Optimizing Payment Network Routing by Refund Prediction

Optimizing Payment Network Routing by Refund Prediction

Van der Voort, H.

Tax, D.J.M. (mentor)
Wolters, B. (mentor)
Houben, G.J.P. (mentor)

Electrical Engineering, Mathematics and Computer Science

Web Information Systems

Information Architecture

2015-03-27

Paying at an online shop is easy for a shopper but the process behind it can become rather complex. The actual transaction travels through the systems of multiple different stakeholders like processors, banks and payment schemes. Multiple available stakeholders allow for optimizations. This work focuses on optimization of transaction routing between payment schemes to minimize transaction costs. Payment routing between schemes is only possible for dual branded payment methods and becomes challenging when schemes differ in functionality. The difference in support for refunds is an example of such a difference which is crucial for retail merchants. In this work a refund predictor for individual transactions is used to predict refund behavior which enables routing between schemes with different support for refunds. The research can be split into two parts; First it is shown that existing evaluation methods cannot deal with the presented criteria, specific to the routing context. The difference in individual transaction costs seem to be uncovered by all methods. In addition an evaluation method, called Current Optimal Instance Score (cOIS), is defined which is based on the realistic loss function from literature. The proposed evaluation method is evaluated in comparison to the loss function from literature. Choosing the right parameters of a predictor using this new evaluation method improves the performance such that 8% of the costs are reduced. For the second part a refund predictor is designed which is able to predict if an individual transaction eventually will be refunded when it enters the system. This predictor is used and optimized for route optimization between schemes which differ in refund support and fees. A sample weight strategy is designed to add some weight to transactions which make the biggest difference in costs. In this research it is shown that transactions costs can be optimized by routing transactions with the knowledge of a refund classifier. Using cOIS, the proposed evaluation method, the final system is validated at a score of 0.487. This result shows that we are halfway in optimizing costs, from the current costs to the theoretical optimum. In practice this system resulted in a cost reduction of 36% which equals to 1.4% of a merchant its profits.

http://resolver.tudelft.nl/uuid:471f8e17-0008-415e-9ce9-53164d0c79bb

2020-03-19

Student theses

master thesis

(c) 2015 Van der Voort, H.