New sequential testing method speeds up random forests 4x with 0.1% disagreement

Ensemble classifiers like random forests combine hundreds of decision trees via majority vote. More trees mean better accuracy but proportionally higher compute costs. In a new paper, Kalman and Moscovich treat this as a sequential testing problem: evaluate base models one by one, and stop as soon as the current majority is decisive enough. They formalize three notions of optimality for early-stopping, each minimizing the number of base models executed while keeping disagreement with the full ensemble bounded. For each setting, they show how to construct and solve a linear program efficiently, yielding provably optimal stopping rules.

Tests on real-world datasets from the UC Irvine repository and the Grinsztajn benchmarks show these methods deliver 4x speedups on most datasets while keeping disagreement at just 0.1%. That means near-identical predictions to the full ensemble, but at a quarter of the computation. For production systems running large ensembles at scale, this is a practical win: lower latency, reduced energy cost, and cheaper inference. The approach is model-agnostic and works for any binary ensemble classifier, not just random forests. The paper is available on arXiv with code links via CatalyzeX and Hugging Face, making it easy for practitioners to adopt.