Customer, Distribution and Marketing Analytics

Introduction and context

Insurance is increasingly managed as a customer relationship rather than as a sequence of isolated policy transactions. A customer may discover an insurer through a comparison website, obtain advice from an agent or broker, purchase through a mobile application, submit a claim through a call centre, and renew through an automated digital journey. Each interaction creates data and each decision affects acquisition cost, persistency, claims experience, profitability, conduct risk and customer trust.

Customer, distribution and marketing analytics applies statistical, economic and behavioural methods to these decisions. Its purpose is not merely to predict who will buy. A mature analytics function estimates which customers are likely to respond, which interventions cause additional response, which channels create sustainable value, and which action should be taken next under commercial, regulatory and ethical constraints.

The central managerial problem is therefore an optimisation problem: allocate scarce attention, advice, incentives and capital to customers and distribution partners in ways that improve value while preserving fairness and treating customers appropriately.

Learning objectives

By the end of this chapter, the reader should be able to:

1. Distinguish acquisition, conversion, churn, renewal, cross-sell and customer lifetime value models.
2. Construct useful customer and distribution metrics and identify their assumptions.
3. Evaluate agents, brokers and digital channels using profitability, quality, persistency and conduct indicators.
4. Explain uplift modelling, experimentation and causal inference in insurance marketing.
5. Design next-best-action and lifetime-value programmes that are operationally and ethically sound.
6. Critically assess bias, leakage, drift, privacy and regulatory risks in customer analytics.

1. The analytical unit: customer, policy, household and journey

Before modelling, an insurer must define the unit of analysis. A policy-level model may predict renewal of a motor policy; a customer-level model may predict whether a person buys home insurance; a household-level view may reveal that several policies are held by related members. These units are not interchangeable.

A customer journey normally contains the following stages:

A customer data mart should preserve time order. It may contain quotes, exposures, contacts, prices, channel, intermediary, product, payment behaviour, complaints, claims, service interactions and consent status. The modelling population, observation window, prediction horizon and outcome definition must be documented before variables are selected.

A crucial distinction is between a prediction and a decision. A propensity score estimates the likelihood of an outcome under observed conditions. It does not necessarily show that an intervention caused the outcome. A customer predicted to renew may renew without a reminder; sending the reminder may add no value. This distinction motivates uplift modelling and causal inference.

2. Core customer models

2.1 Acquisition and conversion

Acquisition analytics estimates the probability that a prospect becomes a customer and the cost of generating that customer. Let p_i be the probability that prospect i converts within a specified horizon. A logistic model is often used when the outcome is binary:

Here, x_ij are predictor variables, β_j are estimated coefficients, and β₀ is the intercept. The model assumes that the log-odds are appropriately related to the predictors and that observations, or the dependence structure, are handled correctly. The output is a probability only after applying the inverse-logit transformation:

The model may support lead prioritisation, but ranking alone is insufficient. A lead with high conversion probability may require expensive human contact, while a medium-probability lead may convert cheaply online. A decision metric should therefore combine probability, expected contribution, contact cost and compliance constraints.

Acquisition cost is commonly measured as:

Attribution is difficult when customers encounter several channels. Last-click attribution is simple but can over-credit the final touchpoint and ignore brand, referral and prior interactions. Incrementality tests are generally more credible for deciding whether a channel caused additional business.

2.2 Churn, renewal and persistency

Churn is the loss of a customer or policy during a defined period. In insurance, cancellation, non-payment, replacement by another insurer and failure to renew may have different causes and should not automatically be combined.

A simple churn rate is:

The denominator must be defined carefully. If policies enter and leave the portfolio throughout the period, exposure-based or survival methods are preferable. A survival function, S(t), denotes the probability of remaining active beyond time t. The hazard at time t describes the instantaneous risk of exit conditional on survival until t. In discrete monthly data, a practical hazard estimate is:

This is not the same as a twelve-month churn probability. If monthly hazards are constant and customers are exposed for twelve months, the approximate annual survival is (1-h)¹², and annual churn is 1-(1-h)¹². This calculation assumes a stable hazard and no competing risks.

