, along with the statsforecast package, HierarchicalForecast ensures high performance and reduces reliance on statistical packages originally implemented in R.
4. Evaluation and Visualization: It offers extensive tools for evaluating forecast accuracy, including a variety of standard and probabilistic metrics like sCRPS and logarithmic scores, alongside visualization aids to better interpret hierarchical structures.
5. Dataset Accessibility: The library provides easily accessible preprocessed datasets with the accompanying metadata, supporting reproducible research and simplifying experimentation.

Evaluation and Benchmarking

The authors present rigorous evaluations of the proposed methods using a benchmark set of hierarchical time series datasets, including datasets like the Australian Labour monthly reports and daily Wiki2 article views. The results demonstrate the framework's capability to improve over previous models in terms of sCRPS, highlighting its potential for enhancing prediction reliability across various domains.

Implications and Future Prospects

The development of the HierarchicalForecast library marks a significant step forward in harmonizing the efforts of multiple forecasting disciplines within the Python ecosystem. By providing a comprehensive suite of tools and methods, it enables researchers to perform comparative analyses seamlessly and engage in innovative research with a robust baseline.

Looking ahead, this framework can stimulate further research into novel hierarchical forecasting algorithms, encouraging the development of models that seamlessly integrate statistical rigor with machine learning adaptability. Collaboration among the forecasting community can further refine the framework, enhancing its applicability to diverse, complex datasets.

In conclusion, the HierarchicalForecast library not only bridges the existing gap between statistical and machine learning approaches in hierarchical forecasting but also sets a solid foundation for future advancements in the field. It epitomizes the synthesis of computational efficiency, methodological comprehensiveness, and empirical rigor, offering researchers a compelling resource for developing, evaluating, and deploying hierarchical forecasting solutions.