Predicting long-term deposit customers using convolutional neural network and data conversion technique

Abstract

The banking industry is the foundation of any nation’s economy, and bank deposits are its primary source of profitability. Bank deposits play a significant role in determining a nation’s saving rate. Globalization has resulted in substantial technological changes, business strategy, and customer service across many industries, including the financial services sector. This study proposes a deep learning model using Residual Network architectural design and transfer learning on a Portuguese banking institution containing 40,811 training data, with 36,202 belonging to label 0 and 4639 belonging to label 1. Clearly, this shows a significant level of bias between the two labels. Hence, a SMOTE method of class balancing was applied. This dataset, in comma-separated value (CSV), was converted into images coupled with the weight transfer from the residual network trained on ImageNet; our fully connected layer was built and trained with the image files. The highest performance reached by the conventional machine learning models, Random Forest (RF), is 90.78% for accuracy, 59.37% precision, 96.78% recall, and 85.28% F1 score, tested on 412 test samples. However, our proposed methodology achieves an outstanding result with an accuracy of 93.00%, 97.00% precision, 90.00% recall, 93.00% F1 score, and 94.00% ROC, with test samples of size 5601. Since long-term deposits are necessary for the banking system to fund the individual, corporate, and industrial loans needed for the country’s growth and development, these results will provide an effective and reliable marketing technique required to determine the target population.