Machine Learning Results

Model comparison, feature importance and honest limitations

Model Performance Comparison

model target RMSE R2 / AUC
Linear Regression CCI 2.361 0.021
Linear Regression Retail 14.890 0.135
Logistic Regression CCI direction 0.469
Decision Tree CCI 2.326 0.046
Random Forest CCI 1.968 0.324
Random Forest (Tuned) CCI 1.923 0.359
Random Forest Retail 5.510 0.871
Gradient Boosting CCI 1.988 0.312
KNN Regression CCI 2.220 0.108
0.871
Random Forest Retail
Best CV R2 in notebook
0.324
Random Forest CCI
Moderate non-linear CCI signal
0.469
Logistic CCI Direction
AUC below random
5.96
Holdout Retail RMSE
R2=0.892, MAE=3.59

Regression and Classification Diagnostics

ML Figure 3: linear regression coefficients for CCI
ML Figure 3: linear regression coefficients for CCI
ML Figure 4: logistic regression confusion matrix
ML Figure 4: logistic regression confusion matrix
ML Figure 11: ROC curve diagnostic
ML Figure 11: ROC curve diagnostic

Feature Importance Figures

Random Forest feature importance for CCI
Random Forest feature importance for CCI
Random Forest feature importance for retail
Random Forest feature importance for retail

Model Comparison and Additional Model Figures

ML Figure 10: model comparison, RMSE and R2
ML Figure 10: model comparison, RMSE and R2
ML Figure 15: extended model comparison
ML Figure 15: extended model comparison
Decision tree visualization
Decision tree visualization
KNN k-selection
KNN k-selection

Clustering and Country Climate Profile Figures

ML Figure 7: K-Means elbow and silhouette choice
ML Figure 7: K-Means elbow and silhouette choice
ML Figure 8: cluster distribution by country
ML Figure 8: cluster distribution by country
ML Figure 9: hierarchical dendrogram of country climate profiles
ML Figure 9: hierarchical dendrogram of country climate profiles

Actual vs Predicted Diagnostic

The retail model sits much closer to the diagonal than linear baselines, but this should be read as predictive association, not causal weather effect.

Why Weak Models Matter

Linear regression and CCI direction classification are intentionally shown. Negative or weak results demonstrate that the project is testing claims rather than only displaying successful models.