KSKasra Sartaee
All research & engineering

research / UCL / 2025–2026

Probabilistic forecasting of geomagnetic indices

Transformer and LSTM forecasters of the Dst index with calibrated probabilistic outputs, benchmarked against a physics-informed baseline.

OSF preregistration

PyTorch

Illustrative forecast fan chart of the Dst indexIllustrative data. 48 hours of observed Dst with a storm dropping to about minus 130 nanotesla and recovering, then a 24 hour forecast: a median with 50% and 90% predictive bands that widen with lead time.
fig.1 Forecast fan chart with the median and 50% / 90% predictive bandssimulated / illustrative

Question

How can geomagnetic forecasting communicate uncertainty as well as a predicted Dst value?

Why it matters

A predictive distribution describes a range of possible outcomes. Its usefulness depends on calibration and comparison with an appropriate baseline.

Method

  • Built Transformer and LSTM forecasters with probabilistic output heads.
  • Calibrated predicted uncertainty with temperature scaling.
  • Compared against a physics-informed hybrid baseline.

Findings and uncertainty

The work focuses on probabilistic outputs and their calibration. Quantitative benchmark results are not presented on this page.

Limitations

The figure is a simulation: synthetic observations, an illustrative forecast and nine correlated ensemble paths. It is accelerated at one simulated hour per three seconds. It is neither a live space-weather feed nor output from the trained research models. Its bands illustrate uncertainty; they do not report measured coverage.

Next questions

How well do calibration and forecast skill hold across different geomagnetic conditions? Published evaluation results would be needed to answer that question here.

Sources

The study design is preregistered on OSF. The registration is not yet public, so access may be restricted.