Astrophysically robust systematics removal using variational inference: application to the first month of Kepler data

Space-based transit search missions such as Kepler are collecting large numbers of stellar light curves of unprecedented photometric precision and time coverage. However, before this scientific goldmine can be exploited fully, the data must be cleaned of instrumental artefacts. We present a new method to correct common-mode systematics in large ensembles of very high precision light curves. It is based on a Bayesian linear basis model and uses shrinkage priors for robustness, variational inference for speed, and a de-noising step based on empirical mode decomposition to prevent the introduction of spurious noise into the corrected light curves. After demonstrating the performance of our method on a synthetic dataset, we apply it to the first month of Kepler data. We compare the results, which are publicly available, to the output of the Kepler pipeline's pre-search data conditioning, and show that the two generally give similar results, but the light curves corrected using our approach have lower scatter, on average, on both long and short timescales. We finish by discussing some limitations of our method and outlining some avenues for further development. The trend-corrected data produced by our approach are publicly available