← back to blog

Recommender Systems in Insurance

Personalizing coverage recommendations while respecting regulatory constraints

Chaïmae SritiAugust 2025

1. Introduction

Recommender systems in insurance serve a fundamentally different purpose than in e-commerce or entertainment. Instead of maximizing engagement or purchases, they must balance personalization with risk management, regulatory compliance, and customer protection.

The core challenge: recommend coverages that match customer needs without creating adverse selection, violating anti-discrimination laws, or undermining the insurer's loss ratio.

Key Use Cases:

  • Cross-sell and upsell recommendations (e.g., suggesting umbrella policy to homeowners)
  • Coverage limit optimization (e.g., recommending appropriate liability limits)
  • Bundling strategies (e.g., auto + home discounts)
  • Retention interventions (e.g., suggesting policy adjustments before renewal)

2. Collaborative Filtering

Collaborative filtering identifies patterns in customer behavior: "Customers similar to you also purchased..."

User-Based Collaborative Filtering:

Find customers with similar profiles and recommend what they purchased.

# Example: User similarity based on demographics
Customer A: Age 35, Married, 2 Cars, Homeowner
Customer B: Age 37, Married, 2 Cars, Homeowner ← Similar
Customer C: Age 24, Single, 1 Car, Renter ← Different

If Customer B has Umbrella policy → Recommend to Customer A

Item-Based Collaborative Filtering:

Find products frequently purchased together.

# Example: Product co-occurrence matrix
                Auto    Home    Umbrella    Life
Auto            1       0.65    0.42        0.28
Home            0.65    1       0.73        0.31
Umbrella        0.42    0.73    1           0.19
Life            0.28    0.31    0.19        1

→ If customer has Auto + Home, recommend Umbrella (0.73 correlation)

⚠️ Challenges in Insurance:

  • Sparse data: Customers don't buy insurance frequently (unlike Netflix watching patterns)
  • Cold start problem: New customers have no purchase history
  • Regulatory risk: Similar customers might be defined by protected classes (age, gender, location)
  • Adverse selection: Recommending high-risk products to high-risk customers worsens loss ratios

3. Content-Based Filtering

Content-based methods recommend products based on customer attributes and product features, rather than relying on behavior of other customers.

Feature Engineering for Recommendations:

Customer Features:
- Risk profile: Credit score, claims history, driving record
- Life stage: Age, marital status, number of dependents
- Assets: Home value, vehicle count, business ownership
- Coverage gaps: Uninsured/underinsured exposures

Product Features:
- Coverage type: Property, Liability, Life, Health
- Premium range: Budget, Standard, Premium
- Complexity: Simple (Term Life) vs Complex (Whole Life)
- Required prerequisites: Must have home to buy umbrella

Example Rule-Based Recommendation:

IF (has_home AND home_value > $500k AND has_auto) THEN
    recommend(Umbrella_Policy,
              min_limit = max(home_value, 1M),
              priority = HIGH)

IF (age > 30 AND has_dependents AND no_life_insurance) THEN
    recommend(Term_Life,
              coverage = 10 * annual_income,
              priority = CRITICAL)

IF (liability_limit < 100k AND net_worth > 250k) THEN
    recommend(Liability_Increase,
              suggested_limit = min(net_worth, 500k),
              priority = MEDIUM)

✓ Advantages:

  • Transparent and explainable to regulators
  • No cold start problem—works for new customers
  • Can encode actuarial expertise and business rules
  • Easier to audit for fairness and compliance

4. Hybrid Approaches

Hybrid systems combine collaborative filtering, content-based methods, and contextual signals to produce more robust recommendations.

Weighted Ensemble:

final_score = (
    0.4 * collaborative_score +    # What similar customers bought
    0.3 * content_score +           # Match customer-product features
    0.2 * business_rules_score +    # Actuarial constraints
    0.1 * contextual_score          # Time, location, external events
)

# Filter recommendations:
- Remove products customer already has
- Enforce prerequisites (e.g., need home for umbrella)
- Check underwriting eligibility
- Respect opt-out preferences
- Cap cross-sell attempts per period

Matrix Factorization with Constraints:

Use collaborative filtering but inject domain constraints as regularization.

# Factorize customer-product matrix into latent features
User embeddings: [risk_aversion, wealth, life_stage, ...]
Product embeddings: [complexity, premium, coverage_breadth, ...]

Prediction: score = user_embedding · product_embedding

# Add constraints during training:
- Penalize recommendations that violate business rules
- Add fairness constraints (demographic parity, equalized odds)
- Incorporate profitability signals (expected LTV - acquisition cost)

Contextual Bandits:

Treat recommendations as a reinforcement learning problem: learn which products to recommend in which contexts to maximize long-term customer value.

