ADAPTIVE PAYMENT ROUTING IN E-COMMERCE: A MACHINE LEARNING APPROACH FOR HIGHER PURCHASE SUCCESS RATES

Abstract

Online stores need to have a high payment flow success rate to keep their customers and keep their business going. Traditional rule-based routers can’t change their behavior in real time when network conditions or issuers change. We suggest a hybrid machine-learning architecture that com- bines (1) an offline XGBoost classifier to predict the chances of success on each route, (2) an online LinUCB contextual bandit for exploration–exploitation routing, and (3) a rolling-window anomaly detector for quick failover. The system was tested on six months’ worth of production transactions, which included almost 5 million records and an 8 percent failure rate. It kept its decision latency under 30 ms and got a 96.3 percent first-attempt approval rate in offline simulations, which is a

3.6 percentage-point improvement over a 92.7 percent rule-based baseline. In a real A/B test with 200,000 transactions, our ML-driven router had a success rate of 96.7% compared to 94.1% for the old system (p < 0.01). It automatically found and rerouted around a gateway outage to keep perfor- mance up. More tests, such as feature-interaction heatmaps, cost-benefit trade-offs, and calibration curves, show that the results are strong, easy to understand, and cost-effective. We end by going over the main points and talking about where things could go in the future, like advanced reinforce- ment learning and multi-objective routing that finds the best balance between fees, delays, and the chance of getting approval.