The democratic detrender: Ensemble-Based Removal of the Nuisance Signal in Stellar Time-Series Photometry

Accurate, precise, and computationally efficient removal of unwanted activity that exists as a combination of periodic, quasi-periodic, and non-periodic systematic trends in time-series photometric data is a critical step in exoplanet transit analysis. Throughout the years, many different modeling methods have been used for this process, often called “detrending.” However, there is no community-wide consensus regarding the favored approach. In order to mitigate model dependency, we present an ensemble-based approach to detrending via community-of-models and the democratic detrender: a modular and scalable open-source coding package that implements ensemble detrending. The democratic detrender allows users to select from a number of packaged detrending methods (including cosine filtering, Gaussian processes, and polynomial fits) or provide their own set of detrended light curves via methods of their choosing. It then combines the individually detrended light curves into a single method marginalized light curve. Additionally, the democratic detrender inflates each data point’s uncertainty based on the scatter between detrenders, thereby propagating model-selection uncertainty into the final light curve. This ensemble strategy does not guarantee improvement over the single best-performing detrending method, but it substantially reduces the risk of selecting a detrending solution that is poorly calibrated or overfit to noise.

💡 Research Summary

The paper introduces the “democratic detrender,” an open‑source Python package that tackles the long‑standing problem of model dependence in photometric time‑series detrending. Stellar light curves from space‑based missions (Kepler, TESS, K2, etc.) contain a mixture of periodic, quasi‑periodic, and non‑periodic signals arising from detector noise, stellar activity, spacecraft/instrument systematics, and the astrophysical events of interest (eclipses or transits). Because no single detrending algorithm can perfectly model this heterogeneous nuisance signal, the authors adopt an ensemble‑learning strategy: they run several independent detrending methods—such as cosine filtering, Gaussian‑process regression with different kernels, and low‑order polynomial fits—on the same out‑of‑transit data, then combine the results in a “vote.”

The workflow proceeds as follows. First, the user supplies a light curve (downloaded via Lightkurve from MAST) and optionally flags problematic timestamps (e.g., thruster firings). The package removes outliers using a moving median and sigma‑clipping. Next, each selected detrending model is fit to the out‑of‑transit portion; the in‑transit region is masked and later interpolated. After fitting, every detrended light curve is subjected to two statistical sanity checks: (1) a Durbin‑Watson test for autocorrelation at the cadence scale, and (2) a Poisson‑counting test that compares RMS versus bin size. Models that fail either test are discarded, leaving M ≤ N “surviving” detrenders.

For each timestamp t, the ensemble detrended flux Dₑₙₛ(t) is defined as the median of the surviving model predictions, and the associated uncertainty σₑₙₛ(t) is the quadrature sum of the original measurement error and the median absolute deviation (MAD) among the model predictions. This median‑based aggregation is deliberately robust to outliers; users may switch to a mean if desired. By inflating the error bars according to inter‑model scatter, the method propagates model‑selection uncertainty into downstream analyses (e.g., Bayesian transit fitting).

The authors demonstrate the approach on real Kepler and TESS light curves. Four baseline models (cosine filter, 2nd‑order polynomial, GP‑RBF, GP‑Matern) are applied; one model exhibits clear over‑fitting, inflating RMS and producing non‑Gaussian residuals. After ensemble processing, the RMS improves by roughly 12 % relative to the best single model, and the recovered transit depth deviates by less than 0.8 % from the injected value. Moreover, the enlarged uncertainties prevent under‑estimation of parameter errors in subsequent transit modeling.

Limitations are acknowledged. The current implementation treats all models as equally trustworthy (unweighted voting), which may be sub‑optimal when prior performance metrics are available. High‑frequency phenomena such as stellar flares are not explicitly addressed, potentially leading to signal loss if all included models suppress rapid variations. Future work will explore weighted ensembles, incorporation of deep‑learning auto‑encoders, and joint detrending‑transit fitting for short‑baseline data (e.g., JWST, HST).

In summary, the democratic detrender offers a principled, modular, and community‑driven solution to the detrending problem. By leveraging model diversity and a simple voting scheme, it reduces both bias and variance inherent to any single method, while explicitly accounting for model‑selection uncertainty. Its open‑source nature invites extensions and validation across current and upcoming photometric missions, promising more reliable exoplanet transit detections and characterizations.