How to Train Your Flare Prediction Model: Revisiting Robust Sampling of Rare Events

We present a case study of solar flare forecasting by means metadata feature time series, treating it as prominent class-imbalance and temporally coherent problem. Taking full advantage pre-flare series in active regions is made possible via the Space Weather Analytics for Solar Flares (SWAN-SF) benchmark dataset; partitioned collection multivariate region properties comprising 4075 spanning over 9 years Dynamics Observatory (SDO) period operations. showcase general concept temporal coherence triggered demand continuity show that lack proper understanding this effect may spuriously enhance models' performance. further address another well-known challenge rare event prediction, namely, issue. The SWAN-SF an appropriate dataset this, with 60:1 imbalance ratio GOES M- X-class flares 800:1 against flare-quiet instances. revisit main remedies these challenges several experiments to illustrate exact impact each have on Moreover, we acknowledge some basic data manipulation tasks such normalization cross validation also performance -- discuss problems well. In framework review primary advantages disadvantages using true skill statistic Heidke score, two widely used verification metrics task. conclusion, advocate benefits vs. point-in-time forecasting, provided above are measurably quantitatively addressed.