We are only just recently finding access to high-quality time series data for use in solar flare prediction. Many previous and current projects use point-in-time measurements. It is possible that a time-series approach will enable new progress on using machine learning for flare prediction. Here we will use a benchmark dataset, named Space Weather ANalytics for Solar Flares (SWAN-SF), released by Angryk et al. [A-1], and made entirely of multivariate time series, aiming to carry out an unbiased flare forecasting and hopefully set the above question to rest.

The SWAN-SF dataset is made of five partitions (see Fig. 1). These partitions are temporally non-overlapping and divided in such a way that each contains approximately an equal number of X- and M- class flares (see Fig. 2). The data points in this dataset are time series slices of physical (magnetic field) parameters extracted from the flaring and flare-quiet regions, in a sliding fashion. That is, for a particular flare with a unique id, k equal-length multivariate time series are collected from a fixed period of time in the history of that flare. This period is called an observation window, denoted by T_obs, and spans over 24 hours. Given that t_i indicates the starting point of the i-th slice of the multivariate time series, the i+1-th slice starts at t_i + τ , where T_obs = 8τ . [Xb]