Model Accuracy
Overview
Our client, one of the leading consumer bank, was facing heavy financial and credibility losses due to increase in fraudulent transactions. Client wanted to identify high risk accounts and avoid frauds.
Approach
TransOrg stitched and cleaned the data from various sources to perform deep exploratory analysis to identify fraudulent accounts.
Transactions from these fraudulent account were analysed to identify patterns at time, amount, merchant level to red flag such transactions in future.
Based on the patterns, TransOrg created machine learning models to predict the likelihood of occurrence of fraud by providing an “anomaly score” for every customer.
Output
- Model accuracy: Top 1.11% scorers capture 43% frauds and Top 10% scorers capture 81% frauds
- A list of accounts with high anomaly score in every prediction cycle was shared with Client’s Anti Fraud Unit for verification
- A scheduler was created in order to give “anomaly score” at regular intervals
- Model performance tracker was also created and retraining was automated, in the case accuracy drops down a certain threshold