Empirical study on the performance of conformal forecasting methods on real-world data

R. Morales Arsenal, C. Bergmeir, L. Escot Mangas

Conformal prediction is a statistical framework that generates valid predictions with quantifiable uncertainty, ensuring reliability in fields like machine learning. By leveraging exchangeability, it remains statistically valid even in finite samples, producing prediction intervals under minimal assumptions, which is crucial for high-stakes decisions. Despite its advantages, conformal prediction faces several challenges, including the need for large calibration datasets and difficulties in scenarios with non-exchangeable data, such as time series forecasting or dynamically evolving datasets. In such cases, traditional conformal prediction methods require adaptations to handle temporal dependencies.This study investigates the impact of violating these assumptions on coverage and the width of conformal intervals, extending the analysis to a diverse set of time series with different characteristics, such as varying levels of autocorrelation, non-stationarity, and structural changes.

Keywords: Forecasting methods, uncertainty, exchangeability, machine learning

Temporary Series
June 13, 2025  11:00 AM
Sala VIP Jaume Morera i Galícia

P. Galeano San Miguel, D. Peña, R. S. Tsay

J. A. Martín Segura, C. Pérez López

A. Lazcano, P. Hidalgo, J. E. Sandubete Galán

J. Cara