Chainladder Modeling
Compare pooled and feature-segmented chainladder estimates using the synthetic triangles generated from the claims data.
Claims Modeled
420
Chainladder applied with and without segmentation
Pooled chainladder
Ultimate: $28,180,995
IBNR: $857,325
Severity-segmented
Ultimate: $28,160,733
IBNR: $837,063
Ground truth (synthetic)
Ultimate: $28,128,598
IBNR: $804,928
Segmentation effect
38.7% closer to true ultimate
Severity segmentation sharpens link ratios and reduces IBNR error
Claim coverage
420
Synthetic triangles built from all claims, segmented by severity
Method comparison
| Method | Ultimate | IBNR | IBNR Error vs. True |
|---|---|---|---|
| Pooled chainladder | $28,180,995 | $857,325 | 6.5% |
| Severity-segmented chainladder | $28,160,733 | $837,063 | 4.0% |
| Synthetic ground truth | $28,128,598 | $804,928 | 0.0% |
Severity breakdown
Segmented chainladder results
| Segment | Claims | Latest Paid | Ultimate | IBNR | Share of segmented IBNR |
|---|---|---|---|---|---|
| Major | 92 | $16,058,199 | $16,746,356 | $688,157 | 82.2% |
| Moderate | 158 | $8,666,700 | $8,801,153 | $134,453 | 16.1% |
| Minor | 170 | $2,598,770 | $2,613,225 | $14,454 | 1.7% |
| Total | 420 | $27,323,670 | $28,160,733 | $837,063 | 100% |
Minor
Fast emergence, low noise, minimal tail.
Moderate
Steady emergence, balanced noise profile.
Major
Longer tail, materially heteroskedastic increments.