Renewal models can use price change, tenure, claims, service contacts, payment history, broker, product fit and prior renewal behaviour. However, variables observed only after a renewal decision must not be used to predict that decision. Such target leakage produces impressive validation results and poor deployment performance.

Retention interventions should be economically justified. The expected value of an intervention may be represented as:

where p_i^retained is the incremental probability of retaining customer i, M_i is the contribution margin associated with retention over the relevant horizon, and C_i is intervention cost. The word incremental is essential: a conventional churn score estimates risk, whereas the intervention decision requires expected change caused by the action.

2.3 Cross-sell and product propensity

Cross-sell analytics estimates the probability that an existing customer purchases a suitable additional product. It should incorporate need, eligibility, affordability, contact permission, product suitability and existing cover. A high propensity does not authorise indiscriminate selling. The recommendation must be appropriate and understandable.

A useful evaluation is the response rate among contacted customers compared with a valid control group. Cross-sell models can be biased if historical salespeople contacted customers selectively. The model may learn the salesperson's behaviour rather than customer need.

2.4 Customer lifetime value

Customer lifetime value (CLV) is the expected discounted future contribution attributable to a customer relationship. A general formulation is:

Here, T is the forecast horizon, P(active_i,t) is the probability that the relationship is active at time t, Revenue_i,t and VariableCost_i,t are expected monetary amounts, r is the periodic discount rate, and acquisition cost is measured consistently. Amounts must use the same currency and time unit; if t is monthly, r must be a monthly rate.

For insurance, contribution should be defined carefully. Premium is not equivalent to profit because claims, reinsurance, commissions, servicing, expenses, taxes and capital costs matter. Depending on purpose, CLV may use expected underwriting contribution, contribution after claims, or a broader economic-profit measure. Claims costs should not be treated as customer value in a way that penalises customers for using entitled coverage or creates unfair treatment.

A simplified constant-margin model illustrates the logic. If annual expected contribution is £120, annual retention probability is 0.80, annual discount rate is 0.05, and the horizon is three years, then:

The three discounted contributions are approximately £91.43, £69.39 and £52.88. Before acquisition cost, CLV is approximately £213.70. This is illustrative and assumes constant contribution, independent retention across years, no additional products, no inflation and no terminal value. A real model would usually use cohort, survival, product and expense assumptions.

CLV is best used for resource allocation and portfolio strategy, not as a justification for inferior service to low-value customers. Fair treatment, contractual obligations and service standards remain applicable to all customers.

3. Distribution analytics

Insurance distribution includes tied agents, independent agents, brokers, bancassurance partners, comparison platforms, affinity partners, call centres and direct digital channels. Evaluation should distinguish volume from quality.

Important indicators include:

Comparing intermediaries requires risk and mix adjustment. A broker specialising in complex commercial risks should not be judged directly against a digital channel selling simple personal lines. Regression, matching, hierarchical models or standardised cohorts can reduce confounding, but they cannot remove all unobserved differences.

Channel profitability may be represented as:

The quantities must be measured on a consistent basis and over an appropriate development period. Early-period profitability can be misleading because claims develop over time. For long-tail business, incurred or ultimate expected claims should be used according to the purpose of analysis.

Agent and broker scorecards should combine leading indicators, such as quote quality and training completion, with lagging indicators, such as complaints, persistency and loss performance. Incentives based only on sales volume can encourage unsuitable sales, misrepresentation or excessive discounting. Balanced scorecards and clawback arrangements can align commercial rewards with customer and portfolio outcomes.

Digital channels require funnel analysis. If 10,000 visitors generate 1,000 quotes and 120 purchases, visitor-to-quote conversion is 10%, quote-to-bind conversion is 12%, and visitor-to-bind conversion is 1.2%. Each ratio identifies a different intervention point. Funnel metrics must account for duplicate visitors, bots, returning users and channel switching.

