📈 economics

From Churn Prediction to Retention Pricing: A Cost-Sensitive Decision Framework for Managing Revenue at Risk in Subscription Services

This paper presents a cost-sensitive decision framework that transforms churn prediction into an actionable retention-pricing strategy by quantifying revenue at risk and demonstrating that broad, low-cost outreach is economically superior to narrow, high-precision targeting for maximizing net benefit in subscription services.

Original authors: DHARANI M

Published 2026-09-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: DHARANI M

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of subscription services, from streaming platforms to mobile phone plans, a company's survival depends on a simple, continuous cycle: bringing in new customers and keeping the ones it already has. The challenge of keeping customers is known as "retention," while the act of a customer leaving is called "churn." For a business, losing a customer is not just a single lost sale; it is the disappearance of a steady stream of income that must be replaced by finding a new person, which is almost always much more expensive. Because of this, companies have long tried to predict who is about to leave so they can intervene. Traditionally, this prediction has been treated as a pure math problem, where the goal is simply to build a computer model that guesses correctly as often as possible. However, being right in a general sense is not the same as being useful in a business sense. A model might be very accurate overall but still fail to guide a company on who to call, what offer to make, or how much that offer should cost. The real question is not just "who will leave?" but "who is it worth trying to save?"

This research takes a different approach, treating the prediction of who will leave not as an end in itself, but as the starting point for a financial decision. The study focuses on a dataset of over 7,000 customers from a telecommunications provider, examining their billing records, how long they have been with the company, and the type of contract they signed. The researchers first built a machine learning system to rank every customer by their likelihood of leaving. Crucially, they designed this system to avoid a common mistake where the computer uses future data during its training, ensuring the rankings are honest and reliable. Once they had a list of customers ranked by risk, they did not stop at the prediction. Instead, they asked a practical question: if the company offers a discount to keep a customer, at what point does the cost of that discount outweigh the value of the customer staying?

The analysis revealed a striking imbalance in where the money is actually at risk. While customers on month-to-month contracts make up only about half of the total customer base, they account for nearly 87 percent of the annual revenue that is currently in danger of being lost. These customers leave at a rate more than fifteen times higher than those who sign up for two-year contracts. Paradoxically, the customers who leave most often are the ones who, on average, generate the least amount of long-term value for the company, while the customers who stay the longest are the most valuable. This creates a complex situation for business leaders: the segment that is most likely to leave is also the one that is cheapest to lose, yet because so many of them are leaving at once, the total amount of money at stake is massive.

To solve this, the researchers tested a strategy of contacting customers with retention offers based on their risk score. They simulated various scenarios, assuming different costs for the offers and different rates of success in convincing people to stay. The results showed that the most profitable strategy is not to wait for a very high certainty that a customer will leave before acting. Instead, the data suggests that companies should cast a much wider net. The optimal strategy involves contacting customers even when their risk of leaving is relatively low, specifically around a 15 percent probability. This might seem counterintuitive, as it means reaching out to many people who would have stayed anyway. However, the math shows that the cost of sending a small discount to a few extra people is far lower than the cost of missing the chance to save a customer who actually leaves. In the simulations, this broad, low-threshold approach generated significantly more net benefit than a narrow, high-precision strategy that only targeted the most obvious cases.

The study concludes that the way companies manage retention needs to shift from a narrow focus on accuracy to a broader focus on economics. Because the cost of losing a customer is so high compared to the cost of a failed retention offer, it makes financial sense to be generous with outreach. The findings also point to a specific pricing strategy: companies should use contract length as a lever to move customers away from the risky, month-to-month group and into longer-term commitments, rather than relying solely on reactive discounts after a customer has already decided to leave. By connecting the prediction of who might leave directly to the cost of saving them, this work provides a clear, reproducible method for businesses to decide how to spend their retention budget, turning a statistical model into a practical guide for protecting revenue.

Technical Summary: From Churn Prediction to Retention Pricing

Problem Statement
Customer churn in subscription businesses represents a fundamental revenue management challenge: the loss of a recurring revenue stream necessitates replacement at a significantly higher acquisition cost. While churn prediction is a well-established field, existing research predominantly treats it as a standalone classification exercise, optimizing for metrics like accuracy or F1-score. This approach fails to bridge the gap between predictive modeling and actionable revenue management decisions. Specifically, a model's accuracy does not dictate the economic viability of a retention campaign, which depends on the trade-off between the cost of an intervention (e.g., a discount offer) and the expected revenue saved. The paper addresses this disconnect by reframing churn prediction as a cost-sensitive decision input rather than a mere classification task.

