Ensemble Conformalized Quantile Regression for Probabilistic Time Series Forecasting.

Area of Science:

• Time Series Analysis
• Machine Learning
• Statistical Forecasting

Background:

• Traditional forecasting methods struggle with nonstationary and heteroscedastic time series.
• Existing conformal prediction (CP) and quantile regression (QR) techniques have limitations for complex data.

Purpose of the Study:

• To introduce a novel probabilistic forecasting method, ensemble conformalized quantile regression (EnCQR).
• To develop distribution-free and valid prediction intervals (PIs) for nonstationary and heteroscedastic time series.
• To enhance forecasting accuracy and PI informativeness.

Main Methods:

• EnCQR utilizes a bootstrap ensemble of quantile regression learners.
• It employs conformal prediction principles adapted for time series by removing exchangeability requirements.
• The method is model-agnostic and can be applied to various forecasting models, including deep learning.

Main Results:

• EnCQR generates prediction intervals that are distribution-free and approximately marginally valid.
• The method effectively handles time series with varying degrees of heteroscedasticity.
• EnCQR demonstrates superior performance compared to standalone QR or CP methods, yielding sharper and more informative PIs.

Conclusions:

• EnCQR is a robust and versatile probabilistic forecasting method for challenging time series data.
• The ensemble approach overcomes limitations of traditional CP for time series.
• EnCQR provides a significant advancement in generating reliable and informative prediction intervals.