ReModels: Quantile Regression Averaging models

Grzegorz Zakrzewski, Kacper Skonieczka, Mikolaj Malkinski, Jacek Mandziuk

Abstract

Electricity price forecasts play a crucial role in making key business decisions within the electricity markets. A focal point in this domain are probabilistic predictions, which delineate future price values in a more comprehensive manner than simple point forecasts. The golden standard in probabilistic approaches to predict energy prices is the Quantile Regression Averaging (QRA) method. In this paper, we present a Python package that encompasses the implementation of QRA, along with modifications of this approach that have appeared in the literature over the past few years. The proposed package also facilitates the acquisition and preparation of data related to electricity markets, as well as the evaluation of model predictions.

Introduction

In the volatile electricity market, point forecasts are often insufficient for effective risk management. Decision-makers require probabilistic forecasts—prediction intervals that quantify uncertainty and account for potential price spikes. While Quantile Regression Averaging (QRA) is the established "gold standard" for this task, the field lacks transparency and reproducibility. Researchers frequently develop custom QRA modifications without sharing code or standardized data-processing steps, hindering comparative analysis. The ReModels Python package addresses this "reproducibility crisis" by providing a unified, open-source framework for QRA and its variants.

Methodology and Package Architecture

The ReModels package streamlines the entire forecasting pipeline, from data acquisition to probabilistic evaluation.

QRA Variants

The core of the package implements nine distinct QRA variants. These models utilize point forecasts from multiple sources as independent variables in a quantile regression to estimate specific percentiles of the future price distribution.

Overview of the nine QRA variants implemented in ReModels, showing the processing pipeline from point predictions to final quantile regression.

The implemented variants include:

  • Standard QRA: The baseline model.
  • Regularized (LQRA): Employs LASSO (L1) penalties for high-dimensional inputs.
  • Factor-based (FQRA, sFQRA): Uses Principal Component Analysis (PCA) for dimensionality reduction.
  • Smoothed (SQRA, SQRM): Incorporates kernel density estimation to ensure a convex cost function.
  • Averaged (QRM, FQRM): Averages point forecasts prior to regression.

Data Preparation and VSTs

A critical component of the ReModels workflow is the handling of extreme price outliers, or "spikes." The package includes eight Variance Stabilizing Transformations (VSTs) to normalize data before modeling.

Mathematical formulas for the eight Variance Stabilizing Transformations (VSTs) supported by the package.

These transformations, such as Box-Cox, Arcsinh, and Logit, allow researchers to systematically determine the most effective method for taming market-specific volatility.

Datasets and Evaluation

The package's utility is demonstrated using data from the German electricity market (ENTSO-E Transparency Platform), covering historical day-ahead prices and load forecasts from 2015 to 2017. ReModels includes a dedicated EntsoeApi class for automated data fetching and a PointModel class compatible with scikit-learn pipelines. Evaluation is conducted using specialized probabilistic metrics, including the Aggregate Pinball Score (APS) and Average Empirical Coverage (AEC).

Key Results

Illustrative comparisons on the German dataset highlight the performance of different variants. For a 50% expected coverage interval, the Smoothed Quantile Regression Averaging with averaging (SQRM) achieved 52.30% actual coverage, outperforming the base QRA (45.17%). At the 90% confidence level, SQRM reached 88.87% coverage, demonstrating high reliability for extreme price bounds. Results also indicate that calibration windows significantly impact error metrics like MAPE, emphasizing the importance of ReModels' automated testing tools.

Conclusion and Future Work

ReModels provides a standardized toolkit that catalyzes research in electricity price forecasting. While currently focused on linear QRA variants, future updates aim to include non-linear models such as Quantile Regression Forests and Neural Networks. Additionally, while optimized for energy markets, the framework's modularity allows for extension into other domains like wind power or nodal voltage forecasting.

Summary updated

Abstract

Electricity price forecasts play a crucial role in making key business decisions within the electricity markets. A focal point in this domain are probabilistic predictions, which delineate future price values in a more comprehensive manner than simple point forecasts. The golden standard in probabilistic approaches to predict energy prices is the Quantile Regression Averaging (QRA) method. In this paper, we present a Python package that encompasses the implementation of QRA, along with modifications of this approach that have appeared in the literature over the past few years. The proposed package also facilitates the acquisition and preparation of data related to electricity markets, as well as the evaluation of model predictions.
BibTeX
@inproceedings{zakrzewski2024remodels,
  author = {Grzegorz Zakrzewski and Kacper Skonieczka and Mikolaj Malkinski and Jacek Mandziuk},
  title = {{ReModels: Quantile Regression Averaging models}},
  booktitle = {SoftwareX},
  year = {2024},
  url = {https://openreview.net/forum?id=z8YKtHD3pi},
}
Figure