Policyholder Churn Prediction
“Which policyholders will not renew?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which policyholders will not renew?
P&C insurers face 10-15% annual non-renewal rates, with acquisition costs of $400-$600 per new policyholder (J.D. Power). For an insurer with 5M policyholders, a 12% churn rate means 600K lost customers and $240-360M in replacement acquisition costs annually. Worse, profitable low-risk policyholders are the most likely to leave because competitors aggressively poach them with lower rates. The signals of impending non-renewal are scattered: rate-shopping behavior (quote requests from competitors), claim dissatisfaction, premium increase reactions, and life changes (moving, marriage, new vehicle) that trigger a review of coverage.
Quick answer
AI predicts which insurance policyholders will not renew by connecting rate-change history, service interactions, competitive market activity, and claims patterns into a relational graph. Unlike simple logistic regression on policyholder attributes, graph-based models detect that a low-loss-ratio customer who received a 12% rate increase and lives in a zip code with active competitor conquest campaigns is 74% likely to leave. Early detection gives retention teams 60-90 days to intervene with targeted offers.
Approaches compared
4 ways to solve this problem
1. Rule-Based Retention Triggers
Flag policyholders based on hardcoded rules: rate increase above X%, more than Y service calls, or approaching renewal date without engagement. Simple to implement and transparent.
Best for
Quick wins when you have obvious churn triggers like large rate increases or recent claim denials.
Watch out for
Rules miss the interaction effects. A 12% rate increase might be fine for a price-insensitive customer but fatal for a low-loss-ratio policyholder in a competitive market. One-size-fits-all rules waste retention budget.
2. Logistic Regression / Survival Models
Statistical models trained on policyholder features like tenure, premium, claims history, and rate-change history. Survival models add time-to-event analysis for more nuanced predictions.
Best for
Building interpretable churn scores when you need to explain the model to business stakeholders or regulators.
Watch out for
Cannot incorporate competitive market dynamics, service-interaction sentiment, or life-event signals without extensive manual feature engineering. Misses the relational context entirely.
3. Gradient-Boosted Trees (XGBoost)
Tree-based models that capture non-linear interactions between policyholder attributes. Can handle mixed data types and missing values well.
Best for
Moderate accuracy improvement over logistic regression with reasonable engineering effort.
Watch out for
Still a flat-table approach. Cannot see that a competitor launched a conquest campaign in the policyholder's zip code, or that the policyholder's household has been shopping for quotes. These cross-table signals require manual joins and aggregations.
4. Relational Deep Learning (Kumo's Approach)
Connects policyholders to rate changes, service interactions, competitive market data, claims history, and life-event signals in a single graph. Learns churn patterns from the full relational context automatically.
Best for
Detecting churn risk 60-90 days before renewal by combining rate-sensitivity signals, competitive pressure, and service-interaction patterns that span multiple data sources.
Watch out for
Requires competitive market data and service-interaction logs to be connected to policyholder records. If these data sources are siloed in different systems, integration work comes first.
Key metric: Multi-line policyholders retain at 90% vs. 70% for single-line (J.D. Power). Targeted retention of profitable at-risk customers saves $60-120M annually for a 5M-policyholder insurer.
Why relational data changes the answer
Flat churn models see each policyholder as an isolated row: premium, tenure, claims count, rate-change percentage. They can predict that policyholders with large rate increases churn more often. But they cannot see that Jennifer Adams received a 12% rate increase, called customer service twice with negative sentiment, lives in a zip code where two competitors just launched aggressive conquest campaigns, and has a loss ratio of 0.28 that makes her extremely attractive to those competitors. These signals come from four different tables, and their interaction effect is what drives the prediction. A 12% increase for a policyholder with no competitive alternatives is manageable. The same increase for Jennifer is a near-certain loss.
Relational learning captures these multi-table interaction patterns without manual feature engineering. The model walks from policyholder to rate changes, to service interactions, to competitive market activity in their zip code, to their claims history and risk profile. It learns that the combination of high rate increase + competitive pressure + low loss ratio + negative service sentiment predicts churn at 74% confidence, while any single factor alone would score below 40%. This precision matters because retention budgets are finite. Spending $200 on a rate adjustment for a profitable customer about to leave is a great investment. Spending the same $200 on a high-loss-ratio customer who was going to renew anyway is a waste.
Predicting churn from a flat policyholder table is like predicting whether a restaurant regular will stop coming by looking only at their visit frequency. You miss that the restaurant just raised prices 12%, a new competitor opened across the street with a grand-opening discount, and the regular complained to the manager about cold food last Tuesday. Every signal matters, and they compound. Relational churn models see the full picture: the price increase, the competitor, and the complaint, all connected to the same customer.
