A hybrid belief rule based decision support system for assessing credit risk for mortgage lending

Mortgage lending institutions assess credit risk to determine the possibility of borrower’s failure to pay their loan obligation. A common decision-making approach adopted by lenders consists of following several administratively pre-established rules and transformation of data gathered from credit bureau to a composite score by machine learning models. With accumulation of data over time, rules have become more complex and the consequence of noncompliance is getting more severe. This situation force lender to minimize the stress and unrecognized biases or errors of complex decision making by human interventions.
This research proposes a hybrid belief rule-based system to make joint utilization of expert knowledge and heterogeneous source of credit risk data available from external agencies, lending institutions internal data and their pre-existing decline and referral rules. Both expert knowledge and credit risk data are independent and complementary and are used to train the optimal expert belief rule-based and data-driven model, respectively to establish the association between the default feature space and default status space. The proposed methodology can determine the nonlinear relationships between default features and can explicitly represent the underwriter’s domain-specific knowledge as well as the judgment from historical data. The result from both the models is fine-tuned by aggregating it by evidence reasoning algorithm.