Using Machine Learning To Find Payment Anomalies
Popular Electronic Payment system with over 17 millions active users approached Vostok AI team to analyze their existing eWallet solution. The idea was to identify why some of the payments have failed and if there are any technical issues that can be fixed. Although the percentage of failed payment was low (~0.25% of the total number), number of transaction per month was huge (75,000,000 payments), thus making it a very critical part of the Customer’s business.
Customer Challenge.
Our customer is publicly traded financial firm. Using their eWallet users can make transfer between people, pay bills, and do online store purchases. The company also provides acquiring services for businesses which allows to accept non-cash payments. Key indicator for acquiring is payment conversion, thus main task from customer side was to analyze why some of the thousands of payments failed and what could be the reasons behind (technical or business).
The challenge was in the number of transactions - more than 100 million of transactions generated within 1.5 months
Solution.
In order to identify signs of anomalies due to which payments are not going through, we decided to separate main scope of work in three main areas of where payment error can be:
On the side of acquiring system;
On the provider side;
On the side of banking system.
This helped us to form separate hypothesis and patterns for each of the areas.
Results.
The final result of analysis brought some valuable business points to management considerations, for instance:
Found top 5 errors while making payments;
Identified 3x main phones which has 60% of the errors;
Identified top 3x telco in terms of payment fails
Large number of payment fails was between 1am and 3am, this is due to incorrect timezone set on a telco side.
The result was a significant change of IT strategy of the company in order to fix existing problems and provide reliable payment infrastructure.