A Simple Model for Predicting Customer Churn

In today's digital business environment, consumers have become savvy smartphone shoppers, moving swiftly among apps, mini-programs, social media, shopping platforms, and company websites as they acquire information, seek views from friends and relatives, follow recommendation posts of KOLs, search for better deals, and eventually shop.As such, every business needs to change and become as sophisticated as their customers, not only to continue to provide suitable products and services, but to utilize new technologies to ensure smooth customer experiences and track their shopping journey in real-time.Among all consumer behaviors, customer defection is the most critical issue they care about because customer churn, if not prevented in time, means the loss of future business.To actively prevent customer defection, businesses need an early detection system.Unfortunately, unlike big companies, it  is difficult for small and medium enterprises (SMEs) to do so because most of them are short in resources, people, and know-how to operate a sophisticated CRM system.

In this paper, we develop a Brownian motion model based on the level, drift and volatility of usage rates of individual customers.An important feature of our model is its simplicity of implementation and interpretation: it can be calibrated with very few observations, at the individual customer level, and using standard business technology (e.g., spreadsheets such as Excel).We build and empirically demonstrate that this approach can simultaneously provide real-time predictions on both future usage rates and churn with performance comparable to several more complex benchmark models.Our approach is simple unlike the “black box” offered by SaaS companies.It leverages key parameters derived from customer's product usage rate and, as we show, these parameters enable useful customer diagnostics linked to future usage and cashflows, and the risks of customer defection.We find that volatility is positively related to and is a better leading indicator of churn than drift.The approach lends itself to simple management and evaluation of customers of SMEs.

Specifically, we contribute to the literature on CRM at a general level and churn management by proposing a simple, easy-to-implement model that uses each customer's historical usage pattern as a leading indicator of churn.The model we develop can greatly aid managers in proactively targeting individual customers at risk of churning.Using a de-classified telecom dataset, we demonstrate that the simple GBM model performs better than several benchmark models and about the same as the more complicated dynamic linear model in predicting customer churn.The model has several managerially desirable features related to implementation, diagnostics, and predictive ability that we discuss below.

Overall, based on several relative and absolute performance metrics, we find that the model performs very well in both short-range and long-range forecasting of whether a customer will defect, and on forecasting usage rates.The model also yields two readily interpretable diagnostics—drift and volatility—that can help monitor customer “health” in managerial dashboards.Customers with negative drift and high volatility, for example, are potentially good targets for intervention.Aside from helping to identify which customers are at risk of defecting, these customer-level risk-return diagnostics can help continuous service providers manage usage-related cash flow risk for active.Firms can segment customers with similar risk-return, i.e., usage rate's drift and volatility, profiles and target them with customized services (e.g., calling plans, add-on phone features).We estimated the impact of customer characteristics, VIP status, changes in features and use of add-on services on these diagnostic parameters.These estimated effects could help managers, for example, in improved assignment of VIP levels to customers to manage usage rate drift and volatility.

We compared the predictive performance of our GBM-based model with several benchmark models commonly used in the marketing literature for churn prediction and found that the individual level GBM model outperformed other models on all but two of the measures considered.We also estimated a dynamic linear model capable of capturing more flexibly the dynamics in usage rates; we found that our GBM-based models performed about the same as this model.Since the predictive performance of the two GBM-based models is equivalent, we favour the “Individual Level” model because it is easier to implement.In cases where managers want to learn about the drivers of usage rate's drift and volatility with the goal of guiding policy toward managing them or making predictions about new customers for whom no usage rate data exist, the “Multilevel” model is appropriate.

We view simplicity, both in estimation and interpretation, as an important strength of the modelling framework, because simple models are more likely to be used by managers.Our model is suitable for most contractual settings, and some non-contractual type settings (such as casual gym memberships), if a customer can be identified across instances and where there is some ongoing usage that can be observed or inferred over time.It would obviously not be useful in contexts where only cross-sectional data are observed or where customers utilize a physical good, service, or experience only sporadically and very infrequently (e.g., wedding dress, birthday party venue).We speculate that our model will become much more relevant to durable consumption, or even health tracking, as increasingly these situations are being digitally transformed and companies can make use of first-party data from such applications.