Random-Forest wrapper feature selection can use out-of-bag (OOB) predictions as a built-in validation signal, but the criteria that decide each forward step accumulate OOB estimation noise. We propose TOP-Gain (Truncated Out-of-bag Probability Gain), which scores a candidate feature by the mean sample-level increase in the OOB probability of the true class, with decreases truncated at zero.

Under a simple noise model, truncation removes about two-thirds of the noise variance contributed by unaffected samples while preserving the signal, and it makes the ranking of candidate features markedly more reproducible. In a pre-registered benchmark on 26 OpenML-CC18 datasets, TOP-Gain achieved the best mean rank in area under the ROC curve (ROC-AUC), accuracy and Macro-F1 among 15 methods; it was significantly better than 13 of 14 baselines and matched the accuracy of VSURF with 21 times less single-core computation.

Within an identical wrapper, it selected significantly more stable feature sets than net-gain, Brier, log-loss and accuracy criteria without loss of accuracy. A confirmatory study on 24 further datasets replicated the advantage over all baselines.

The accuracy gains were broad rather than tied to specific dataset characteristics, and, as pre-registered, the stability gain grew as relevant features became sparser