What techniques handle imbalanced data in financial classification?

Q: What techniques handle imbalanced data in financial classification?

Imbalance (<5% positive) causes most classifiers to default to majority class, since accuracy is trivially maximized. Four primary techniques:\n\n1. Resampling: oversample minority (random duplication or SMOTE), undersample majority (random removal or Tomek links, ENN).\n\n2. Class weighting: weight minority class more in loss function. Most sklearn and XGBoost models support class_weight='balanced' or scale_pos_weight.\n\n3. Algorithmic: ensemble methods like BalancedRandomForest, EasyEnsemble, or focal loss in neural nets.\n\n4. Threshold adjustment: train normally, then tune threshold on validation (see Q 2527).\n\nExample: Trevorden Bank has 220,000 transactions with 340 fraud (0.15% positive).\n\nBaseline logistic: AUC=0.71, but recall at 0.5 threshold = 3% (predicts all negative).\n\nSMOTE to 50/50: AUC=0.84, recall at 0.5 = 68%. But introduces synthetic overfitting, validation AUC drops to 0.78.\n\nClass weight 660:1: AUC=0.82, recall at 0.5 = 72%, precision 12%.\n\nXGBoost with scale_pos_weight=660 + threshold tuning: AUC=0.86, precision 38%, recall 71%. Best tradeoff.\n\nBalancedRandomForest: AUC=0.85, similar to XGBoost but faster training.\n\nTrevorden deploys XGBoost + threshold 0.27 as production. Monthly model monitoring: minority class prevalence can drift (fraud rate varies 0.12%-0.28% seasonally), so they retrain quarterly with updated scale_pos_weight. Critical rule: NEVER apply SMOTE or class weighting before train/test split — it causes leakage. Always resample only within training folds of cross-validation.

Question

AcadiFi · Accepted Answer

Imbalance (<5% positive) causes most classifiers to default to majority class, since accuracy is trivially maximized. Four primary techniques:

1. Resampling: oversample minority (random duplication or SMOTE), undersample majority (random removal or Tomek links, ENN).

2. Class weighting: weight minority class more in loss function. Most sklearn and XGBoost models support class_weight='balanced' or scale_pos_weight.

3. Algorithmic: ensemble methods like BalancedRandomForest, EasyEnsemble, or focal loss in neural nets.

4. Threshold adjustment: train normally, then tune threshold on validation (see Q 2527).

Example: Trevorden Bank has 220,000 transactions with 340 fraud (0.15% positive).

Baseline logistic: AUC=0.71, but recall at 0.5 threshold = 3% (predicts all negative).

SMOTE to 50/50: AUC=0.84, recall at 0.5 = 68%. But introduces synthetic overfitting, validation AUC drops to 0.78.

Class weight 660:1: AUC=0.82, recall at 0.5 = 72%, precision 12%.

XGBoost with scale_pos_weight=660 + threshold tuning: AUC=0.86, precision 38%, recall 71%. Best tradeoff.

BalancedRandomForest: AUC=0.85, similar to XGBoost but faster training.

Trevorden deploys XGBoost + threshold 0.27 as production. Monthly model monitoring: minority class prevalence can drift (fraud rate varies 0.12%-0.28% seasonally), so they retrain quarterly with updated scale_pos_weight. Critical rule: NEVER apply SMOTE or class weighting before train/test split — it causes leakage. Always resample only within training folds of cross-validation.

What techniques handle imbalanced data in financial classification?

Master Level II with our CFA Course

Related Questions

Practice Questions