Claim Severity Prediction

Problem

Estimate the expected cost of an insurance claim at FNOL (First Notice of Loss) or during the claims triage process. Severity models inform: reserving (how much capital to set aside), triage (route to fast-track vs specialist handling), settlement negotiation, and fraud detection. The claim severity distribution is highly skewed — most claims are small, but a small proportion of complex claims account for the majority of total cost.

Typically modelled as a two-part model:

Severity | claim is open: Predicted ultimate cost given a claim has been filed
Pure premium = Frequency × Severity (combined with frequency model for pricing)

Users / Stakeholders

Role	Decision
Claims adjuster	Triage priority; settlement authority level
Reserve actuary	Case reserve adequacy; IBNR estimation
Claims manager	Resource allocation; SLA management
Pricing actuary	Pure premium for product pricing
SIU (Special Investigations Unit)	Severity outliers as fraud signals

Domain Context

Actuarial tradition: Severity modelling uses GLMs with log-link (Gamma distribution for continuous severity; Tweedie for combined frequency-severity). Regulatory capital requires actuarial certification of reserves.
IBNR: Incurred But Not Reported claims are estimated statistically from development triangles. ML supplements but rarely replaces chain-ladder methods for reserving.
Long-tailed lines: Workers’ comp, liability, and casualty claims may take 5–10 years to fully develop. Early severity estimates have high uncertainty.
Covariates available at FNOL: Only limited information is available at First Notice of Loss (cause code, reported injury type, claimant demographics, policy). More features become available as investigation proceeds.
Regulatory: Insurance is regulated at state/country level. Rate filings may require actuarial approval. Model documentation standards (ASOP 56 — Modeling in the US). GDPR applies to personal data in EU.
Data: Closed claims with known ultimate are training labels. Open claims are right-censored. Severity is zero-inflated (small claims settled at zero after investigation).

Inputs and Outputs

FNOL features:

Claim: cause_code, injury_type_code, body_part_code, accident_description_nlp
Policy: coverage_type, limit, deductible, policy_age, prior_claims_count
Claimant: age, occupation_class, injury_severity_index
Location: state, urban/rural, jurisdiction_factor
Context: day_of_week_filed, time_to_report (days), represented_by_attorney_flag
Prior development: initial_reserve_set, n_medical_visits, litigation_flag

Output:

predicted_severity:  Expected ultimate incurred cost
severity_percentile: Position in severity distribution (e.g., P75 = complex claim)
triage_tier:         FAST_TRACK / STANDARD / COMPLEX / LARGE_LOSS
reserve_recommendation: Point estimate + uncertainty range for reserving
fraud_flag_score:    Severity outlier relative to expected (feeds SIU)

Decision or Workflow Role

FNOL received → initial triage model scores claim
  ↓
FAST_TRACK (< P50 severity, no red flags) → automated settlement flow
STANDARD → assigned adjuster with case reserve guidance
COMPLEX (> P90 severity) → specialist team with higher authority
LARGE_LOSS → executive escalation + dedicated team
  ↓
Claim develops → feature updates → reserve updates
  ↓
Claim closes → actual = training label → model refresh
  ↓
Severity model feeds pricing actuaries via pure premium calculation

Modeling / System Options

Approach	Strength	Weakness	When to use
GLM Gamma with log-link	Actuarially validated; interpretable; regulator-familiar; handles skewed distribution	Misses interactions; manual feature engineering	Primary model for regulated reserving; pricing filings
Tweedie regression	Handles combined zero-inflated frequency × severity in one model	Less interpretable than two-part model	Pure premium for pricing
LightGBM	Captures non-linear interactions; higher predictive accuracy	Requires careful calibration; model risk overhead	Triage decisions (non-regulatory); fraud scoring
Quantile regression	Provides prediction intervals (e.g., P10/P50/P90)	Does not model full distribution	Reserving uncertainty quantification
Log-normal OLS	Simple; analytically tractable	Smearing correction required for back-transform	Quick baseline
Neural network	Highest raw accuracy on large datasets	Black box; actuarial acceptance barriers	Large direct carriers with 1M+ claims; challenger

Recommended: Gamma GLM for actuarial/regulatory use cases. LightGBM for triage and operational routing. Quantile regression for reserving uncertainty.

Deployment Constraints

Latency: Triage decision at FNOL should complete in <5 seconds. Batch reserve updates: overnight run.
Interpretability: Adjusters need to understand why a claim was scored as complex. SHAP reason codes are useful for high-severity flags.
Actuarial certification: In many jurisdictions, reserve models require sign-off by a qualified actuary. ML models need documentation and validation that meets ASOP standards.
Update cadence: Pricing models retrained annually (or at rate filing). Triage models can be retrained quarterly.

Risks and Failure Modes

Risk	Description	Mitigation
Right-censoring	Open claims not settled yet; unknown true ultimate	Develop only on settled claims; truncation correction
Attorney representation bias	Attorney involvement inflates severity; proxy for litigation jurisdiction	Explicit feature; jurisdiction-specific models
Reserve setting incentive	Adjusters under-reserve to hit targets; corrupts training labels	Independent actuarial label review
Distribution shift	Inflation, legal environment changes, pandemic effects	Periodic recalibration; economic index adjustment
Small sample in specialty lines	Niche products have few claims	Pooling with similar lines; credibility weighting

Success Metrics

Metric	Target	Notes
Reserve adequacy ratio	95–105% of ultimate	Actuarial KPI; ratio of predicted to actual
MAE (log-scale)	< 0.3 log-£	Model accuracy on held-out settled claims
Triage accuracy	> 85% correct tier	Fraction of complex claims correctly routed
SIU referral precision	> 30% confirmed fraud	Of claims flagged, fraction with confirmed fraud
Adjuster satisfaction	> 4.0/5 tool rating	Model utility for frontline users

References

Frees, E.W. et al. (2014). Predictive Modeling Applications in Actuarial Science. Cambridge University Press.
Tweedie, M.C.K. (1984). An index which distinguishes between some important exponential families.

Notes

Explorer

claim_severity_prediction