Methodology
The study utilizes the Telco Customer Churn dataset, comprising 7,043 customer records with demographic, service, contract, and billing attributes. The methodology is structured into four distinct components:

Leakage-Free Probability Modeling:
To ensure valid performance estimation, the author implemented a strict "leakage-free" pipeline. The dataset was split into training (80%) and test (20%) sets using stratified sampling before any resampling or scaling. Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training fold. Six classifiers (Logistic Regression, K-Nearest Neighbors, Random Forest, Gradient Boosting, Histogram-based Gradient Boosting, and Multi-Layer Perceptron) were tuned via GridSearchCV with stratified 3-fold cross-validation, optimizing for the F1-score. The resulting probability scores serve as the demand-side signal for the retention decision.
Revenue-at-Risk and CLV Quantification:
Using actual billing data, the author calculated the annual revenue at risk by contract segment (Month-to-month, One-year, Two-year). Additionally, a first-order Customer Lifetime Value (CLV) estimate was derived per segment by multiplying average monthly charges by average tenure. This metric, while not discounting future cash flows or modeling survival probability explicitly, serves to illustrate relative value differentials.
Cost-Sensitive Retention-Threshold Framework:
The core decision framework formulates the selection of customers for retention offers as a threshold optimization problem. For a given probability threshold $t$ , the model calculates:
- Revenue Saved: The sum of monthly charges for true churners identified above the threshold, multiplied by an assumed retention success rate and a 12-month horizon.
- Campaign Cost: The sum of costs for all contacted customers (true positives and false positives), calculated as a fraction of their monthly billing over a 3-month offer duration.
- Net Benefit: The difference between Revenue Saved and Campaign Cost.
Since historical campaign outcomes (offer cost and success rates) are unavailable in the dataset, the study treats these as assumed parameters. A sensitivity analysis sweeps across three plausible scenarios (e.g., 15% offer cost with 0.4 retention success) to identify the optimal threshold $t^*$ that maximizes Net Benefit.
Segment-Level Analysis:
The analysis explicitly breaks down metrics by contract type to identify asymmetries in churn exposure and customer value.

Key Results

Model Performance: Under the leakage-free protocol, the Neural Network (MLP) achieved the highest recall (0.746) and F1-score (0.622) on the test set. Given the revenue management context where the cost of a false negative (losing a customer) exceeds the cost of a false positive (wasting an offer), the MLP was selected as the probability source for the decision framework.
Revenue Concentration: Month-to-month subscribers account for 87% of the total annual revenue at risk ($1.45 million of $1.67 million), despite comprising only 55% of the customer base. This segment exhibits a churn rate of 42.7%, compared to 2.8% for two-year contract holders.
Value vs. Risk Asymmetry: An inverse relationship was observed between churn exposure and per-customer value. Month-to-month customers have the lowest estimated CLV ($1,198), while two-year customers have the highest ($3,448).
Optimal Threshold: Across all sensitivity scenarios, the Net Benefit is maximized at a low probability threshold of approximately 0.15. This implies that contacting roughly 900 customers (out of 1,409 in the test set) yields a higher economic return than narrow, high-precision targeting. At a threshold of 0.15, the model captures 356 of 374 true churners in the test set. As the threshold increases (e.g., to 0.5), the net benefit declines (e.g., by 15% in the base case) because the marginal gain in precision is outweighed by the lost revenue from missing true churners.

Significance and Contributions
The paper claims to provide a reproducible bridge between predictive analytics and revenue management decision-making. Its primary contributions are:

Segment-Level Revenue Quantification: Moving beyond aggregate churn rates to quantify revenue at risk by contract type using actual billing data.
Methodological Rigor: Establishing a leakage-free classification pipeline that produces well-calibrated probability rankings suitable for economic decision-making.
Decision Framework: Introducing a cost-sensitive threshold framework that explicitly trades offer costs against expected revenue, demonstrating that broad, low-cost outreach is often economically superior to narrow targeting under typical subscription cost structures.
Pricing Implications: Highlighting that retention spend should be prioritized by segment based on total revenue at risk rather than individual CLV, and suggesting that contract-length incentives (e.g., discounts for longer terms) can serve as a proactive demand-management lever to shift customers from high-risk/low-value segments to lower-risk/higher-value segments.

The author maintains a modest stance, noting that the study relies on assumed parameters for offer costs and success rates rather than measured campaign data. Consequently, the findings are presented as a sensitivity range to guide budgeting and pricing design, with the recommendation that future work incorporate real A/B-test data to refine these estimates.

Technical Summary: From Churn Prediction to Retention Pricing

More like this