{
  "best_model": "XGBoost",
  "results": {
    "LogisticRegression": {
      "cv_auc": 0.9217,
      "cv_gini": 0.8434,
      "cv_brier": 0.1127,
      "holdout_auc": 0.9228,
      "holdout_gini": 0.8456,
      "holdout_brier": 0.1114
    },
    "XGBoost": {
      "cv_auc": 0.9226,
      "cv_gini": 0.8452,
      "cv_brier": 0.1113,
      "holdout_auc": 0.9241,
      "holdout_gini": 0.8483,
      "holdout_brier": 0.1098
    },
    "XGBoost_cal": {
      "cv_auc": 0.9225,
      "cv_gini": 0.8449,
      "cv_brier": 0.0897,
      "holdout_auc": 0.9243,
      "holdout_gini": 0.8486,
      "holdout_brier": 0.0883
    },
    "LightGBM": {
      "cv_auc": 0.9225,
      "cv_gini": 0.8449,
      "cv_brier": 0.1111,
      "holdout_auc": 0.9239,
      "holdout_gini": 0.8477,
      "holdout_brier": 0.1097
    },
    "LightGBM_cal": {
      "cv_auc": 0.9225,
      "cv_gini": 0.8451,
      "cv_brier": 0.0896,
      "holdout_auc": 0.9242,
      "holdout_gini": 0.8484,
      "holdout_brier": 0.0883
    }
  },
  "best_xgb_params": {
    "n_estimators": 563,
    "max_depth": 3,
    "learning_rate": 0.04448759145940924,
    "subsample": 0.5705311592018347,
    "colsample_bytree": 0.761636704321766,
    "min_child_weight": 6,
    "gamma": 0.8972162546507014,
    "reg_alpha": 2.9651828345346862,
    "reg_lambda": 4.2804546466141
  },
  "best_lgbm_params": {
    "n_estimators": 628,
    "num_leaves": 54,
    "max_depth": 3,
    "learning_rate": 0.04550221443019292,
    "subsample": 0.8675755613520427,
    "colsample_bytree": 0.5811729741100575,
    "min_child_samples": 32,
    "reg_alpha": 8.55907729876928,
    "reg_lambda": 3.292460209578426
  },
  "top_features": [
    {
      "feature": "affordability_surplus",
      "importance": 0.742099940776825
    },
    {
      "feature": "combined_risk",
      "importance": 0.5698344707489014
    },
    {
      "feature": "credit_score",
      "importance": 0.3508100211620331
    },
    {
      "feature": "expense_understatement_flag",
      "importance": 0.324170857667923
    },
    {
      "feature": "worst_status_12m",
      "importance": 0.21787478029727936
    },
    {
      "feature": "affordability_margin",
      "importance": 0.20726099610328674
    },
    {
      "feature": "worst_status_3m",
      "importance": 0.18635587394237518
    },
    {
      "feature": "debit_order_returns_3m",
      "importance": 0.1527206152677536
    },
    {
      "feature": "affordability_surplus_ratio",
      "importance": 0.1525910347700119
    },
    {
      "feature": "age",
      "importance": 0.11965403705835342
    }
  ],
  "scale_pos_weight": 3.26,
  "n_train_samples": 133108,
  "n_holdout_samples": 33278,
  "train_default_rate": 0.2346,
  "holdout_default_rate": 0.2346,
  "excluded_proxy_score_features": [
    "origination_score",
    "client_management_score",
    "early_stage_cure_score",
    "late_stage_recovery_score",
    "score_band"
  ],
  "excluded_policy_rule_features": [
    "nca_affordability_pass"
  ],
  "policy_rule": {
    "name": "nca_affordability_pass",
    "description": "Applicants failing affordability are declined outside the model."
