The right way to Enhance Buyer Retention in FinTech

Introduction

, there’s a sample that any rising group shortly learns: buying a brand new consumer, or bringing again a consumer who has already churned, is way more durable and dearer than retaining somebody who remains to be utilizing the product. In digital banking, that is particularly seen as a result of buyer acquisition prices are excessive.

Person retention in fintech will depend on many elements: service high quality, product performance, loyalty mechanics, communication, and different components of the consumer expertise. On this article, I’ll give attention to one sensible instance: how ML-based personalization can assist work with at-risk customers and apply loyalty mechanics extra exactly.

The duty had three steps. First, we would have liked to discover ways to establish customers who have been more likely to churn. Then we needed to outline the retention mechanic, which means what precisely may deliver the consumer’s curiosity again to the product. After that, we would have liked to optimize the intervention itself: to know which customers within the threat phase would really be affected by the retention provide.

On this article, I’ll share a sensible information to constructing an analytical system for this sort of process. It’s based mostly on two ML fashions: a pre-churn mannequin, which identifies customers in danger, and an uplift mannequin, which helps decide which ones are actually delicate to a retention provide. I’ll clarify the sensible worth of every stage: why it was wanted from a product perspective, what drawback it solved, and what conclusions we have been in a position to attract from the analysis.

Preparation: Goal and Options

Let’s begin with the churn definition. I labored on the expansion of a debit card product, the place consumer conduct is finest described by means of transactional exercise. So we settled on a sensible definition: churn means no card funds for 30 days. This definition is commonly utilized in monetary merchandise, and the logic behind it’s pretty clear.

If an individual has not made a single transaction in a month, the cardboard has successfully dropped out of their pockets. Formally, the account should be open, however from a product perspective, the consumer is not lively. Accordingly, the mannequin goal is a binary flag: whether or not the consumer made no less than one cost inside the subsequent 30 days.

Subsequent, let’s transfer to characteristic assortment. For this process, I collected a number of dozen options describing the consumer’s conduct within the service. I can not disclose the precise manufacturing characteristic set, however I’ll share the principle ideas for constructing a characteristic illustration of a consumer in a fintech product based mostly on transactional exercise and consumer profile information.

• Person profile. Fundamental consumer traits: age, area, system sort, related merchandise, tariff plan, subscriptions or loyalty program standing, account standing equivalent to new consumer, lively consumer, and so forth.
• Exercise aggregates over time home windows. Behavioral metrics calculated over 7, 30, and 90 days: variety of transactions, turnover, common transaction quantity, variety of distinctive spending classes, variety of lively days, and so forth. Completely different home windows assist seize adjustments in consumer conduct.
• Time intervals. How a lot time has handed for the reason that final or first cost, what the common hole between transactions was, and the way the rhythm of exercise adjustments over time. In observe, the sign {that a} consumer “has not appeared for some time” typically seems to be one of many strongest predictors of churn.
• Calendar options. Day of the week and month. With out these options, the mannequin might confuse pure drops in exercise with precise churn. Monetary companies are extremely delicate to seasonal patterns, so it’s higher to account for them upfront.
• Derived options. Ratios between time home windows, shares of transactions of a sure sort, normalized and relative metrics. These options assist evaluate purchasers with completely different baseline ranges of exercise and keep away from penalizing customers who merely spend much less.

Pre-churn Mannequin

Let’s begin with the pre-churn mannequin. Our purpose at this stage is to establish the phase of customers for whom the present loyalty mechanics are not working. In different phrases, customers who’re dropping curiosity within the service beneath the present stage of loyalty incentives.

At this stage, the duty is to establish customers with an elevated threat of leaving. For this, we’d like a mannequin that predicts the likelihood of a future cost. If the expected likelihood is low, the consumer falls into the pre-churn phase. In that case, the service ought to take note of the consumer prematurely, whereas churn can nonetheless be prevented.

Coaching the baseline mannequin

The mannequin is constructed on the collected options and predicts the likelihood of a cost over a 30-day horizon. Primarily, this can be a binary classification process with likelihood estimation. Person traits are handed as enter, and the output is the likelihood that the consumer will make no less than one cost inside the subsequent 30 days.

$p_base(x) = P(Y = 1 | X = x)$

the place Y ∈ {0, 1} signifies whether or not the consumer will make a cost inside the subsequent 30 days, and X denotes the matrix of consumer options.

Many ML instruments can be utilized for this sort of process. In observe, gradient boosting fashions typically ship sturdy efficiency with an inexpensive quantity of analytical effort required for mannequin coaching: XGBoost [1], LightGBM [2], and CatBoost [3].

Customers can then be ranked by threat stage. The decrease the expected likelihood of cost, the upper the chance of churn:

$risk_score(x) = 1 – p_base(x)$

A consumer enters the pre-churn phase if their predicted likelihood of cost is beneath a particular threshold t:

$S_prechurn = {x : p_base(x) < t}$

The brink t is chosen based mostly on the phase measurement, the retention finances, and the suitable stage of threat. In sensible phrases, this implies selecting a threshold that retains loyalty prices inside the promo marketing campaign finances.