4. Uplift modelling and next-best action

4.1 Propensity versus uplift

A propensity model estimates the likelihood of purchase, renewal or response. An uplift model estimates the difference between outcomes with and without an intervention for comparable customers. Let Y(1) be the potential outcome if treated and Y(0) the potential outcome if not treated. Individual treatment effect is:

Both potential outcomes cannot normally be observed for the same person. The average treatment effect (ATE) is:

The conditional average treatment effect for customers with characteristics X is:

Uplift segmentation commonly identifies four groups: persuadables, sure things, lost causes and sleeping dogs. A persuadable customer responds because of treatment; a sure thing responds anyway; a lost cause does not respond; a sleeping dog may respond negatively or be harmed by contact. This framework helps avoid spending on sure things and disturbing sleeping dogs.

4.2 Experimentation

The strongest practical design is a randomised controlled experiment. Eligible customers are randomly assigned to treatment and control, with the treatment receiving an intervention and the control receiving business-as-usual or no intervention. Randomisation makes treatment assignment independent of potential outcomes in expectation.

For a binary outcome, an intention-to-treat estimate is:

If 900 of 10,000 treated customers respond and 750 of 10,000 control customers respond, the estimated incremental response is 9.0% − 7.5% = 1.5 percentage points, or 150 additional responses per 10,000 eligible customers. Statistical uncertainty should be reported with confidence intervals or appropriate tests, not only with a point estimate.

Experiments require a clear unit of randomisation, stable treatment delivery, adequate sample size, pre-specified outcomes, and monitoring for interference and attrition. If household members share a treatment, randomising individuals may contaminate the control group; household-level randomisation may be more appropriate.

4.3 Observational causal inference

Randomisation is not always feasible. Observational methods include regression adjustment, propensity scores, inverse probability weighting, matching, interrupted time series and difference-in-differences. These methods depend on assumptions such as consistency, appropriate measurement, positivity and conditional exchangeability.

A propensity score is the probability of treatment given observed covariates:

Inverse probability weighting can create a pseudo-population in which observed covariates are balanced. It does not correct for unmeasured confounding, poor overlap, treatment misclassification or outcomes measured after treatment inappropriately. Causal claims must therefore be proportionate to the design and diagnostics.

Difference-in-differences compares changes over time in a treated group with changes in a comparison group:

Its key identifying condition is the parallel-trends assumption: absent treatment, treated and comparison groups would have followed similar trends. This assumption should be assessed using pre-intervention data and subject-matter reasoning.

4.4 Next-best action

Next-best-action systems select an intervention—such as a renewal reminder, coverage explanation, service call, payment support or cross-sell suggestion—subject to customer need, eligibility, consent, capacity and expected value. A conceptual decision rule is:

The maximisation is over permissible actions. IncrementalExpectedValue should reflect treatment effect, contribution, operational capacity and risk, not merely predicted response. Business rules must override the model where legal, safety or conduct requirements apply.

A production system should include contact-frequency limits, suppression after complaint or vulnerability indicators, explanation pathways, monitoring, and human review for consequential decisions. Reinforcement learning may eventually optimise sequences of actions, but it requires careful reward design, exploration controls and off-policy evaluation. It should not be deployed merely because it is technically sophisticated.

5. Measuring marketing effectiveness

Marketing measurement should connect exposure to incremental outcomes and economic value. Relevant metrics include reach, response, conversion, incremental conversion, cost per incremental customer, return on marketing investment and retention impact.

Return on marketing investment can be expressed as:

The numerator must be incremental contribution, not total revenue attributed to marketing. A channel can have high attributed sales but low incrementality if it mainly captures customers who would have purchased anyway.

Attribution models include first-touch, last-touch, linear, position-based, algorithmic and media-mix approaches. Each has limitations. Multi-touch attribution often relies on user-level tracking and can be distorted by missing identifiers, walled gardens and selection bias. Media-mix modelling works at aggregated levels and can estimate broader effects but needs sufficient variation and careful controls.

A robust measurement architecture combines:

• randomised holdouts for important campaigns;
• consistent campaign and cost metadata;
• pre-registered outcome definitions;
• customer-level and aggregate analyses;
• financial reconciliation to policy and accounting systems;
• post-campaign checks for cancellations, complaints and claims quality.

6. Data, model development and validation

A reliable workflow begins with a business decision, not an algorithm. The analyst should specify the action, eligible population, prediction time, outcome horizon, cost of errors and constraints. Data preparation then includes identity resolution, consent filtering, missingness analysis, temporal joins and exclusion of post-outcome information.

