2nd Place Solution: No Explicit Physics Modeling

This presentation describes the 2nd winning solution for the Ariel Data Challenge 2025. The method uses linear regression with features based on transit depth approximations at different phases of the light curve. Transit boundaries are detected through second derivative analysis, and polynomial fitting is applied to out-of-transit regions for detrending. Features include depth estimates from averaged AIRS frequencies, adjacent frequency windows, and the FG1 channel, with calculations of average, mid-transit, and percentile depths. Slope features capture transit wall steepness, and a separate model handles edge-case outliers. Uncertainty estimation employs bootstrapping combined with validation errors. The approach shows that careful feature engineering with polynomial approximations can effectively model transit characteristics for exoplanet atmospheric retrieval.