Validation

First, when validating this sort of mannequin, it is very important verify whether or not it can work in the identical mode by which it is going to be utilized in manufacturing. To do that, it’s helpful to carry out a number of newer months for an extra mannequin verify. The mannequin is skilled on deeper historic information, after which we consider how effectively it predicts churn in a newer time window.

Second, what issues right here is just not a lot the standard of the ultimate classification, however the high quality of the expected likelihood itself. We have to monitor how effectively the mannequin ranks customers by their threat of churn, for instance utilizing ROC-AUC and different rating metrics. It’s also vital to look individually at likelihood calibration: whether or not any bias seems within the predicted chances. For this, we are able to use a calibration curve and a Hosmer–Lemeshow curve [5].

In our case, the mannequin’s rating high quality barely degraded even on contemporary information. Nonetheless, calibration did deteriorate: over time, the expected chances turned shifted relative to the precise occasion charge beneath the affect of unobserved elements. It is a sample I often see in comparable fashions, so I wish to emphasize it individually.

Calibration

There’s a pretty efficient method to tackle the issue of deteriorating mannequin calibration. It entails constructing a separate calibration mannequin on prime of the principle pre-churn mannequin. The thought is straightforward: the calibration mannequin takes the prediction of the principle mannequin as its solely characteristic and transforms it right into a extra correct likelihood on contemporary information. Logistic regression is used for this step.

$p_calibrated(x) = g(p_base(x))$

the place p_base(x) is the prediction of the baseline pre-churn mannequin, and g is the calibration mannequin.

The principle benefit of this method is its simplicity. The calibration mannequin may be up to date often, is fast to validate, and doesn’t require a full redesign of the whole ML system. Because of this, we get a course of that helps preserve the mannequin steady over time with out fixed handbook upkeep.

Supply for the Pre-Churn Phase

After constructing the pre-churn mannequin, we launched a randomized experiment: a part of the pre-churn phase obtained an elevated cashback provide, whereas the remainder remained within the common stream.

The experiment confirmed that the route was proper: the elevated cashback provide helped enhance consumer retention, and we noticed statistically vital beneficial properties in retention. Nonetheless, we weren’t glad with the price of retaining one consumer. It turned out to be even increased than the price of buying new customers. This was pushed by a big share of natural customers inside the phase. Together with genuinely at-risk customers, the pre-churn mannequin additionally chosen customers who have been much less attentive to retention mechanics. In different phrases, some customers obtained the cashback provide although they have been already going to make a cost, whereas others didn’t reply to it in any respect.

This led to the principle conclusion of this stage: figuring out {that a} consumer is within the threat zone is just not sufficient. We additionally want to know whose conduct can really be modified by the communication, and whose conduct will stay the identical. For that, we’d like the following stage: an uplift mannequin, which we’ll transfer on to now.

Uplift Mannequin

The purpose of the uplift mannequin is to estimate how a lot a selected provide can change consumer conduct. It identifies the purchasers whose likelihood of creating a cost will increase probably the most particularly due to the provide. That is the place the price saving comes from: the finances is just not spent on the whole threat phase, however solely on the customers whom the provide can really assist retain.

Such a mannequin can’t be skilled solely on consumer options and the actual fact of cost. It wants experimental information, the place a part of the viewers obtained the provide at random and one other half remained within the management group. Solely then can the mannequin observe the distinction in response between the 2 situations and be taught to separate customers who’re delicate to the intervention from those that are detached. The A/B experiment we ran earlier supplied precisely this information, so we may transfer straight to mannequin constructing.

Mannequin coaching

In essence, uplift is the distinction between two chances of the goal motion: with and with out the intervention.

$τ(x) = P(Y = 1 | X = x, T = 1) – P(Y = 1 | X = x, T = 0)$

the place τ(x) is the anticipated uplift for consumer x, Y ∈ {0, 1} signifies whether or not the consumer made a cost inside 30 days, X denotes the consumer options, and T ∈ {0, 1} signifies whether or not the intervention was utilized. In our case, T = 1 signifies that the consumer obtained the cashback provide, whereas T = 0 signifies that the consumer didn’t obtain the provide.

There are a number of methods to estimate this distinction. I’ll describe two approaches [4]: the T-learner and the S-learner.

In a T-learner, two impartial binary classification fashions are skilled. The primary predicts the likelihood of cost for customers who obtained the provide. The second predicts the likelihood of cost for customers within the management group.

$p₀(x) = P(Y = 1 | X = x, T = 0)$

$p₁(x) = P(Y = 1 | X = x, T = 1)$

After that, uplift is calculated because the distinction between the 2 predictions for a similar consumer:

$τ(x) = p₁(x) – p₀(x)$

In different phrases, the mannequin estimates how a lot the provide will increase the likelihood of cost for a selected consumer.