  },
  "feature_rationale": {
    "gross_income": "Primary repayment capacity indicator.",
    "net_income": "Disposable income after statutory deductions.",
    "discretionary_income": "Cash remaining after living expenses and debt obligations.",
    "debt_service_ratio": "Debt obligations relative to net income; core affordability measure.",
    "living_to_income": "Share of income absorbed by living expenses.",
    "discretionary_ratio": "Free-cash-flow ratio after living expenses.",
    "gross_to_net_ratio": "Relationship between gross and take-home pay.",
    "combined_risk": "Simple interaction between credit score risk and debt burden.",
    "credit_score": "Core bureau-style payment-risk signal.",
    "score_norm": "Normalised credit score for model stability.",
    "score_risk": "Inverse credit score representation.",
    "living_expenses": "Declared household expense burden.",
    "food_ratio": "Food spend relative to income.",
    "housing_ratio": "Housing spend relative to income.",
    "transport_ratio": "Transport spend relative to income.",
    "food_expense": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "transport_expense": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "accommodation_expense": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "utilities_expense": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "medical_expense": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "credit_bureau_monthly_debt": "Estimated third-party debt repayment obligation.",
    "bureau_open_accounts": "Credit exposure breadth.",
    "bureau_utilisation": "Credit utilisation stress indicator.",
    "minimum_living_expense": "Normative affordability floor.",
    "affordability_expense_basis": "Verified expense basis used for affordability.",
    "affordability_surplus": "Residual affordability after verified expenses and debt.",
    "affordability_margin": "Residual affordability as a share of income.",
    "expense_understatement_flag": "Flags declared expenses that appear understated.",
    "bureau_debt_to_income": "Bureau debt burden ratio.",
    "affordability_surplus_ratio": "Residual affordability scaled by income.",
    "affordability_basis_ratio": "Verified expense basis scaled by income.",
    "expense_gap_ratio": "Gap between verified and declared expenses.",
    "norm_to_declared_ratio": "Norm-based living expense pressure.",
    "predicted_monthly_income": "Alternative income estimate for consistency checking.",
    "household_gross_income": "Broader household affordability signal.",
    "cpa_commitments": "Estimated committed deductions / payment obligations.",
    "nlr_commitments": "Non-loan recurring commitments proxy.",
    "contactability_index": "Collections and operational contactability proxy.",
    "likely_payment_day": "Likely point in the month when repayment cash flow arrives.",
    "payment_gap_days": "Expected delay between due date and payment timing.",
    "income_prediction_gap": "Gap between declared and estimated income.",
    "household_to_income_ratio": "Household support buffer relative to applicant income.",
    "cpa_to_income_ratio": "Committed deductions pressure on income.",
    "nlr_to_income_ratio": "Recurring commitments pressure on income.",
    "contactability_risk_gap": "Composite operational risk based on contactability and utilisation.",
    "monthly_instalment": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "loan_amount_requested": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "loan_term_months": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "worst_status_3m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "worst_status_12m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "worst_status_24m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "times_30_dpd_12m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "times_60_dpd_12m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "times_90_dpd_12m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "months_since_worst": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "consecutive_on_time": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "paid_up_accounts": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "written_off_accounts": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "income_variance_pct": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "credit_history_months": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "no_credit_history_flag": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "salary_consistency_3m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "debit_order_returns_3m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "debit_order_returns_12m": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "cash_withdrawals_pct": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "gambling_transactions_flag": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "garnishee_order_flag": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "legal_action_flag": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "fraud_indicator_score": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "instalment_to_income": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "debt_to_income_multiple": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "assets_to_income_multiple": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "avg_balance_to_income": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "min_balance_to_income": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "credit_history_years": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "age": "Lifecycle and employment stability proxy.",
    "income_band": "Segment view of affordability.",
    "salary_frequency": "Repayment timing alignment proxy.",
    "employment_type": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "education_level": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "dependants": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "residence_type": "Feature retained because it contributes to credit risk segmentation and can be explained operationally.",
    "loan_product_type": "Feature retained because it contributes to credit risk segmentation and can be explained operationally."
  },
  "adverse_action_reasons": [
    {
      "feature": "norm_to_declared_ratio",
      "coefficient": 0.6329,
      "reason": "Elevated risk contribution from norm to declared ratio."
    },
    {
      "feature": "expense_gap_ratio",
      "coefficient": 0.488,
      "reason": "Elevated risk contribution from expense gap ratio."
    },
    {
      "feature": "affordability_basis_ratio",
      "coefficient": 0.4851,
      "reason": "Elevated risk contribution from affordability basis ratio."
    },
    {
      "feature": "employment_type_unemployed",
      "coefficient": 0.4754,
      "reason": "Elevated risk contribution from employment type unemployed."
    },
    {
      "feature": "expense_understatement_flag",
      "coefficient": 0.4322,
      "reason": "Declared expenses appear understated versus verified expense basis."
