Open Access

Privacy-Aware Multi-Lender Routing for Point-of-Sale Finance using Reinforcement Learning: A Simulation Study

4 Independent Researcher Jersey City, NJ, USA

Abstract

Many point-of-sale lending platforms run a credit application through a waterfall: prime lender first, then near-prime and sub-prime lenders if declined. Each fallback means a call to the checkout by a lender and another institution receiving the applicant’s data. This paper investigates if a reinforcement learning policy can select the lender or choose not to route at all, solely based on the credit score and the requested amount. A Proximal Policy Optimization (PPO) agent is trained in a synthetic three-lender simulator to trade conversion against latency, lenders contacted and declines, where routing is modeled as a Markov decision process. Over 5 seeds, each evaluated on 10,000 held-out simulated applications, PPO raised net conversion from 62.7% to 70.8% relative to the waterfall, cut mean lender latency from 2,944 to 1,476 ms, and reduced lenders contacted from 2.41 to 0.74 per application. The conversion gain depends on the abandonment of slow sessions, which is assumed by the simulator. Without abandonment the waterfall converts more than PPO. The gains are also dampened by stronger baselines. 73.2% change for single-lender routing and parallel broadcast. The learned policy never takes into account the prime or near-prime lender. It sends applicants to the sub-prime lender or declines 26.2% of them outright It acts very much like a single-lender rule with a tuned score cut-off. Solving the same decision process exactly shows that the reward does favor matching applicants to lender tiers, yet neither longer training nor a price penalty led PPO to find that behavior. The paper also discusses the limitations of the simulator and the consumer protection issues raised by early declines and tier selection.

Keywords

References

Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 (General Data Protection Regulation), Art. 5(1)(c) and Recital 26, Official Journal of the European Union, L 119, pp. 1–88, May 2016.
California Consumer Privacy Act of 2018, as amended by the California Privacy Rights Act of 2020, Cal. Civ. Code § 1798.100(c).
R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018.
J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv:1707.06347, 2017.
A. Agrawal and H. Patil, “A control-theoretic approach to dynamic payment routing for success rate optimization,” arXiv:2510.16735, 2025.
L. Li, W. Chu, J. Langford, and R. E. Schapire, “A contextual-bandit approach to personalized news article recommendation,” in Proc. 19th Int. Conf. World Wide Web (WWW), Raleigh, NC, USA, 2010, pp. 661–670, doi: 10.1145/1772690.1772758.
S. Lessmann, B. Baesens, H.-V. Seow, and L. C. Thomas, “Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research,” Eur. J. Oper. Res., vol. 247, no. 1, pp. 124–136, 2015, doi: 10.1016/j.ejor.2015.05.030.
P. Henderson, R. Islam, P. Bachman, J. Pineau, D. Precup, and D. Meger, “Deep reinforcement learning that matters,” in Proc. AAAI Conf. Artif. Intell., vol. 32, no. 1, 2018, pp. 3207–3214, doi: 10.1609/aaai.v32i1.11694.
Fair Credit Reporting Act, 15 U.S.C. § 1681b(a)(2)–(3), “Permissible purposes of consumer reports.”
Equal Credit Opportunity Act (Regulation B), 12 C.F.R. § 1002.9(a)(1)(i) and (g), “Notifications.”
A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-Baselines3: Reliable reinforcement learning implementations,” J. Mach. Learn. Res., vol. 22, no. 268, pp. 1–8, 2021.
M. Towers et al., “Gymnasium: A standard interface for reinforcement learning environments,” in Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2025, arXiv:2407.17032.

Most read articles by the same author(s)

1 2 3 4 5 6 7 8 9 10 > >> 

Similar Articles

21-30 of 75

You may also start an advanced similarity search for this article.