Scoring AI‐generated policy recommendations with Risk‐Adjusted Gain in Net Present Happiness

Macroeconomic Adverse Selection in Machine Learning Models of Credit Risk

Machine learning algorithms have come to dominate several industries. After decades of resistance from examiners and auditors, machine learning is now mov- ing from the research desk to the application stack for credit scoring and a range of other applications in credit risk.

Current expected credit loss procyclicality: it depends on the model

When the underlying system or process that is being observed is based upon observations versus age, vintage (origination time) and calendar time, Cox proportional hazards models can exhibit instabilities because of embedded assumptions.

n a world where big data is everywhere, no one has big data relative to the economic cycle. Data volume needs to be thought of along two dimensions. (1) How many accounts / transactions / data fields do we have? (2) How much time history do we have? Few, if any, big data sets include history covering one economic cycle (back to 2005) or two economic cycles (back to 1998). Therefore, unstructured learning algorithms will be unable to distinguish between long-term macroeconomic drivers and point-in-time variations across accounts or transactions. This is the colinearity problem that is well known in consumer lending.

In this paper, we explore the role of consumer risk appetite in the initiation of credit cycles and as an early trigger of the US mortgage crisis. We analyze a panel data set of mortgages originated between 2000 and 2009 and follow their performance up to 2014. After controlling for all of the usual observable effects, we show that a strong residual vintage effect remains. This vintage effect correlates well with consumer mortgage demand, as measured by the Federal Reserve Board’s Senior Loan Officer Opinion Survey, and with changes in mortgage pricing at the time the loan was originated. Our findings are consistent with an economic environment in which the incentives of low-risk consumers to obtain a mortgage decrease when the cost of obtaining a loan rises. As a result, mortgage originators generate mortgages from a pool of consumers with changing risk profiles over the credit cycle. The unobservable component of the shift in credit risk, relative to the usual underwriting criteria, may be thought of as macroeconomic adverse selection.

This study evaluates various models for implementing the Current Expected Credit Loss (CECL) method, using a large mortgage data set from Fannie Mae and Freddie Mac

In order to stress test loan portfolios for the impacts of climate change, historical events need to be analyzed to create templates to stress test for future events. Using the 2012 Midwestern US drought as an example, this work creates a stress-testing template for future droughts. The analysis connects weather and crop yield data to impacts on local macroeconomic conditions by comparing drought-impacted agricultural counties with nearby urban counties. After measuring the net macroeconomic impacts of the drought, this was used as an overlay with existing macroeconomic stress models to stress test a lender in a different part of the US for possible drought impacts. Having a library of such climate events would allow lenders to stress test their portfolios for a wide range of possible impacts.

Unintended bias against protected groups has become a key obstacle to the widespread adoption of machine learning methods

Given the unfortunate global headlines about coronavirus, lenders are starting to ask about stress testing the financial impacts of possible pandemics. For some, that is dusting off old plans. For others, it is devising new plans. For myself, it is reliving the past.

The COVID-19 pandemic created abnormal credit risk conditions that did not align well with pre-2020 credit scores. Since the pandemic, most organizations have either excluded the period 2020-2021 from their modeling or included it without adjustment, leaving it as noise in the data. Model validators and examiners have been divided about requiring one of these approaches or defaulting to model developer judgment. None of this is ideal from a model development perspective. We have found that a technical solution is available. Our analysis uses lifecycle and environment outputs from an Age-Period-Cohort analysis as fixed offsets to the credit score development. Panel data is used, so the credit score is developed with a discrete time survival model approach. We tested logistic regression and stochastic gradient boosted regression trees as estimators with the panel data and APC inputs. For this research, we used Fannie Mae data. The APC model was estimated on the full available history, from 2005 through 2024. The origination scores were estimated on two-year periods from 2016 through 2024 and tested on all other periods, including a score that was developed on the full period. All models were also tested on comparably prepared data from Freddie Mac for cross-validation.