A Causal, Data-Driven Approach to Modeling the Kepler Data

TL;DR

The paper introduces the Causal Pixel Model (CPM), a data-driven, pixel-level approach for modeling Kepler data that effectively captures spacecraft variability while preserving transit signals, improving exoplanet detection over existing methods.

Contribution

The novel CPM method models Kepler pixel data using causal, data-driven techniques to better isolate planetary transits from spacecraft and stellar variability.

Findings

CPM outperforms Kepler PDC in exoplanet detection tasks.

The method effectively captures spacecraft-induced variability.

Hyper-parameters are optimized via cross-validation for broad applicability.

Abstract

Astronomical observations are affected by several kinds of noise, each with its own causal source; there is photon noise, stochastic source variability, and residuals coming from imperfect calibration of the detector or telescope. The precision of NASA Kepler photometry for exoplanet science---the most precise photometric measurements of stars ever made---appears to be limited by unknown or untracked variations in spacecraft pointing and temperature, and unmodeled stellar variability. Here we present the Causal Pixel Model (CPM) for Kepler data, a data-driven model intended to capture variability but preserve transit signals. The CPM works at the pixel level so that it can capture very fine-grained information about the variation of the spacecraft. The CPM predicts each target pixel value from a large number of pixels of other stars sharing the instrument variabilities while not containing any information on possible transits in the target star. In addition, we use the target star's future and past (auto-regression). By appropriately separating, for each data point, the data into training and test sets, we ensure that information about any transit will be perfectly isolated from the model. The method has four hyper-parameters (the number of predictor stars, the auto-regressive window size, and two L2-regularization amplitudes for model components), which we set by cross-validation. We determine a generic set of hyper-parameters that works well for most of the stars and apply the method to a corresponding set of target stars. We find that we can consistently outperform (for the purposes of exoplanet detection) the Kepler Pre-search Data Conditioning (PDC) method for exoplanet discovery.