Impact of Imbalanced Data on Bank Telemarketing Calls Outcome Forecasting using Machine Learning

This paper presents a study on bank telemarketing call outcome prediction. This work uses a publicly available dataset and machine learning techniques to predict if a client will subscribe a bank term deposit. The dataset used contains various information regarding the bank client, the telemarketing calls and the outcome of the client subscription to a bank term deposit resulting from the call. The authors tested the results of previous research studies which suggest that bank telemarketing calls outcomes could be predicted using machine learning. Neural Networks, Naive Bayes, Logistic Regression, Random Forest (RF) and AdaBoost methods have been tested. This study aims to analyse the impact of data imbalance. On the one hand, we have tested the machine learning methods in imbalanced data. On the other hand, we have used these models in perfectly balanced data. The results using balanced data suggest the use of RF which reported the best results of 81.3% accuracy, 81.3% F1-score and 88.2% AUC. This work shows a critical impact concerning severe data imbalance on telemarketing call sales outcome prediction. Moreover, the presented findings can support future research activities in this field. The parameterization of the proposed machine learning methods is presented to allow the reproduction of the results.