01Why it matters
A single number that moves satellites and power grids
F10.7 is the solar radio flux at a 10.7 cm wavelength, measured daily in Penticton, Canada, in solar flux units. When it rises, the upper atmosphere expands and drags satellites down, geomagnetic storms push currents into transformers, the ionosphere scrambles GPS and aviation signals, and particle radiation flips bits in satellite electronics. The costs are not hypothetical: the 1989 Quebec blackout left 6 million people without power for 9 hours, the 2003 Halloween storms destroyed the ADEOS-II satellite, and in 2022 a Starlink launch lost 40 satellites to re-entry within days.
- U.S. electricity losses estimated at $1–20 billion per extreme event scenario
- U.S. satellite industry losses estimated at $2–80 billion per extreme event scenario
- Source: Abt Associates (2017) report for NOAA
02The gap
Even the official forecast misses when it matters most
In May 2024 the Gannon G5 storm, the strongest in two decades, hit Earth. NOAA's official outlook under-forecast the peak day by 58 sfu, and over the pre-peak week (May 3–10) its RMSE was 66.3 sfu. Wonjun framed a question: can financial-engineering methods, built to forecast markets from their own history, be combined with AI to do better? He then made it testable on three independent axes: statistical significance against an ablated baseline, operational superiority over NOAA's real archived forecasts, and preservation of the Sun's physical spectral signature.
03Borrowing from Wall Street
Treating solar flux like a price chart
Traders compute moving averages, Bollinger Bands, RSI and MACD from price history to anticipate momentum and reversals. Wonjun computed the same 12 indicators on F10.7. Because many of them overlap (SMA20 and EMA20 correlate at 0.97, and the Bollinger midline duplicates a moving average), he compressed them with PCA into 5 to 7 orthogonal components explaining at least 95% of the variance. He also switched the target to a log-return: instead of predicting tomorrow's flux, the model predicts its rate of change, which is zero-centred and stable whether the Sun is quiet or stormy.
- Log-difference alone cut baseline LSTM RMSE from 15.77 to 10.56 sfu
- PCA fit on training data only, so nothing leaks from the future
04Physics as a prior
Letting eighty years of solar physics do some of the work
Earlier studies decomposed F10.7 with data-driven methods such as EMD and VMD, whose modes have no physical meaning. Wonjun instead subtracted four periods that solar physicists already know: the 11-year Schwabe cycle, the 27.275-day Carrington rotation, its 13.638-day half, and the 155-day Rieger period. The neural network then only has to learn the residual, focusing its capacity on the genuinely unpredictable part of the signal.
- Schwabe cycle: 11 years
- Carrington rotation: 27.275 days
- Half-Carrington: 13.638 days
- Rieger period: 155 days
05The experiment
315 models, one winner
To prove each ingredient mattered, he built a 21-variant ablation grid: LSTM, GRU and Transformer architectures, raw-level vs log-return targets, baseline vs PCA features, across horizons of 1, 3, 7 and 27 days, plus Fourier hybrids, a persistence model and ARIMA. Each variant was trained with three random seeds over five walk-forward splits, 315 runs and about 100 hours of consumer GPU time, with fixed hyperparameters and no tuning loop. The best model reached an RMSE of 9.704 sfu, 38.5% better than the plain LSTM baseline.
06Head to head
Beating NOAA on its own archive
Academic papers usually compare against other neural networks. Wonjun went further and digitised 260 of NOAA SWPC's archived 27-day outlook reports, producing 6,961 forecast pairs from 2020 to 2024. Across 2024, at the fair 7-day horizon, his model's RMSE of 35.48 sfu beat NOAA's 38.70 sfu by 8%; his 1-day model reached 15.28 sfu. Inside the 29-day Gannon storm window, the 7-day model beat NOAA by 20% (32.96 vs 41.40 sfu), and the 1-day model cut error by 74% to 10.89 sfu.
- Diebold–Mariano (HLN) tests, combined with Stouffer's Z, p ≤ 0.002
- −24% RMSE versus NOAA across the 5-year archive
07Honest uncertainty
How sure is the forecast?
A forecast without an error bar is hard to act on. Wonjun compared three ways of building 90% prediction intervals. MC Dropout (one seed, 50 stochastic passes) covered only 32.3% of real outcomes; a three-seed Deep Ensemble reached 44.5%. Adding split conformal calibration, which widens intervals using the 90th-percentile residual on validation data, lifted coverage to 75.9%. That is a 43.6-point gain, yet still 14.1 points short of the nominal 90%, a shortfall he reports openly rather than hides.
08Opening the black box
Accurate for the right reasons
Four explainability lenses checked whether the model was meaningfully accurate, not just numerically close. Permutation importance showed the most recent five days of F10.7 dominate, with financial indicators and Carrington phase contributing, and the ranking was identical across all three seeds. Integrated Gradients revealed the model shifts attention with horizon and with quiet vs storm conditions. A spectral check confirmed the four solar periods survive in the predictions with no spurious frequencies, and SSIM showed storm rise, peak and decay shapes are preserved.
09What's next
Limits, and the road to a full forecasting system
Accuracy plateaus around 45 days, a limit that helioseismic far-side imaging and CME or flare event flags may lower but chaos ultimately bounds. MC Dropout coverage drops at solar minimum, so regime-aware conformal calibration will need the next minimum (around 2031) to validate. Transfer to sunspot number reached R² = 0.78 without native training, and a full-cycle reanalysis from 1976 would need roughly ten times the GPU hours. Downstream, F10.7 forecasts tracked yearly total solar irradiance at R² = 0.85, linking the work to climate-model solar forcing.
