A Study on Improving Banking Process for Predicting Prospective Customers of Term Deposits using Explainable Machine Learning Models

Abstract

This study has been carried out to provide a sophisticated machine learning approach for the marketing prediction problem in banking systems. Classical machine learning models like linear regression, logistic regression, decision tree, KNN, and decision rules are useful to machine learning engineers as well as business managers as these are easy to explain and provide good feedback to businesses. Individually, these models have not performed as per business expectations in real-time scenarios. Complex ML models like ANN can perform better and have higher accuracy as well, but the majority of these models are complex, and these ML approaches struggle to provide necessary feedback and act as black-box models. This trade-off of useful feedback for businesses and the accurate result makes such models not so desirable in such domains where both are required by businesses. Thus, the objectives of this study are (i) to suggest a suitable machine learning model to the banks in order to provide the explainable AI-based solution for predicting prospective customers of term deposits, (ii) to provide explainable AI-based solution which banks can use to gain insights and feedback and use them to improve the banking process further with a goal of not losing even a single potential customer. The present study meets that challenge and the requirements by analyzing the dataset taken from UCI ML repository, where the data are of Portuguese banks for term deposits telemarketing campaign. To achieve the objective of providing a solution to the banks for predicting prospective customers of term deposits and further to gain insights and feedback to improve the banking process with a goal of not losing even a single potential customer, we have set the target recall score as 0.99, which may act as a benchmark for ML models used in this study. Further, all the ML models used in the study have the same count for the training dataset which enables us to evaluate model performance objectively. In the evaluation, we have used recall score, FPR value, and AUC score as metrics for model evaluation as accuracy as any other metric alone cannot help us with evaluation and reach the correct conclusion. This study focused on interpretability and feedback to the banks. ML models used for this study are good examples of explainable AI. Thus, this approach can support real-world problems of the financial and banking sector which involves predicting how likely a person can be turned into a customer.