Why XGBoost

• Strong performance across classification, regression and ranking tasks.
• Tree ensembles capture non-linear relationships and feature interactions.
• Works well with SHAP and other model-explanation methods.

Feature Engineering

• Use domain knowledge and time-derived features where relevant.
• Consider lagged features for time-dependent problems.
• Be cautious with transformations that make explanations harder to interpret.

Imbalanced Data

• Consider class weighting and suitable evaluation metrics.
• Apply resampling only within training folds to avoid data leakage.
• Evaluate on data that represents the intended production population.

Training

• Tune n_estimators, learning_rate, max_depth, subsample and column-sampling parameters.
• Use reg_alpha and reg_lambda for regularisation.
• Apply cross-validation and early stopping where appropriate.

Evaluation and Explanation

• Select metrics appropriate to the task and class balance.
• Compare feature importance with SHAP-based global and local explanations.
• Monitor model stability over time and avoid causal interpretations.