Editing AI for Risk Management in Banking (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''Credit risk model with SHAP explainability for regulatory compliance:'''
<syntaxhighlight lang="python">
import pandas as pd
import numpy as np
from sklearn.model_selection import TimeSeriesSplit
from sklearn.metrics import roc_auc_score
from sklearn.calibration import CalibratedClassifierCV
import lightgbm as lgb
import shap

# Load loan application + performance data
df = pd.read_csv("loan_data.csv")  # LendingClub or internal loan data
df['default'] = (df['loan_status'].isin(['Charged Off', 'Default'])).astype(int)

# Feature engineering
df['dti_ratio'] = df['dti'] / 100
df['installment_to_income'] = df['installment'] / (df['annual_inc'] / 12)
df['credit_history_years'] = (pd.to_datetime(df['issue_d']) -
                               pd.to_datetime(df['earliest_cr_line'])).dt.days / 365

features = ['loan_amnt', 'int_rate', 'dti_ratio', 'installment_to_income',
            'fico_range_low', 'delinq_2yrs', 'inq_last_6mths', 'open_acc',
            'pub_rec', 'revol_util', 'total_acc', 'credit_history_years']

X, y = df[features].fillna(df[features].median()), df['default']

# Time-series CV (critical: must not use future data for credit models)
tscv = TimeSeriesSplit(n_splits=5)
ginis = []
for train_idx, val_idx in tscv.split(X):
    model = lgb.LGBMClassifier(n_estimators=300, max_depth=5, learning_rate=0.05,
                                min_child_samples=50, reg_lambda=1.0)
    model.fit(X.iloc[train_idx], y.iloc[train_idx])
    preds = model.predict_proba(X.iloc[val_idx])[:, 1]
    ginis.append(2 * roc_auc_score(y.iloc[val_idx], preds) - 1)

print(f"Mean Gini: {np.mean(ginis):.3f} ± {np.std(ginis):.3f}")  # Target: >0.40

# SHAP for regulatory explainability (SR 11-7 compliance)
model.fit(X, y)
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X)

# Global: feature importance for model documentation
shap.summary_plot(shap_values, X)

# Individual: explain a specific loan decision (adverse action notice)
loan_idx = 42
print("\nLoan decision explanation:")
for feat, val in sorted(zip(features, shap_values[loan_idx]), key=lambda x: abs(x[1]), reverse=True)[:5]:
    direction = "↑ risk" if val > 0 else "↓ risk"
    print(f"  {feat}: {direction} (SHAP={val:.3f}, value={X.iloc[loan_idx][feat]:.2f})")

# Calibration: PD must be well-calibrated for IFRS 9
calibrated = CalibratedClassifierCV(model, cv='prefit', method='isotonic')
calibrated.fit(X, y)  # Ensures predicted PD = actual default rate at each score level
</syntaxhighlight>

; Banking risk AI tools
: '''Credit scoring''' → FICO Score (traditional), zest.ai, Scienaptic AI (ML credit)
: '''Market risk''' → MSCI RiskManager, Axioma Qontigo, Bloomberg PORT
: '''Stress testing''' → Moody's Analytics, S&P Global Market Intelligence
: '''Explainable ML''' → InterpretML (EBM), SHAP, IBM AI Fairness 360
: '''Systemic risk''' → SRISK (NYU Stern), BIS network analysis tools
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">