Context: Customer profile + time + channel
Action: Recommend product X
Reward: +1 if purchased, +5 if retained, -2 if churned

Thompson Sampling or UCB to balance:
- Exploitation: Recommend known high-converting products
- Exploration: Test new recommendations to discover better strategies

5. Regulatory Constraints

Insurance recommendations must navigate strict regulations around fairness, transparency, and consumer protection.

Protected Classes:

  • Cannot use race, religion, national origin, gender (in most states)
  • Age restrictions vary by product and jurisdiction
  • Credit-based features face increasing scrutiny
  • ZIP code can be proxy for protected classes → must audit for disparate impact

Explainability Requirements:

Regulators and customers may ask: "Why was this product recommended to me?"

  • Black-box models (deep neural nets) are harder to defend
  • Provide reason codes: "Recommended because you own a home valued over $500k"
  • Allow customers to challenge or opt-out of automated recommendations

Anti-Steering Regulations:

In some jurisdictions, insurers cannot systematically recommend cheaper or worse coverage to certain demographic groups.

  • Monitor recommendation distribution across protected classes
  • Ensure high-value products are recommended equitably
  • Test for disparate impact using A/B tests and statistical parity metrics

Fairness Metrics for Recommendations:

# Demographic Parity:
P(recommend_premium_product | group=A) ≈ P(recommend_premium_product | group=B)

# Equalized Odds:
P(accept_recommendation | qualified, group=A) ≈ P(accept_recommendation | qualified, group=B)

# Calibration:
For customers scored at 70% likelihood to need umbrella policy,
~70% should actually need it (across all groups)

6. Practical Applications

Use Case 1: Cross-Sell at Renewal

Goal: Recommend additional products when customer renews existing policy.

Pipeline:
1. Identify renewal cohort 60 days before expiration
2. Score each customer for cross-sell propensity (logistic regression)
3. Generate top-3 product recommendations per customer
4. Filter by underwriting rules and profitability thresholds
5. Serve recommendations via email, mobile app, or agent dashboard
6. Track conversion and adjust weights monthly

Metrics:
- Cross-sell conversion rate: 8% → 12% after recommendation system
- Average products per customer: 1.4 → 1.7
- Customer retention: +3% (bundled customers churn less)

Use Case 2: Coverage Gap Analysis

Goal: Identify customers who are underinsured and recommend appropriate increases.

Example:
Customer has $100k liability on auto policy
Net worth estimated at $800k (home value + retirement accounts)
→ High risk of being sued beyond coverage limits

Recommendation:
- Increase liability to $300k (+$15/month)
- OR add umbrella policy with $1M limit (+$20/month)

Delivery:
- Show calculator: "If sued for $500k, you'd pay $400k out of pocket"
- Emphasize protection of assets
- Provide easy one-click upgrade option

Use Case 3: Life Event Triggers

Goal: Detect life events and proactively recommend relevant coverage.

Signals:
- New vehicle added → Recommend comprehensive/collision
- Address change to more expensive home → Recommend higher dwelling coverage
- New driver added (teenage child) → Recommend higher liability limits
- Marriage detected → Recommend life insurance, umbrella policy
- Business ownership → Recommend commercial policy

Timing:
- Trigger recommendation within 30 days of detected event
- Use gentle nudge messaging (not aggressive sales)
- Respect communication preferences and frequency caps

Implementation Best Practices:

  • Start simple: Rule-based recommendations first, add ML incrementally
  • A/B test everything: Measure impact on conversion, retention, profitability, fairness
  • Monitor for drift: Customer behavior and product mix change over time
  • Audit for fairness: Quarterly reviews of recommendation distribution across demographics
  • Respect customer preferences: Allow opt-out, control frequency, provide explanations
  • Integrate with underwriting: Don't recommend products customer won't qualify for
  • Measure long-term value: Not just immediate conversion, but retention and profitability

Evaluation Metrics:

Business Metrics:
- Conversion rate: % of recommendations accepted
- Revenue per recommendation: Average premium increase
- Customer lifetime value: Long-term retention + cross-sell
- Loss ratio impact: Are recommended products profitable?

Model Metrics:
- Precision@K: Of top-K recommendations, how many convert?
- Recall@K: Of all products customer needs, how many in top-K?
- NDCG: Normalized Discounted Cumulative Gain (ranking quality)
- Coverage: % of customers receiving relevant recommendations

Fairness Metrics:
- Demographic parity across protected classes
- Equalized opportunity (true positive rate parity)
- Calibration (predicted propensity matches actual conversion)

Recommender systems in insurance must balance competing objectives: personalization vs. fairness, short-term conversion vs. long-term profitability, automation vs. transparency. Success requires not just strong ML models, but deep integration with underwriting, regulatory compliance, and customer experience design.