Data should be split by time where deployment is temporal. Random cross-validation can overstate performance when customer behaviour, prices or campaigns change. Evaluation should include discrimination, calibration, stability and business value. For binary propensity models, discrimination may be measured by area under the ROC curve, while calibration compares predicted probabilities with observed frequencies. Neither metric alone establishes profitability or fairness.

Class imbalance is common in churn and fraud-related outcomes. Accuracy can be misleading when events are rare. Precision, recall, lift, calibration and expected cost should be selected according to the decision. Thresholds should reflect intervention capacity and asymmetric consequences.

Model monitoring should examine population drift, feature drift, calibration drift, outcome drift, treatment effects, approval rates, complaint rates and subgroup performance. Champion–challenger testing, version control, reproducible feature definitions and documented model cards support governance.

7. Ethical, legal and governance considerations

Customer analytics processes personal data and can influence prices, access, communication and service. Governance should address lawful basis, purpose limitation, data minimisation, retention, security, consent and rights of access or explanation, according to applicable law and regulatory standards.

Potential sources of unfairness include historical discrimination, proxy variables, unequal data quality, selective marketing, differential digital access and feedback loops. Removing protected attributes does not guarantee fairness because postcode, language, occupation or device type may act as proxies. Fairness assessment should be risk-based and consider legally protected groups, vulnerable customers and the practical effect of decisions.

Insurance-specific concerns include:

• using claims history or behavioural data in ways that penalise legitimate claims;
• targeting financially vulnerable customers with unsuitable products;
• opaque automated decisions;
• excessive contact or dark-pattern design;
• conflicts in intermediary incentives;
• using sensitive data without necessity;
• treating correlation as evidence of customer intent.

Explainability should be useful rather than merely technical. A customer-facing explanation should identify relevant factors, available choices and review routes without revealing security-sensitive information. Human oversight must be meaningful: reviewers need authority, information and time to challenge the system.

8. Applied case: renewal and retention in a motor portfolio

Background

This illustrative case concerns Meridian Mutual, a fictional motor insurer with 400,000 annual renewal opportunities. The insurer has rising acquisition costs and a renewal rate of 78%. It sends the same reminder sequence to nearly all customers 30 days before renewal.

Problem

Management wants to reduce avoidable lapses without encouraging inappropriate discounts or disadvantaging customers who need service support. A conventional churn model identifies customers likely to lapse, but the marketing team discovers that many high-risk customers renew even without intervention. The model therefore has limited value for deciding whom to contact.

Analysis

The analytics team defines the unit as a policy renewal opportunity and freezes predictors 35 days before renewal. The outcome is renewal within 30 days after expiry. It excludes policies already cancelled, records with invalid contact permission and cases subject to manual underwriting. Historical campaign data are reviewed for leakage and inconsistent treatment.

A six-week randomised pilot assigns eligible opportunities to standard communication, enhanced explanation of price and cover, or service-oriented contact for customers with unresolved service events. Randomisation occurs within product and channel strata. Primary outcomes are renewal and contribution after communication cost; secondary outcomes include complaints, payment failure and subsequent cancellation.

The pilot produces the following illustrative results:

The incremental renewal rate is 2.0 percentage points, equal to 200 additional renewals per 10,000 eligible policies. Suppose incremental contribution per retained policy is £145 and intervention cost is £3 per treated policy. Incremental contribution is 200 × £145 = £29,000; intervention cost is £30,000. The net incremental value is therefore −£1,000 for the full 10,000-policy intervention. The intervention is not economically justified at this average effect.

The team then estimates heterogeneous treatment effects using pre-specified features and validates them in a holdout period. It finds that a targeted segment has a 5 percentage-point incremental renewal effect and an average contribution of £160. If 2,000 customers are targeted, expected additional renewals are 100 and incremental contribution is £16,000. At £3 per contact, cost is £6,000, giving an estimated net value of £10,000, before uncertainty and implementation costs.

Approach and solution

Meridian deploys a constrained targeting rule: contact only customers with positive estimated incremental value, valid consent, no recent complaint suppression flag and a suitable communication channel. Customers with unresolved service problems are routed to trained service staff rather than offered a discount. A randomised holdout remains in place to measure continuing incrementality.

