7th Place Solution: A Hybrid Physics-Based Method with Neural and Tree-Based Model Corrections

The solution is proposed by Horikita Saku and Takaito. We propose a hybrid physics-based approach that combines physical modeling with neural and tree-based corrections. Starting from a physically inspired baseline that models transit signals and wavelength-dependent variations, we build initial predictions using polynomial fitting. A neural network model then performs residual correction, refining the predictions and estimating uncertainty through quantile regression. Finally, a gradient boosting Decision Trees model adjusts sigma scales to account for sample-level noise bias. This three-stage design — physics foundation, neural refinement, and boosting calibration — balances interpretability and performance. Pseudo labeling is applied as a final enhancement to stabilize training and improve leaderboard results. Our method demonstrates that integrating physical priors with modern machine learning yields both accuracy and scientific consistency.