What Are The Practical Applications Of Reinforcement Learning?
Reinforcement Learning for Household Finance: Designing Optimal Policy via Responsiveness
Artificial Intelligence & Machine Learning
In a Q-learning model-free setting, the mortgage servicer can learn the default incentive of the borrower from soft information and responsiveness during communication; and in turn undertake appropriate actions to maximize her reward. This reward maximization incentivizes the effort of the servicer and the loan outcome for the borrower via this optimal policy is not worse off.
Moreover, by implementing the optimal policy, the servicer can preempt borrower’s adversarial behavior, thereby increasing borrower cooperation. Our approach differs from the conventional approach which depends on the qualitative legal and industry expert judgement. Furthermore, we also provide evidence that past experiences are not the dominant deciding factors for the optimal action of the servicer in a high learning environment.
Q-learning (model-free reinforcement learning) characterizes the optimal policy of the servicer where the servicer maximizes the overall profit through the life of the loans.
The servicer uses the quantified soft information about the borrowers from their communications (call transcripts) with the borrowers.
Hence, the servicer can alleviate the information asymmetry between the borrower and the servicer
Reinforcement Learning (RL) also learns the noisy borrower type over time via her responsiveness.
To the best of our knowledge, we are the first to use the state-of-the-art RL techniques to design optimal mortgage servicer intervention policy.
We formulate the global optimum set of actions of the servicer and find that the optimal actions from QL are better than locally optimal harsh (sticks) and lenient set (carrots) of optimal actions. Hence, this model-free approach of learning from experience is better than sticks and carrots.
Lastly, we also document a serious finding that in a high learning environment, the discount factor does not dictate the optimal action of the servicer.
Written by Arka Bandyopadhyay
