Reliable inference of exoplanet light curve parameters using deterministic and stochastic systematics models

TL;DR

This paper evaluates the reliability of different systematics models in exoplanet light curve analysis, advocating for marginalisation and stochastic models like Gaussian processes to improve parameter inference accuracy.

Contribution

It introduces marginalisation over multiple systematics models and discusses stochastic models as robust alternatives for exoplanet light curve analysis.

Findings

Model selection criteria often fail to identify the correct systematics model.

Marginalisation over models better accounts for uncertainties in systematics.

Gaussian processes reliably recover transit parameters across various systematics.

Abstract

Time-series photometry and spectroscopy of transiting exoplanets allow us to study their atmospheres. Unfortunately, the required precision to extract atmospheric information surpasses the design specifications of most general purpose instrumentation, resulting in instrumental systematics in the light curves that are typically larger than the target precision. Systematics must therefore be modelled, leaving the inference of light curve parameters conditioned on the subjective choice of models and model selection criteria. This paper aims to test the reliability of the most commonly used systematics models and model selection criteria. As we are primarily interested in recovering light curve parameters rather than the favoured systematics model, marginalisation over systematics models is introduced as a more robust alternative than simple model selection. This can incorporate uncertainties in the choice of systematics model into the error budget as well as the model parameters. Its use is demonstrated using a series of simulated transit light curves. Stochastic models, specifically Gaussian processes, are also discussed in the context of marginalisation over systematics models, and are found to reliably recover the transit parameters for a wide range of systematics functions. None of the tested model selection criteria - including the BIC - routinely recovered the correct model. This means that commonly used methods that are based on simple model selection may underestimate the uncertainties when extracting transmission and eclipse spectra from real data, and low-significance claims using such techniques should be treated with caution. In general, no systematics modelling techniques are perfect; however, marginalisation over many systematics models helps to mitigate poor model selection, and stochastic processes provide an even more flexible approach to modelling instrumental systematics.