Mining Bad Credit Card Accounts from OLAP and OLTP

TL;DR

This paper presents a method for detecting potentially bad credit card accounts by combining precomputed risk assessments from historical data with real-time transaction analysis, enabling timely identification of risky accounts.

Contribution

It introduces a hybrid approach that precomputes risk probabilities from offline data and updates them with real-time OLTP data for accurate, timely detection of bad accounts.

Findings

Effective risk assessment by combining offline and online data.

Reduced false positives in bad account detection.

Timely identification of risky accounts for preventive action.

Abstract

Credit card companies classify accounts as a good or bad based on historical data where a bad account may default on payments in the near future. If an account is classified as a bad account, then further action can be taken to investigate the actual nature of the account and take preventive actions. In addition, marking an account as "good" when it is actually bad, could lead to loss of revenue - and marking an account as "bad" when it is actually good, could lead to loss of business. However, detecting bad credit card accounts in real time from Online Transaction Processing (OLTP) data is challenging due to the volume of data needed to be processed to compute the risk factor. We propose an approach which precomputes and maintains the risk probability of an account based on historical transactions data from offline data or data from a data warehouse. Furthermore, using the most recent OLTP transactional data, risk probability is calculated for the latest transaction and combined with the previously computed risk probability from the data warehouse. If accumulated risk probability crosses a predefined threshold, then the account is treated as a bad account and is flagged for manual verification.