Editing AI for Insurance Underwriting (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''Telematics driving risk score model:'''
<syntaxhighlight lang="python">
import pandas as pd
import numpy as np
import lightgbm as lgb
from sklearn.model_selection import GroupKFold
from sklearn.metrics import roc_auc_score, brier_score_loss
from sklearn.calibration import calibration_curve

def compute_telematics_features(trips_df: pd.DataFrame) -> pd.DataFrame:
    """Aggregate trip-level telematics data to policy-level features."""
    policy_features = trips_df.groupby('policy_id').agg(
        total_miles=('distance_miles', 'sum'),
        total_trips=('trip_id', 'count'),
        avg_speed=('avg_speed_mph', 'mean'),
        max_speed=('max_speed_mph', 'max'),
        pct_night_driving=('is_night', 'mean'),         # 11pm-5am: higher risk
        pct_highway=('pct_highway', 'mean'),
        harsh_braking_rate=('harsh_braking_events', lambda x: x.sum() / x.count()),
        rapid_accel_rate=('rapid_accel_events', lambda x: x.sum() / x.count()),
        hard_cornering_rate=('hard_corner_events', lambda x: x.sum() / x.count()),
        distraction_rate=('phone_use_events', lambda x: x.sum() / x.count()),
        weekend_driving_pct=('is_weekend', 'mean'),
        rush_hour_pct=('is_rush_hour', 'mean'),         # More exposure to other risky drivers
    ).reset_index()

    # Composite behavior score (higher = riskier)
    policy_features['behavior_score'] = (
        policy_features['harsh_braking_rate'] * 30 +
        policy_features['rapid_accel_rate'] * 20 +
        policy_features['hard_cornering_rate'] * 15 +
        policy_features['distraction_rate'] * 35 +      # Highest weight
        policy_features['pct_night_driving'] * 10
    )
    return policy_features

# Train claim frequency model
features = ['total_miles', 'avg_speed', 'max_speed', 'pct_night_driving',
            'harsh_braking_rate', 'rapid_accel_rate', 'distraction_rate',
            'behavior_score', 'driver_age', 'vehicle_age', 'zip_risk_score']

model = lgb.LGBMClassifier(
    n_estimators=300, max_depth=6, learning_rate=0.05,
    scale_pos_weight=10  # Claims are rare: ~5-10% annual frequency
)
# GroupKFold to prevent data leakage across policy periods
gkf = GroupKFold(n_splits=5)
aucs = []
for train_idx, val_idx in gkf.split(X, y, groups=df['policy_id']):
    model.fit(X.iloc[train_idx], y.iloc[train_idx])
    preds = model.predict_proba(X.iloc[val_idx])[:, 1]
    aucs.append(roc_auc_score(y.iloc[val_idx], preds))
print(f"Mean AUC: {np.mean(aucs):.3f}")  # Target: >0.70 for meaningful pricing lift

# UBI premium relativity
df['claim_prob'] = model.predict_proba(X)[:, 1]
df['base_premium'] = 1200  # Annual base premium
df['ubi_premium'] = df['base_premium'] * (df['claim_prob'] / df['claim_prob'].mean())
print(f"Premium range: ${df['ubi_premium'].quantile(0.1):.0f} - ${df['ubi_premium'].quantile(0.9):.0f}")
</syntaxhighlight>

; Insurance AI tools
: '''Telematics''' → Cambridge Mobile Telematics, TrueMotion (Verisk), Octo Telematics
: '''Property imagery''' → Cape Analytics, Nearmap, EagleView, Verisk Aerial Analytics
: '''Automated underwriting''' → Zesty.ai (homeowners), Lapetus Solutions (life/mortality)
: '''Claims AI''' → Tractable (auto damage), Snapsheet (claims processing), Guidewire
: '''CAT modeling''' → RMS (Moody's), AIR Worldwide (Verisk), CoreLogic
</div>

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