Underwriting Support

Problem

Assist human underwriters in assessing risk and determining appropriate terms (accept/decline, premium, coverage conditions) for insurance applications — particularly in commercial and specialty lines where risks are complex, heterogeneous, and require expert judgement. ML models augment underwriter decision-making rather than replacing it: surfacing risk signals, generating recommended premiums, flagging inconsistencies, and automating routine renewals.

Users / Stakeholders

Role	Decision
Underwriter	Accept/decline risk; set premium; apply endorsements
Underwriting manager	Portfolio quality monitoring; underwriter performance
Pricing actuary	Model performance vs technical price
Distribution (broker)	Quote turnaround speed; competitiveness
Compliance / Conduct	Fair pricing; prohibited rating factors

Domain Context

Judgement-heavy lines: Commercial property, liability, D&O, marine — each risk is unique. Models provide signals; underwriter makes final decision. Model must earn underwriter trust.
Technical price vs market price: The model outputs a technically appropriate premium based on expected loss. Underwriters adjust up or down based on market conditions, relationship, and judgement.
Protected characteristics: UK: EHRC guidance prohibits rating on age for some products. EU: Gender Directive prohibits gender-based pricing. Models must be audited for proxy discrimination.
Data richness varies: Personal lines (motor, home) have millions of data points. Commercial specialty (D&O, trade credit) may have thousands of claims over 20 years. Model complexity must match data availability.
Competitive intelligence: Underwriters need to know if the technical price is competitive in the market. Pricing model output is one input; market positioning is another.
Actuarial standards: ASOP 56 (US), APS X3 (UK) — actuarial professional standards for insurance modelling. Pricing models require actuarial validation.

Inputs and Outputs

Personal lines (motor):

Driver: age, gender (where permitted), occupation, driving history, NCB
Vehicle: make, model, engine_size, year, value, usage_type
Location: postcode density, theft_rate, flood_zone
Policy: coverage_type, voluntary_excess, named_drivers
Telematics (optional): mileage, time_of_day, braking_behaviour

Commercial lines:

Insured: SIC code, revenue, employees, years_in_business, public/private
Risk: premises type, construction, occupation, geographic_spread
Claims history: prior 5-year claims by peril and amount
Financial health: credit score, CCJs, director changes
Exposure: sum_insured, turnover_insured, contractual_liability

Output:

technical_premium:    Model-derived expected loss cost + expense + profit margin
risk_grade:           A / B / C / D / DECLINE (risk quality tier)
premium_range:        Acceptable range for underwriter negotiation
risk_flags:           Specific concerns (high claims freq, flood zone, concentration)
comparable_risks:     Similar risks in portfolio for reference
recommended_terms:    Suggested endorsements, sublimits, exclusions

Decision or Workflow Role

Application/renewal received (broker portal / direct)
  ↓
Automated data enrichment: postcode lookups, credit check, claims history
  ↓
Risk model: technical price + risk grade + flags
  ↓
Straight-through rules: 
  Grade A + small commercial → auto-quote (no UW intervention)
  Grade B/C → UW review with model output as reference
  Grade D / flagged → senior UW + manual referral
  ↓
Underwriter reviews, adjusts, quotes
  ↓
Broker/customer accepts → bound risk
  ↓
Claims development vs technical price → model performance feedback

Modeling / System Options

Component	Approach	Notes
Frequency model	GLM Poisson / Negative Binomial	Count of claims per policy year
Severity model	GLM Gamma / Tweedie	Average claim cost given claim occurs
Pure premium	Frequency × Severity or Tweedie combined	Main pricing output
Risk grade	Logistic regression / XGBoost classifier	Ordinal risk tier
Anomaly / flag	Isolation Forest or rule layer	Unusual risk characteristics
Telematics UBI	LSTM or gradient boosting on trip data	Usage-based insurance

Recommended: GLM Gamma + Poisson (frequency-severity) for actuarial pricing. LightGBM challenger for higher accuracy. Both require actuarial validation.

Deployment Constraints

Regulatory: Rate filings (US) or actuarial sign-off (UK) required before deployment. Cannot change rating factors without regulatory process in some jurisdictions.
Explainability: Underwriters and brokers must understand premium components. Waterfall charts showing factor contributions are standard in insurance pricing tools.
Auditability: Every quote must log the model version, input data, and output. Regulatory audit can request individual rating decisions years later.
Refresh cadence: Personal lines: annual rate review. Commercial: may be triggered by claim events or market changes.

Risks and Failure Modes

Risk	Description	Mitigation
Proxy discrimination	Postcode correlated with ethnicity → indirect discrimination	Fairness audit; geographic smoothing
Adverse selection	Model over-prices low-risk customers → they leave; under-prices high-risk → they stay	Portfolio monitoring; Lorenz curve analysis
Model overfit	Small data (specialty lines) → overfit to historical anomalies	Conservative regularisation; actuarial validation
Underwriter override concentration	UWs always override model for certain risks → model never improves	Track override patterns; recalibrate based on outcomes

Success Metrics

Metric	Target	Notes
Loss ratio improvement	> 3pp vs no-model baseline	Primary actuarial KPI
Gini coefficient	> 0.45	Discriminatory power
Combined ratio	< 95%	Profitability: loss + expense ratios
Technical price accuracy	± 5% of actual development	Calibration metric
UW productivity	+20% quotes per UW per day	Operational efficiency
Adverse action compliance	100%	No discriminatory rating factors

References

Frees, E.W. & Valdez, E. (1998). Understanding Relationships Using Copulas. NAAJ.
Tweedie GLMs: de Jong, P. & Heller, G.Z. (2008). Generalized Linear Models for Insurance Data. Cambridge.

Notes

Explorer

underwriting_support