Autoencoder-LSTM Ensembles for Predictive Analysis of Financial Churn Patterns

In the financial sector, customer churn is a significant problem that influences the organization’s revenue growth. Churn means that customers have left their current providers which is a negative factor for the providers or financial institutions; therefore, accurate churn prediction helps institutions to increase customer loyalty and satisfaction by improving the quality of service. In this paper, a new method of using Autoencoder and LSTM network for predicting financial churn patterns is introduced. The Autoencoder is used to reduce dimensions and for unsupervised learning acts as a pass through that allows the model to learn the latent patterns in customer behaviour and transactions. Since the LSTM has demonstrated its efficacy in analyzing temporal dependencies of a sequence in time series, it is used to predict churn by analyzing sequential customer activity. The ensemble approach combines the merits of the two models improving the predictive power and especially the stability in a situation of data scarcity and high noise levels. Experimental comparison with conventional predictive models such as decision trees and logistic regression reveal that the Autoencoder-LSTM is more accurate and sensitive than traditional models by comparison. In other words, the proposed model enhances the early identification of potential churners and such institutions can then take appropriate measures of containing it, and in the process, offer relevant retention solutions. The paper describes the benefits of utilizing deep learning models for churn prediction for the financial industry and provides information on how such methods might be used to improve customer retention. Possible future research avenues consist of improving the model interpretability and generalizing the approach to other industries with comparable churn prediction requirements. The accuracy obtained from the proposed method is higher at 99.7% compared to other methods.