Intuitive Feature Engineering and Machine Learning Performance Improvement in the Banking Domain

Abstract

Performance tuning of the machine learning models is very important, especially in the banking domain. In line with the new age, they are moving away from their conventional methods to target customers for credit cards, loans, etc. products. The transactional data, customer information, which was collected over the years, have a huge scope of applying data mining techniques to extract useful information for maximizing the return on investment, cost optimization and fraud detection. For a successful deployment of a machine learning model multiple out of time validations are performed and stability is strictly evaluated. Here, we propose a cardinal method for intuitive use case related feature engineering, tuning the hyper parameters, best model selection and diagnosing the model for further improvements. The stability of a model plays a huge factor, as we expect the deployed model to work well for the next 6–8 months, then up for re-tuning based on the data distribution and model performance. Statistical and data driven methods are used to develop sophisticated features and achieved minimal accuracy variation across time periods. Implementation of our methods for the use cases like customer attrition from the bank in the next 6 months and detection of the in-bound calls from the customer to the call centre to enquire about balance, transaction details, etc. are discussed. Achieved 47% & 55% recall score in the top 2 deciles respectively for the use cases.