Reinforcement learning for household finance: Designing policy via responsiveness

We use model-free reinforcement learning (RL) to explore how a mortgage servicer can optimize her actions toward a borrower. Unlike conventional heuristic approaches, our methodology eliminates reliance on subjective and qualitative judgments from industry and legal experts. We are the first to incorporate post-securitization soft information and the borrower’s responsiveness to the servicer to estimate an RL-based policy rule. When maximizing her reward, the servicer dynamically learns the type of borrower, allowing it to anticipate and mitigate adversarial behavior. This, in turn, fosters greater borrower cooperation and improves overall outcomes.