The scorecard includes renewal, net contribution, complaints, payment outcomes, cancellation within 60 days and subgroup calibration. The model is reviewed quarterly and after major pricing or product changes.

Outcome and learning points

The case demonstrates that a high churn probability is not equivalent to high retention opportunity. The economically relevant quantity is incremental value after intervention cost. It also shows why customer analytics must integrate marketing, service, actuarial, distribution, compliance and finance perspectives. A campaign that increases renewal but produces complaints, unsuitable cover or rapid cancellation may not create sustainable value.

9. Contemporary developments and best practice

Privacy-enhancing technologies, clean rooms, federated learning and synthetic data may reduce the need to centralise identifiable information, although they do not eliminate governance obligations. Generative AI can assist campaign drafting, summarisation and analyst productivity, but customer-facing use requires controls against hallucination, discriminatory language, unauthorised personalisation and disclosure of confidential information.

Real-time decisioning is expanding through event streams from websites, applications and service systems. It can improve relevance but increases the risk of excessive surveillance and inconsistent treatment. Event-driven systems should have clear purpose boundaries and auditable decision logs.

Best practice is to:

1. Begin with a defined customer or distribution decision.
2. Separate prediction, causal effect and optimisation questions.
3. Establish time-valid data and prevent leakage.
4. Measure incrementality through randomisation whenever feasible.
5. Use contribution and customer outcomes, not volume alone.
6. Evaluate channels and intermediaries on comparable, risk-adjusted cohorts.
7. Apply consent, suitability, vulnerability and contact-frequency rules before optimisation.
8. Monitor calibration, drift, subgroup effects, complaints and downstream outcomes.
9. Retain holdouts and document assumptions, uncertainty and limitations.
10. Give accountable managers authority to override models and investigate anomalies.

Chapter Summary

Customer, distribution and marketing analytics supports decisions across the insurance relationship: acquisition, conversion, onboarding, renewal, churn prevention, cross-sell and service. Propensity models estimate likely outcomes, while uplift models and causal designs estimate the additional effect of an intervention. This distinction is fundamental to profitable and responsible targeting.

Customer lifetime value combines expected future contribution, retention, discounting and acquisition cost, but requires consistent financial definitions and should never replace fair treatment. Distribution analytics must evaluate agents, brokers and digital channels using productivity, persistency, economics, customer outcomes and risk quality, with adjustment for portfolio mix and claims development.

Randomised experiments provide strong evidence of incrementality; observational causal methods can help when experiments are infeasible but rely on demanding assumptions. Next-best-action systems should maximise constrained incremental value rather than raw response probability. Effective implementation depends on temporal data design, leakage prevention, calibration, monitoring, privacy, fairness, explainability and accountable human governance.

Key Terms

Customer lifetime value: Expected discounted future contribution from a customer relationship after relevant costs.

Churn: Exit of a customer or policy during a defined period.

Conversion: Movement from a prospect, quote or lead to a purchased policy.

Persistency: Continuation of a policy or customer relationship over time.

Propensity model: Model estimating the probability of an observed outcome.

Uplift model: Model estimating the incremental effect of an intervention.

Potential outcome: Outcome that would occur under a specified treatment or non-treatment condition.

Average treatment effect: Average difference between treated and untreated potential outcomes.

CATE: Conditional average treatment effect for customers with specified characteristics.

Incrementality: Outcome caused by an intervention beyond what would otherwise occur.

Randomised controlled experiment: Study assigning units randomly to treatment and control conditions.

Propensity score: Probability of receiving treatment conditional on observed covariates.

Next-best action: Selected permissible intervention expected to create the greatest incremental value.

Customer acquisition cost: Attributable cost of acquiring a new customer.

Marketing attribution: Allocation of observed outcomes across marketing touchpoints.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Target leakage: Use of information unavailable at the time a prediction would be made.

Distribution analytics: Analysis of intermediary and channel performance, quality and economics.

Difference-in-differences: Causal estimator comparing changes in treated and comparison groups.

Conduct risk: Risk that business practices produce unfair, unsuitable or harmful customer outcomes.

