Browsing by Subject "supervised learning"

Lenders assess the credit risk of loan applicants from both affordability and indebtedness perspective. The affordability perspective involves assessing the applicant’s disposable income after accounting for regular household expenditures and existing credit commitments, a measure called money-at-disposal or MaD. Having an estimate of the applicant’s expenditures is crucial, but simply asking applicants for their expenditures could lead to inaccuracies. Thus, lenders must produce their own estimates based on statistical or survey data about household expenditures, which are then passed to the MaD framework as input parameters or used as control limits to ascertain expenditure information reported by the applicant is truthful or at least adequately conservative. More accurate expenditure estimates in the loan origination would enable lenders to quantify mortgage credit risk more precisely, tailor loan terms more aptly, and protect customers against over-indebtedness better. Consequently, this would facilitate the lenders to be more profitable in their lending business as well as serve their customers better. But there is also a need for interpretability of the estimates stemming from compliance and trustworthiness motives. In this study, we examine the accuracy and interpretability of expenditure predictions of supervised models fitted to a microdataset of household consumption expenditures. To our knowledge, this is the first study to use such a granular and broad dataset to create predictive models of loan applicants’ expenditures. The virtually uninterpretable "black box" models we used, aiming at maximizing predictive power, rarely did better accuracy-wise than interpretable linear regression ones. Even when they did, the gain was marginal or in predicting minor expenditure categories that contributed only a low share of the total expenditures. Thus, ordinary linear regression is what we suggest generally provides the best combination of predictive power and interpretability. After careful feature selection, the best predictive power was attained with 20-54 predictor variables, the number depending on the expenditure category. If a very simple interpretation is needed, we suggest either a linear regression model of three predictor variables representing the number of household members, or a model based on the means within 12 "common sense groups" that we divided the households in. An alternative solution with a predictive power somewhere between the full linear regression model and the two simpler models is to use decision trees providing easy interpretation in the form of a set of rules.