Forecasting Solar Flares with the EVEREST Transformer

Antanas Žilinskas, Robert Noel Shorten, Jakub Marecek

Presentation Attendance: Yes, we will present in-person

Keywords: Time-series forecasting, Solar flare prediction, Rare-event forecasting, Transformer models, Uncertainty quantification, Extreme value theory

TL;DR: A transformer model for solar flare forecasting that jointly models rare-event uncertainty and tail risk to deliver calibrated, high-sensitivity predictions.

Abstract: Solar flare forecasting is a rare-event time-series problem characterised by severe class imbalance, long-range temporal dependencies, and the need for calibrated probabilities and tail-aware behaviour. We present EVEREST, a compact Transformer forecaster trained with auxiliary objectives that improve calibration and tail sensitivity while retaining a single-head inference path. EVEREST integrates (i) a single-query attention bottleneck, (ii) an evidential Normal–Inverse–Gamma head on logits, (iii) an extreme-value head based on Generalized Pareto exceedances, and (iv) a lightweight precursor head for anticipatory supervision. On SHARP–GOES, EVEREST achieves strong True Skill Statistic (TSS) and low Expected Calibration Error (ECE) without inference overhead.

Track: Research Track (max 4 pages)

Submission Number: 114