LoBoost: Fast Model-Native Local Conformal Prediction for Gradient-Boosted Trees

A research team including Vagner Santos and Victor Coscrato has published LoBoost, a novel method for adding conformal prediction capabilities directly to gradient-boosted decision trees (GBDTs). GBDTs like XGBoost and LightGBM are dominant for tabular data but traditionally output only point predictions without uncertainty estimates. LoBoost solves this by cleverly reusing the trained ensemble's internal leaf structure to define multiscale calibration groups. Each data point is grouped with others that share similar paths through the first k trees, allowing for local residual quantile calibration that adapts to data heterogeneity without the computational overhead of fitting auxiliary models or complex data splitting schemes.

The technical innovation lies in LoBoost's model-native approach: it requires no retraining, no auxiliary nuisance models, and maintains the standard train/calibration split. Experiments across multiple datasets show it delivers prediction intervals with competitive quality while actually improving test Mean Squared Error (MSE) in most cases. Most impressively, it achieves calibration speedups of 10-100x compared to adaptive conformal methods that require fitting additional models. For data scientists using production GBDT models, this means adding reliable uncertainty quantification becomes computationally trivial rather than prohibitive, enabling safer deployment in risk-sensitive applications like finance and healthcare.