🧲 Tormenta

A machine-learning nowcast of the geomagnetic Kp index (0–9) from solar-wind measurements taken about an hour upstream of Earth. CT-378 Machine Learning, NED University, Fall 2026 · Alizeh & HRK.

Open the live app Model & model card Source code

Results on unseen years (2020 to Jan 2026)

0.573RMSE (Kp units)

0.808R²

34%lower error than persistence

80.8%coverage of the 80% range

Nowcast of the May 2024 storm compared with observed Kp — The May 2024 storm, the strongest in 20 years, which the model never saw in training. Predicted peak 8.6, observed 9.

How it works

Inputs: interplanetary magnetic field (|B|, Bx, By, Bz), solar-wind speed, density and temperature, and F10.7 solar flux. No geomagnetic indices are used, so the nowcast does not wait for ground magnetometers.
Features: 66 physics-based features: coupling functions (Newell, V·Bs, electric field), pressure, lags, 3–24 h trailing statistics, position inside the 3-hour Kp block, and time of day and year.
Model: scikit-learn HistGradientBoosting, chosen over Linear Regression, KNN, SVR and Random Forest by time-ordered cross-validation. Quantile models with a conformal correction give the 80% range.
Data: NASA OMNI2 hourly solar wind (1998–2025) with GFZ Potsdam definitive Kp as labels. Live inputs come from NOAA SWPC's propagated real-time solar wind.
Leakage safeguards: causal features only, the same feature code for training and live use (unit-tested), decisions made on validation years only, and the test years scored once per model version.

Model comparison (test years)

Model	RMSE	R²	RMSE on storm hours
Persistence (previous Kp)	0.862	0.566	1.368
Linear Regression	0.665	0.742	1.470
KNN	0.657	0.748	1.337
SVR	0.605	0.786	1.079
Random Forest	0.606	0.785	0.972
HistGradientBoosting	0.573	0.808	0.846

Error and bias by storm level for each model — Error by storm level. Every model under-predicts the largest storms, and HistGradientBoosting the least.

Live check

Live nowcast compared with GFZ and NOAA Kp over five days — Real-time snapshot, 25–30 Oct 2026: RMSE 0.59 against GFZ quick-look Kp, in line with the test error.

Limitations

The largest storms (G3+) are under-predicted by about 0.6 Kp on average, because they are rare in the training data.
Sudden storm onsets show up 1–2 hours late within a 3-hour Kp block.
Lead time is at most about 1 hour, the travel time of the solar wind to Earth.
For education and awareness only, not an operational warning service. Use NOAA SWPC or GFZ for that.