    },
    {
      "feature": "instalment_to_income",
      "coefficient": 0.4282,
      "reason": "Elevated risk contribution from instalment to income."
    },
    {
      "feature": "debt_service_ratio",
      "coefficient": 0.4001,
      "reason": "High debt obligations relative to income."
    },
    {
      "feature": "combined_risk",
      "coefficient": 0.3837,
      "reason": "Elevated risk contribution from combined risk."
    }
  ],
  "fairness_summary": {
    "gender": [
      {
        "group": "Female",
        "sample_size": 16547,
        "observed_default_rate": 0.2337,
        "mean_predicted_pd": 0.35120001435279846,
        "approval_rate": 0.3619,
        "auc": 0.9242
      },
      {
        "group": "Male",
        "sample_size": 16731,
        "observed_default_rate": 0.2356,
        "mean_predicted_pd": 0.35010001063346863,
        "approval_rate": 0.3655,
        "auc": 0.9241
      }
    ],
    "province": [
      {
        "group": "Eastern Cape",
        "sample_size": 3522,
        "observed_default_rate": 0.2785,
        "mean_predicted_pd": 0.3783000111579895,
        "approval_rate": 0.3339,
        "auc": 0.9252
      },
      {
        "group": "Free State",
        "sample_size": 2116,
        "observed_default_rate": 0.2585,
        "mean_predicted_pd": 0.3671000003814697,
        "approval_rate": 0.3327,
        "auc": 0.9161
      },
      {
        "group": "Gauteng",
        "sample_size": 8758,
        "observed_default_rate": 0.1955,
        "mean_predicted_pd": 0.3158999979496002,
        "approval_rate": 0.4064,
        "auc": 0.927
      },
      {
        "group": "KwaZulu-Natal",
        "sample_size": 5303,
        "observed_default_rate": 0.2285,
        "mean_predicted_pd": 0.3465999960899353,
        "approval_rate": 0.3696,
        "auc": 0.9221
      },
      {
        "group": "Limpopo",
        "sample_size": 2918,
        "observed_default_rate": 0.2834,
        "mean_predicted_pd": 0.4041000008583069,
        "approval_rate": 0.3036,
        "auc": 0.9278
      },
      {
        "group": "Mpumalanga",
        "sample_size": 2456,
        "observed_default_rate": 0.2492,
        "mean_predicted_pd": 0.3691999912261963,
        "approval_rate": 0.3359,
        "auc": 0.9208
      },
      {
        "group": "North West",
        "sample_size": 2095,
        "observed_default_rate": 0.2625,
        "mean_predicted_pd": 0.3747999966144562,
        "approval_rate": 0.3403,
        "auc": 0.9184
      },
      {
        "group": "Northern Cape",
        "sample_size": 1749,
        "observed_default_rate": 0.2813,
        "mean_predicted_pd": 0.3968000113964081,
        "approval_rate": 0.3145,
        "auc": 0.9188
      },
      {
        "group": "Western Cape",
        "sample_size": 4361,
        "observed_default_rate": 0.2006,
        "mean_predicted_pd": 0.3190000057220459,
        "approval_rate": 0.3972,
        "auc": 0.9228
      }
    ],
    "marital_status": [
      {
        "group": "Divorced",
        "sample_size": 8471,
        "observed_default_rate": 0.2375,
        "mean_predicted_pd": 0.349700003862381,
        "approval_rate": 0.3726,
        "auc": 0.9259
      },
      {
        "group": "Married",
        "sample_size": 8084,
        "observed_default_rate": 0.2381,
        "mean_predicted_pd": 0.3571000099182129,
        "approval_rate": 0.358,
        "auc": 0.9251
      },
      {
        "group": "Single",
        "sample_size": 8458,
        "observed_default_rate": 0.2319,
        "mean_predicted_pd": 0.3481000065803528,
        "approval_rate": 0.3574,
        "auc": 0.922
      },
      {
        "group": "Widowed",
        "sample_size": 8265,
        "observed_default_rate": 0.2311,
        "mean_predicted_pd": 0.3481000065803528,
        "approval_rate": 0.3668,
        "auc": 0.9235
      }
    ],
    "age_band": [
      {
        "group": "18-25",
        "sample_size": 4186,
        "observed_default_rate": 0.489,
