Submitted to arXiv on 9 October, the study combines Monte Carlo Tree Search with Generalized Nested Rollout Policy Adaptation to select a smaller subset from 1,072 engineered features. Candidate subsets are scored on out-of-sample validation performance rather than training error; the method can also use an objective that accounts for subset size and multicollinearity. The first page lists both authors at LAMSADE, Université Paris Dauphine-PSL and CNRS in Paris, France.

For an explanatory regression of weekly NASDAQ-100 returns covering January 2000 to March 2026, the authors report an out-of-sample R² of 0.6961 for their method and 0.6313 for an optimized Lasso comparison. The paper specifically describes the target as contemporaneously observed explanatory regression, so the result does not represent prospective return-forecasting performance. This is a preprint finding within a particular financial dataset and experimental protocol. The work proposes a search-based alternative for selecting interpretable sparse models when many correlated features are available.