Review Questions

Short Answer Questions

1. Why is a churn propensity score insufficient for selecting retention interventions?
2. Define customer lifetime value and identify four components that should be considered in its calculation.
3. What is the difference between a policy-level and a household-level analytical unit?
4. State two limitations of last-touch marketing attribution.
5. Why should distribution channels be compared using risk- and mix-adjusted measures?

Long Answer Questions

1. Explain how an insurer could design an acquisition-to-renewal analytics framework, including outcomes, data windows and governance controls.
2. Compare propensity modelling, uplift modelling and randomised experimentation in terms of purpose, assumptions and managerial use.
3. Discuss the construction and limitations of customer lifetime value in a multi-product insurance portfolio.
4. Explain how agents, brokers and digital channels can be evaluated through a balanced distribution scorecard.
5. Analyse the ethical and regulatory risks of automated next-best-action systems and propose appropriate safeguards.

Analytical / Application Questions

1. An insurer contacts 20,000 customers. The treatment renewal rate is 82% and the control renewal rate is 79%. The contribution per incremental renewal is £130 and contact cost is £2.50 per customer. Calculate incremental renewals, incremental contribution, intervention cost and net value. State whether the intervention is justified before uncertainty is considered.
2. Design a randomised experiment to test whether a personalised renewal explanation improves retention. Specify the unit of randomisation, treatment, control, primary outcome, secondary outcomes and at least three threats to validity.
3. A broker has high sales volume but low persistency and high complaint rates, while a digital channel has lower volume but higher persistency and lower cost. Develop a decision framework that avoids selecting a channel solely on sales volume and identifies the additional data required.

Multiple Choice Questions with Answers

Question 1

Which measure most directly estimates the additional response caused by a marketing intervention?

A. Propensity score

B. Uplift estimate

C. Last-touch attribution

D. Churn rate

Correct Answer: B. Uplift estimate

Question 2

In a retention experiment, the treatment renewal rate is 80% and the control renewal rate is 77%. What is the estimated incremental renewal rate?

A. 3 percentage points

B. 23 percentage points

C. 77 percentage points

D. 157 percentage points

Correct Answer: A. 3 percentage points

Question 3

Which information is an example of target leakage in a renewal model?

A. Policy tenure measured before renewal

B. Prior-year renewal status

C. A complaint recorded after the renewal decision

D. Premium quoted before renewal

Correct Answer: C. A complaint recorded after the renewal decision

Question 4

The principal purpose of a control group in a marketing experiment is to estimate:

A. Total customer revenue

B. The outcome absent the intervention

C. The insurer’s market share

D. The number of available leads

Correct Answer: B. The outcome absent the intervention

Question 5

Which group in the classic uplift segmentation responds because of the intervention?

A. Sure things

B. Lost causes

C. Sleeping dogs

D. Persuadables

Correct Answer: D. Persuadables

Question 6

If 1,000 treated customers have a 9% response rate and 1,000 control customers have a 7% response rate, how many incremental responses are estimated?

A. 2

B. 20

C. 70

D. 90

Correct Answer: B. 20

Question 7

Which factor is essential when calculating customer lifetime value?

A. Only the initial premium

B. Expected future contribution and discounting

C. Only the number of marketing impressions

D. The agent’s sales ranking alone

Correct Answer: B. Expected future contribution and discounting

Question 8

What does calibration assess in a probability model?

A. Whether predicted probabilities correspond to observed frequencies

B. Whether every variable is normally distributed

C. Whether the model has the highest possible accuracy

D. Whether treatment was randomly assigned

Correct Answer: A. Whether predicted probabilities correspond to observed frequencies

Question 9

Which is a major limitation of observational causal inference?

A. It cannot use customer data

B. It always requires a binary outcome

C. Unmeasured confounding may remain

D. It cannot estimate differences over time

Correct Answer: C. Unmeasured confounding may remain

Question 10

A balanced distribution scorecard should include which combination?

A. Sales volume only

B. Commission only

C. Productivity, persistency, economics and customer outcomes

D. Advertising impressions only

Correct Answer: C. Productivity, persistency, economics and customer outcomes