        "mean_predicted_pd": 0.6173999905586243,
        "approval_rate": 0.1285,
        "auc": 0.9294
      },
      {
        "group": "26-35",
        "sample_size": 10747,
        "observed_default_rate": 0.2934,
        "mean_predicted_pd": 0.41530001163482666,
        "approval_rate": 0.2871,
        "auc": 0.929
      },
      {
        "group": "36-45",
        "sample_size": 11946,
        "observed_default_rate": 0.1549,
        "mean_predicted_pd": 0.27079999446868896,
        "approval_rate": 0.434,
        "auc": 0.895
      },
      {
        "group": "46-55",
        "sample_size": 5376,
        "observed_default_rate": 0.1112,
        "mean_predicted_pd": 0.21119999885559082,
        "approval_rate": 0.5227,
        "auc": 0.9013
      },
      {
        "group": "56-69",
        "sample_size": 1023,
        "observed_default_rate": 0.1554,
        "mean_predicted_pd": 0.24629999697208405,
        "approval_rate": 0.477,
        "auc": 0.8965
      }
    ]
  },
  "logistic_top_coefficients": [
    {
      "feature": "num__norm_to_declared_ratio",
      "coefficient": 0.632876030698465,
      "abs_coefficient": 0.632876030698465
    },
    {
      "feature": "num__affordability_surplus_ratio",
      "coefficient": -0.566326020731039,
      "abs_coefficient": 0.566326020731039
    },
    {
      "feature": "num__affordability_margin",
      "coefficient": -0.5633530644186399,
      "abs_coefficient": 0.5633530644186399
    },
    {
      "feature": "num__expense_gap_ratio",
      "coefficient": 0.4880477657244633,
      "abs_coefficient": 0.4880477657244633
    },
    {
      "feature": "num__affordability_basis_ratio",
      "coefficient": 0.4850638030279616,
      "abs_coefficient": 0.4850638030279616
    },
    {
      "feature": "num__discretionary_ratio",
      "coefficient": -0.47951119867144354,
      "abs_coefficient": 0.47951119867144354
    },
    {
      "feature": "cat__employment_type_unemployed",
      "coefficient": 0.4753633249003756,
      "abs_coefficient": 0.4753633249003756
    },
    {
      "feature": "num__expense_understatement_flag",
      "coefficient": 0.43216524370028375,
      "abs_coefficient": 0.43216524370028375
    },
    {
      "feature": "num__instalment_to_income",
      "coefficient": 0.4281508092747665,
      "abs_coefficient": 0.4281508092747665
    },
    {
      "feature": "num__debt_service_ratio",
      "coefficient": 0.4000660671070147,
      "abs_coefficient": 0.4000660671070147
    },
    {
      "feature": "num__combined_risk",
      "coefficient": 0.38374024083432967,
      "abs_coefficient": 0.38374024083432967
    },
    {
      "feature": "num__minimum_living_expense",
      "coefficient": -0.2978228162072123,
      "abs_coefficient": 0.2978228162072123
    },
    {
      "feature": "num__affordability_surplus",
      "coefficient": 0.2972818784366496,
      "abs_coefficient": 0.2972818784366496
    },
    {
      "feature": "num__nlr_to_income_ratio",
      "coefficient": -0.27900972083655295,
      "abs_coefficient": 0.27900972083655295
    },
    {
      "feature": "num__written_off_accounts",
      "coefficient": 0.27178552357998814,
      "abs_coefficient": 0.27178552357998814
    },
    {
      "feature": "num__debit_order_returns_3m",
      "coefficient": 0.2567028708006428,
      "abs_coefficient": 0.2567028708006428
    },
    {
      "feature": "num__worst_status_12m",
      "coefficient": 0.25415366744494733,
      "abs_coefficient": 0.25415366744494733
    },
    {
      "feature": "num__monthly_instalment",
      "coefficient": -0.2516098365573718,
      "abs_coefficient": 0.2516098365573718
    },
    {
      "feature": "num__age",
      "coefficient": -0.24622515365377046,
      "abs_coefficient": 0.24622515365377046
    },
    {
      "feature": "cat__employment_type_student",
      "coefficient": -0.2195824956393162,
      "abs_coefficient": 0.2195824956393162
    }
  ],
  "runtime_seconds": 5245.32,
  "runtime_human": "87m 25s"
}