Article

Reinforcement Learning-Driven Optimization of Payment Gateway Efficiency in FinTech Platforms

---

Abstract

Payment gateways in modern FinTech platforms operate under non-stationary conditions where authorization outcomes, latency, and provider availability drift due to network congestion, issuer behavior, and partial incidents. This paper proposes an offline reinforcement learning framework for adaptive payment orchestration that selects routes and control parameters under compliance-constrained action masking and stability governance. Using a held-out 30-day test window, the learned policy improves overall gateway performance relative to baseline routing by increasing authorization rate from 0.9236 to 0.9281, corresponding to a +0.49 percent relative gain. Efficiency improves through a reduction in median latency from 312.4 ms to 289.7 ms, a 7.27 percent decrease, and a reduction in tail latency (p95) from 921.8 ms to 872.6 ms, a 5.34 percent decrease. Reliability strengthens with a timeout rate reduction from 0.0068 to 0.0052, a 23.53 percent decrease, and a retry rate reduction from 0.081 to 0.067 average retries per transaction, a 17.28 percent decrease. Unit economics improve as cost per successful authorization decreases from 1.012 to 0.995, a 1.68 percent reduction in normalized cost. Segment-level analysis shows the strongest gains in high-volume card traffic, where p95 latency decreases from 910 ms to 858 ms and timeout rate decreases from 0.0065 to 0.0049. Under incident windows, robustness improves materially as p95 latency drops from 1120 ms to 1015 ms while timeout rate decreases from 0.0112 to 0.0086 and retry rate decreases from 0.124 to 0.094. Ablation confirms that rolling telemetry features and conservative offline regularization are primary contributors to stable improvements, while removing share caps increases volatility despite stronger in-sample latency reductions.

Keywords: Payment Orchestration, Payment Routing, Offline Reinforcement Learning, Operational Resilience, FinTech Infrastructure

How to Cite:

Setiawan, I. & Wulandari, A. N., (2025) “Reinforcement Learning-Driven Optimization of Payment Gateway Efficiency in FinTech Platforms”, FinTech Innovation Journal 1(4), 361-387. doi: https://doi.org/10.63913/ftij.v1i4.98