CASE STUDY: OMAP

OMAP

How we replaced legacy actuarial pricing models with a hybrid neural network architecture — uncovering hidden risk segments and measurably reducing loss ratios across OMAP's auto insurance portfolio.

The Challenge

OMAP (Ontario Mutual Automobile Plan) relied on traditional Generalized Linear Models (GLMs) to price auto insurance policies — the industry standard for decades. But GLMs are inherently limited to capturing linear, predefined variable interactions. As OMAP's dataset grew to include telematics feeds, geospatial risk indicators, and complex policyholder behavioral data, the models failed to capture non-linear patterns that were clearly present in the loss experience. The result: systematic mispricing of both high-risk and low-risk segments, leading to adverse selection, inflated loss ratios, and an eroding competitive position in a market where pricing precision is everything.

COMBINED ACTUARIAL NEURAL NETWORK (CANN) — DUAL-STREAM ARCHITECTUREINPUT FEATURESAge / GenderTerritoryVehicle ClassDriving RecordTelematics ScoresGeospatial DensityTime-of-Day UsageCross-Var TensorsStandard / RegulatedHigh-DimensionalSTREAM 1: GLMGeneralized Linear ModelTraditional actuarial rating structure✓ Fully interpretable✓ Regulatory compliant→ Base PremiumSTREAM 2: DNNDeep Neural NetworkNon-linear pattern recognition✓ Non-linear interactions✓ Learns residual from GLM→ Adjustment Multiplier×OPTIMIZEDPREMIUMBase × Multiplier= Final PriceTrained on 3+ years historical claims dataHoldout validation · Production drift monitoring

The Method: Combined Actuarial Neural Networks (CANN)

We deployed a Combined Actuarial Neural Network (CANN) — a hybrid architecture that preserves the interpretability and regulatory compliance of traditional GLMs while layering in the pattern-recognition power of deep learning. The key insight: rather than replacing the GLM entirely (which would create regulatory and explainability challenges), the neural network learns the residual error — the systematic patterns the GLM misses. This gives actuaries full transparency on the base rate structure while capturing complex feature interactions that linear models cannot represent.

The Solution

The architecture runs two parallel inference streams. The GLM stream processes standard regulatory rating variables (age, territory, vehicle class, driving record) and produces a base premium. The neural network stream ingests high-dimensional features — telematics driving scores, geographic accident density heatmaps, time-of-day usage patterns, and cross-variable interaction tensors — and outputs a multiplicative adjustment factor. Both streams converge at inference time into a single optimized premium. We trained the network on 3+ years of historical claims data with careful holdout validation to prevent overfitting, and built a monitoring layer to track model drift in production.

PRICING ENGINE RESULTS — RISK SEGMENT DISCOVERYBEFORE: GLM ONLY■ Actual Risk■ GLM PriceRisk segments treated uniformlyLow-mileage UrbanActual: Low RiskStandard SuburbanActual: MediumHigh-telematicsActual: High RiskYoung UrbanActual: MediumAFTER: CANN■ Actual Risk■ CANN PriceHidden segments identified & repricedLow-mileage UrbanActual: Low RiskStandard SuburbanActual: MediumHigh-telematicsActual: High RiskYoung UrbanActual: Medium→

End State & Outcomes

The CANN model identified risk micro-segments that the GLM had completely missed — including a cluster of low-mileage urban drivers who were being significantly overcharged, and a subset of high-telematics-score policyholders whose actual risk was 2–3x higher than the GLM predicted. OMAP was able to reprice these segments accurately: offering more competitive rates to genuinely low-risk drivers (improving retention) while correctly pricing high-risk cohorts (reducing loss exposure). Overall, the new pricing engine delivered a measurable reduction in combined loss ratios and gave OMAP a data-driven edge that competitors using traditional GLMs alone cannot match.