Tune Capacity and Stop Early

More correction can become memorization

Important controls include number of boosting iterations, learning rate, leaf count or depth, minimum examples per leaf, and regularization. Early stopping monitors validation performance and stops adding trees once improvement stalls, preventing unnecessary compute and later overfitting.

iteration:        20   50   90   140
training loss:   .42  .29  .19  .12
validation loss: .45  .34  .31  .36  <- stop near 90

Boosted models can output excellent rankings but poorly calibrated probabilities. If decisions interpret scores literally as risk, assess calibration separately and fit any calibrator using held-out evidence.

Scenario: A fraud model's validation AUC improves while predicted 90% risks materialize only 55% of the time. Ranking improved, but probability meaning did not. The threshold policy must not assume calibration.

Note: Use histogram-based gradient boosting in this CPU-only track for efficient scikit-learn examples; no external boosting library is required.