Week 8: The Week I Stopped Chasing Better Models and Started Asking Why

A data science learning journal entry documenting a shift from chasing better ML model performance to understanding model behavior. After training LightGBM and finding it only marginally better than Random Forest and XGBoost, the author turned to SHAP for model explainability. Comparing SHAP explanations across all three models revealed consistent feature importance (notably 4-hour RSI), building confidence in feature engineering. A stacking ensemble performed slightly worse on a held-out validation set, which the author frames as a valuable lesson about honest evaluation and generalization rather than a failure.