We study the prediction of solar flare size and time-to-flare using 38 features describing magnetic complexity of the photospheric magnetic field. This work uses support vector regression to formulate a mapping from the 38-dimensional feature space to a continuous-valued label vector representing flare size or time-to-flare. When we consider flaring regions only, we find an average error in estimating flare size of approximately half a \emph{geostationary operational environmental satellite} (\emph{GOES}) class. When we additionally consider non-flaring regions, we find an increased average error of approximately 3/4 a \emph{GOES} class. We also consider thresholding the regressed flare size for the experiment containing both flaring and non-flaring regions and find a true positive rate of 0.69 and a true negative rate of 0.86 for flare prediction. The results for both of these size regression experiments are consistent across a wide range of predictive time windows, indicating that the magnetic complexity features may be persistent in appearance long before flare activity. This is supported by our larger error rates of some 40 hr in the time-to-flare regression problem. The 38 magnetic complexity features considered here appear to have discriminative potential for flare size, but their persistence in time makes them less discriminative for the time-to-flare problem.

We study the photospheric magnetic field of ~2000 active regions in solar cycle 23 to search for parameters indicative of energy build-up and subsequent release as a solar flare. We extract three sets of parameters: snapshots in space and time- total flux, magnetic gradients, and neutral lines; evolution in time- flux evolution; structures at multiple size scales- wavelet analysis. This combines pattern recognition and classification techniques via a relevance vector machine to determine whether a region will flare. We consider classification performance using all 38 extracted features and several feature subsets. Classification performance is quantified using both the true positive rate and the true negative rate. Additionally, we compute the true skill score which provides an equal weighting to true positive rate and true negative rate and the Heidke skill score to allow comparison to other flare forecasting work. We obtain a true skill score of ~0.5 for any predictive time window in the range 2-24hr, with a TPR of ~0.8 and a TNR of ~0.7. These values do not appear to depend on the time window, although the Heidke skill score (<0.5) does. Features relating to snapshots of the distribution of magnetic gradients show the best predictive ability over all predictive time windows. Other gradient-related features and the instantaneous power at various wavelet scales also feature in the top five ranked features in predictive power. While the photospheric magnetic field governs the coronal non-potentiality (and likelihood of flaring), photospheric magnetic field alone is not sufficient to determine this uniquely. Furthermore we are only measuring proxies of the magnetic energy build up. We still lack observational details on why energy is released at any particular point in time. We may have discovered the natural limit of the accuracy of flare predictions from these large scale studies.