In an S-learner, a single mannequin is used. It takes the consumer options and an extra characteristic indicating whether or not the consumer obtained the intervention or not.

$a(x, t) = P(Y = 1 | X = x, T = t)$

To estimate uplift, we make predictions for a similar consumer beneath two situations: first as if the consumer obtained the provide, after which as if the consumer didn’t.

$τ(x) = a(x, 1) – a(x, 0)$

In observe, each approaches produced comparable outcomes, so the selection between them was not essential. The subsequent step is intuitive: we take the customers with the best predicted uplift, which means these for whom the provide will increase the likelihood of cost probably the most.

Validation

After constructing the uplift mannequin, we evaluated its effectiveness utilizing an A/B experiment. The experiment design was as follows.

1. We chosen the pre-churn phase. The scale of the phase was decided by the accessible marketing campaign finances.
2. We cut up the phase into two teams: one with out the uplift mannequin and one with the uplift mannequin.
3. Within the group with out the uplift mannequin, the provide was distributed in the usual approach: some customers obtained cashback, whereas others remained within the management group.
4. Within the group with the uplift mannequin, the mannequin first chosen customers with the best predicted uplift. Then, inside this chosen phase, some customers obtained the provide, whereas others remained within the management group.
5. Lastly, we in contrast the outcomes. This allowed us to judge two methods: an ordinary provide distribution throughout the pre-churn phase and a focused distribution after filtering customers by means of the uplift mannequin.

Because of the experiment, we noticed that including the uplift mannequin to the retention system can considerably cut back the price of retaining a consumer. The principle impact comes from lowering the share of natural customers: presents cease being distributed at scale to customers who both would have made a cost with out the provide anyway, or wouldn’t have responded to it in any case.

Buyer Retention System

Let me describe how the ultimate retention system labored. In essence, it was a sequential choice course of for retention intervention.

On the first stage, the pre-churn mannequin p_calibrated(x) estimated the likelihood of a cost inside the subsequent 30 days. If this likelihood was beneath a particular threshold, the consumer entered the pre-churn phase:

$S_prechurn = {x : p_calibrated(x) < t}$

On the second stage, the uplift mannequin τ(x) was utilized inside the pre-churn phase. It estimated how a lot the cashback provide would change the likelihood of cost for a selected consumer and chosen those that have been more than likely to reply to the provide:

$Goal = {x ∈ S_prechurn : τ(x) > u}$

the place u is the minimal anticipated uplift at which it is sensible to indicate the provide to the consumer.

For ongoing operation of the system, we took into consideration that each fashions, the pre-churn mannequin and the uplift mannequin, would should be recalibrated often. Meaning the system wanted a relentless stream of contemporary experimental information. So we designed the provide distribution logic in a approach that might generate this information repeatedly.

Some customers have been excluded from the retention system solely, which means no retention mechanic was utilized to them. This group confirmed the actual churn stage with out intervention and served as a reference level for calibrating the pre-churn mannequin. Throughout the pre-churn phase itself, there was one other cut up: some customers didn’t obtain the provide, whereas others obtained it randomly, independently of the uplift mannequin. This random provide allocation supplied clear information on response to the intervention, which was then used to recalibrate the uplift mannequin.

Conclusion

In the long run, we constructed a whole machine learning-based retention system. It depends on two fashions. The primary one, the pre-churn mannequin, identifies customers with a low likelihood of cost, which means those that are already within the threat zone. The second, the uplift mannequin, seems to be inside this phase and finds prospects who’re really seemingly to reply to a personalised provide, equivalent to elevated cashback.

This mixture solves two issues directly. On the one hand, we are able to see prematurely who’s dropping curiosity within the service and react earlier than the consumer really leaves. Alternatively, the advertising finances is spent extra exactly: presents usually are not despatched to each consumer within the threat group, however solely to these whose conduct is more likely to be influenced by the intervention.

Because of this, the system helps not simply retain prospects, however do it in an economically significant approach. The corporate spends much less on customers who would have stayed anyway and concentrates the finances the place the retention intervention produces an actual incremental impact.

Sources

[1] Tianqi Chen and Carlos Guestrin. XGBoost: A Scalable Tree Boosting System. Proc. twenty second ACM SIGKDD, 2016. https://xgboost.readthedocs.io/

[2] Microsoft Company. LightGBM Documentation, 2026. https://lightgbm.readthedocs.io/

[3] Yandex. CatBoost Documentation, 2026. https://catboost.ai/docs/en/

[4] Sören R. Künzel, Jasjeet S. Sekhon, Peter J. Bickel, and Bin Yu. Metalearners for Estimating Heterogeneous Remedy Results Utilizing Machine Studying. Proceedings of the Nationwide Academy of Sciences, 116(10), 2019. https://www.pnas.org/doi/10.1073/pnas.1804597116

[5] Van Calster et al. Tutorial on Calibration Measurements and Calibration Fashions for Scientific Prediction Fashions. JAMIA, 27(4), 621–633, 2020. https://tutorial.oup.com/jamia/article/27/4/621/5762806