Learn how to combine pre-churn scoring and uplift modeling for effective customer retention in fintech.
Introduction
In the fintech landscape, retaining customers is significantly more cost-effective than acquiring new ones. This principle is particularly important in digital banking, where customer acquisition expenses can be exorbitant.
Customer retention hinges on various elements such as the quality of service, the functionality of products, effective loyalty systems, and transparent communication with users. This article discusses a concrete example of leveraging pentagon-set-to-deploy-chatgpt-to-enhance-operations-for-3-million-personnel/">intelligence-is-reshaping-the-landscape-of-cybercrime/">machine learning (ML) to address the challenge of customer retention.
In addressing customer retention, we can break down our approach into three main steps: identifying users at risk of churning, determining the best retention strategies to re-engage them, and optimizing interventions to ensure maximum effectiveness.
This discussion illustrates a system built using two ML models: a pre-churn model that predicts which users are at risk of discontinuing their usage and a subsequent uplift model that assesses the potential impact of targeted retention offers. Let’s delve into how these models function and their significance in developing an effective customer retention system.
Every fintech company must first establish a clear definition of churn. In our case, we adopted a practical approach for a debit card product: churn is defined as a user not making any card payments for a duration of 30 days. This straightforward definition allows us to focus on transactional activity as a key indicator.
With this definition in mind, our objective is to predict which users will cease making transactions within the next month. We accomplish this by developing a binary classification model that estimates the likelihood of a user making at least one payment within the specified timeframe.
The creation of this pre-churn model depends on a curated set of user features. While proprietary information about our feature set cannot be disclosed, I will share fundamental principles for developing features in a fintech environment. Typical features often include transactional history, customer engagement levels, and demographic data.
To construct the model, different machine learning tools may be utilized. Gradient boosting algorithms such as XGBoost, LightGBM, and CatBoost are known for their efficient performance in such tasks, coupled with manageable training resources.
Once the model is in place, users are ranked based on their risk scores. A user is categorized into the pre-churn segment if their predicted probability of making a payment falls below a determined threshold. Factors influencing this threshold selection include the size of the segment, budgetary constraints for retention measures, and the acceptable risk level for the organization.
Validation of this model involves ensuring its effectiveness under real-world conditions. A robust strategy is to set aside more recent data for validation purposes. Initially, we train the model on historical data, then assess its predictive capabilities using a more recent dataset.
It's crucial to prioritize the quality of the predicted probabilities, not just the accuracy of classifications. Monitoring metrics such as ROC-AUC, which indicate how effectively the model distinguishes between users who are likely to churn and those who are not, is particularly important.
Calibration is another vital aspect we must consider. Over time, predicted probabilities may diverge from actual user behavior. To mitigate this drift, we can implement a calibration model that adjusts the primary pre-churn model’s predictions on new data, typically using logistic regression.
Once the pre-churn model is established, it’s common to initiate an experiment where a portion of the pre-churn segment receives a tailored retention offer, while others receive standard communications. If successful, this may indicate that targeted retention measures can indeed reinvigorate inactive users.
Initial findings from our experiments indicated that enhanced loyalty offers, such as increased cashback, could drive user re-engagement. However, we identified that the associated costs of retention could exceed the expenses involved in acquiring new customers, particularly due to the presence of organic users within the pre-churn segment.
Essentially, this meant that while some users were genuinely at risk, others were already inclined to remain active irrespective of the retention offer, thus creating inefficiencies.
Our key takeaway from this phase was clear: it is insufficient to merely identify users in the risk zone. We also need to ascertain which users' behaviors can be influenced by targeted interventions versus those who are likely to remain indifferent. This realization led us to incorporate an uplift model into our retention system.
The uplift model is designed to quantify how much a specific retention offer can modify user behavior. The objective is to pinpoint users whose likelihood of engagement increases due to receiving a personalized offer. This targeted approach aids in conserving budget resources, allowing businesses to concentrate their retention efforts on users who show genuine responsiveness.
In order to train the uplift model, we require experimental data wherein a segment of users is randomly exposed to the offer while another remains in a control group. This setup generates the necessary data to observe differential responses, forming the foundation of our uplift learning.
Uplift is calculated as the disparity in payment likelihood between users who receive the intervention versus those who do not. There are two prominent methodologies for deriving uplift: the T-learner and the S-learner.
The T-learner involves training two distinct models—one for users receiving the offer and another for those in the control group. The uplift is then determined by subtracting the probabilities predicted for users in each group.
Conversely, the S-learner uses a single model that incorporates user features along with an indicator for whether the user received the intervention. By predicting outcomes based on both scenarios for the same user, uplift can be derived accordingly. In practice, both methods yield comparable results.
After implementing the uplift model, we conducted an A/B experiment to assess its overall efficiency. The results demonstrated that integrating the uplift model into our retention framework led to a meaningful reduction in the costs associated with retaining users. A significant contribution to this financial efficiency stemmed from the halting of broad outreach to organic users who would have completed transactions without any assistance.
The final retention mechanism employed a sequential selection strategy for executing retention interventions. Initially, the pre-churn model estimated payment probabilities for users. Those falling below the predetermined threshold entered the pre-churn segment. Next, we applied the uplift model to further refine the subset, identifying users most likely to respond positively to the offer. This strategic filtering ensured that retention measures were directed only toward those most likely to benefit.
Looking ahead, regular recalibration for both the pre-churn and uplift models is integral to maintaining system effectiveness. To facilitate this, we designed an architecture where ongoing fresh experimental data is collected consistently. This approach not only enhances model accuracy but also ensures retention strategies remain relevant in rapidly changing market conditions.
Moreover, excluding certain users from retention interventions serves a useful purpose. This group establishes a reference point, providing a clear picture of the actual churn rate without any intervention. Coupled with randomized offer allocations, we gather clean data to inform recalibrations of the uplift model.
In summary, this comprehensive ML-based retention system demonstrates how fintech companies can effectively pinpoint users at risk of leaving, while simultaneously understanding which individuals would respond positively to targeted offers. By adeptly utilizing data-driven insights, businesses can optimize their marketing budgets, focusing on users who truly need intervention.
What factors contribute to customer churn in fintech? Major factors include high acquisition costs, poor service quality, lack of personalized experiences, and insufficient communication.
How can I measure the success of a retention strategy? You can measure success through key performance indicators like increased user activity, reduced churn rates, and improved customer satisfaction scores.
Is machine learning essential for customer retention in fintech? While not mandatory, machine learning provides valuable tools for predicting user behavior and targeting retention efforts effectively, leading to considerable savings and improved customer satisfaction.