When Less Is More? Deep Reinforcement Learning-Based Optimization of Debt Collection

Artificial intelligence (AI) presents opportunities to revolutionize financial services. In the current study, we focused on the microloan debt-collection context wherein collectors usually follow a strict sequence of “harsh” debt-collection actions, such as notifying borrowers' social contacts about their delinquencies. We applied a deep reinforcement learning (DRL) algorithm to examine whether such actions are necessary and derived optimal collection strategies on a fine-grained dataset. Installment-level and loan-level optimized results reduced the frequency of “harsh” actions by 49.05% and 60.19%, respectively. This suggests that microloan platforms should deploy collection actions more cautiously. More interestingly, we showed that installment-level and loan-level optimizations suggest somewhat different debt collection patterns across installments in a loan: installment-level optimization overall recommends avoiding the use of any “harsh” actions; by contrast, loan-level optimization recommends using very few actions in early stages but increasing the intensity of applications of (“harsher”) actions in late stages. Generally, loan-level optimization yields a higher recovery rate and greater economic gains than installment-level optimization or other commonly-used debt-collection strategies. This is probably owed to the fact that our DRL algorithm could capture the potential correlations across installments when optimizing at the loan level. Despite the overall superiority of loan-level optimization, our heterogeneity analysis further revealed that installment (loan)-level optimization should be used when borrowers with good socio-economic backgrounds fail to repay early (late) installments in a loan duration. Otherwise, the original intense strategy is more effective. Our findings offer concrete, actionable, and personalized guidance on debt-collection practice