How KumoRFM solves this
Relational intelligence built for insurance data
Kumo connects policyholders to their policy details, claims history, billing patterns, service interactions, rate changes, and competitive market data. The model identifies that Policyholder PH-6601 received a 12% rate increase, called customer service twice with billing questions, and lives in a zip code where a competitor just launched an aggressive acquisition campaign. These signals predict non-renewal 60-90 days before the renewal date, giving retention teams time to offer proactive rate adjustments or coverage enhancements to keep profitable customers.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
POLICYHOLDERS
| policyholder_id | name | policy_type | premium | tenure_years | loss_ratio |
|---|---|---|---|---|---|
| PH-6601 | Jennifer Adams | Home + Auto | $3,200 | 6.4 | 0.28 |
| PH-6602 | Mark Stevens | Auto Only | $1,800 | 2.1 | 0.65 |
| PH-6603 | Diana Lee | Home + Auto + Umbrella | $5,400 | 11.2 | 0.15 |
RATE_CHANGES
| policyholder_id | effective_date | old_premium | new_premium | pct_change |
|---|---|---|---|---|
| PH-6601 | 2025-07-01 | $2,860 | $3,200 | +11.9% |
| PH-6602 | 2025-08-01 | $1,720 | $1,800 | +4.7% |
| PH-6603 | 2025-06-01 | $5,200 | $5,400 | +3.8% |
SERVICE_INTERACTIONS
| policyholder_id | channel | type | sentiment | timestamp |
|---|---|---|---|---|
| PH-6601 | Phone | Billing Question | Negative | 2025-09-05 |
| PH-6601 | Phone | Coverage Question | Neutral | 2025-09-12 |
| PH-6603 | App | Document Request | Positive | 2025-09-10 |
COMPETITIVE_MARKET
| zip_code | competitor | campaign_type | avg_savings_offered | start_date |
|---|---|---|---|---|
| 90210 | Geico | Conquest | $400-$600 | 2025-08-15 |
| 90210 | Progressive | Switch & Save | $300-$500 | 2025-09-01 |
| 10001 | None Active | N/A | N/A | N/A |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(POLICYHOLDERS.STATUS = 'non_renewed', 0, 90, days) FOR EACH POLICYHOLDERS.POLICYHOLDER_ID WHERE POLICYHOLDERS.STATUS = 'active'
Prediction output
Every entity gets a score, updated continuously
| POLICYHOLDER_ID | PREMIUM | CHURN_PROB | LOSS_RATIO | RETENTION_ACTION |
|---|---|---|---|---|
| PH-6601 | $3,200 | 0.74 | 0.28 | Proactive Rate Review |
| PH-6602 | $1,800 | 0.31 | 0.65 | Standard Renewal |
| PH-6603 | $5,400 | 0.08 | 0.15 | Loyalty Reward |
Understand why
Every prediction includes feature attributions — no black boxes
Policyholder PH-6601 (Jennifer Adams)
Predicted: 74% probability of non-renewal
Top contributing features
Rate increase magnitude
+11.9%
28% attribution
Competitor conquest campaigns in zip
2 active
24% attribution
Negative service interactions
2 calls, 1 negative
21% attribution
Low loss ratio (attractive to competitors)
0.28
16% attribution
No bundling discount applied
Missing auto bundle
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about policyholder churn prediction
How does AI predict policyholder churn in insurance?
AI predicts policyholder churn by analyzing connected signals across rate changes, service interactions, competitive market activity, claims history, and life events. Graph-based models detect that specific combinations of these factors (like a large rate increase plus active competitor campaigns in the policyholder's area) predict non-renewal 60-90 days before the renewal date.
What is the cost of policyholder churn for insurance companies?
Acquiring a new policyholder costs $400-$600 (J.D. Power). For an insurer with 5M policyholders and 12% annual churn, that is 600K lost customers and $240-360M in replacement costs per year. Worse, the most profitable low-risk policyholders are the ones most likely to leave because competitors target them aggressively.
How early can AI detect insurance customer churn?
Graph-based models can flag at-risk policyholders 60-90 days before renewal. The signals appear when rate increases are applied, when service interactions turn negative, when competitors launch local campaigns, or when life events (home purchase, new vehicle) trigger coverage shopping. Earlier detection means more time for effective retention interventions.
What retention strategies work best for at-risk insurance policyholders?
The most effective strategies are personalized: proactive rate reviews for price-sensitive customers, bundling offers for single-line policyholders, loyalty rewards for long-tenure customers, and coverage enhancements for customers who recently had claims. AI-driven targeting doubles conversion rates compared to untargeted campaigns because it matches the right intervention to the right customer.
Bottom line: Retain 25-35% of at-risk profitable policyholders with targeted rate adjustments, saving $60-120M in annual acquisition costs for a 5M-policyholder insurer.
Related use cases
Explore more insurance use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
