XGBoost vs LightGBM: Speed Differs Sharply, Accuracy Barely at All

A controlled benchmark on 20,000 rows and 30 features gave XGBoost and LightGBM identical tuning budgets of 15 randomized search trials with 3-fold cross-validation. The two models finished just 0.00020 AUC apart on the test set, a gap small enough to be attributed to random seed variance. LightGBM completed the same tuning run 2.23 times faster than XGBoost, while XGBoost returned predictions 2.2 times faster at inference. Even when LightGBM was given its saved time back as extra search trials under a fixed 30-second wall-clock budget, it fitted 1.59 times more candidates yet gained no meaningful accuracy improvement. The findings suggest the practical choice between the two libraries should be driven by whether training speed or inference speed matters more, not by expected accuracy differences.