Regulating Algorithmic Credit Scoring Presented by: Mrs. Langenbucher Date: 31 March 2022 Credit Scoring – from intuition to A.I. - Historically, loan decisions were based on a mix of qualitative and quantitative information - There was a risk of cognitive errors and implicit biases distorting the assessment of credit default risks - Currently, most established agencies rely on a fixed number of input variables (e.g. free income, past credit history) to produce standardized scores - Novel scoring agencies vastly expand input variables to produce a variety of alternative scores Algorithmic Credit Scoring - Novel agencies collect various alternative data (e.g. web browser data, purchasing patterns, location of the computer used, Facebook friends, typos in text messages, taste in music, font types found on the electronic device, time needed to fill out an application, diligence in charging smartphone) - Then, they make the A.I. establish correlations between "input" data and a "definition of success" in order to generate alternative credit scores Promises and Risks - Algorithmic scoring promises better access for the "credit invisible" who don't have the credit history to inform the traditional factors - Concerns include • Data privacy • Fair lending • Explainability of use of data/credit decisions • Institutional design of regulators Data Privacy - EU • Omnibus approach: GDPR - Processing of data is illegal, unless exemptions allow for data processing - • Consent, Art. 6 • Special safeguards for protected data, Art. 9 • Information on data stored and on „purpose of the processing", Art. 13-15 • Focused approach: Fair Credit Reporting • Act - Processing of data is legal, unless prohibitions apply • „Consumer reporting agency" • „Consumer report" • A „permissible purpose" is required for handing over data to a third party • Consumer have rights to access information and to have the denial of credit US explained An Example: Upstart Fair Lending Direct discrimination/ Intentional disparate treatment EU • • • • Antidiscrimination Directives ECJ: Fundamental Human Rights' impact on private law Varying regimes of Member States law Indirect discrimination/ Defenses Disparate impact: facially neutral rule creates a disadvantage • Accepted general Objectively justified by principle • a legitimate aim and • (1) Define two groups • appropriate and necessary means • (2) Fix acceptable disparity Direct discrimination/ Intentional disparate treatment US Equal Credit Opportunities Act • No discriminatory treatment • Explanation why loan request was denied • Civil liability Indirect discrimination/ Disparate impact: facially neutral rule creates a disadvantage • Supreme Court has accepted disparate impact doctrine in housing & employment • Consumer Financial Protection Bureau and some courts suggest applying ECOA Defenses • • No alternative practice with less discriminatory results Manifest business necessity Why Can't We "Just Outlaw Using The Variable Race"? - In order to "outlaw" using the variable race, we must first divide our data into variables which are correlated with race, not-correlated with race and race itself: • - y = βo+ Xrace × β1+ Xcorrelated × β2 + Xnot-correlated × β3 Arguably the doctrine of anti-discrimination calls for an elimination of the correlated, as well as of the race matrix, rendering the equation to: • - y = Bo+Xnot-correlated X β3 However, studies have found that nearly all variables corelate with race (and those that do not are usually not very informative), thus, this approach renders the algorithm inutile. Institutional Design - - US • Consumer Financial Protection Bureau • Federal Trade Commission • Office of Financial Research e.g. Germany • BaFin - unclear jurisdiction for anti- discrimination or data protection • Data protection authorities • Anti-discrimination - public authorities, courts Next Steps - - - Refine the GDPR regime, notably • consent" • information on the purpose of the processing"/"explainability" Refine indirect discrimination doctrine, notably • „legitimate aim"-1: private actors and indirect discrimination • „legitimate aim"-2: financial stability and indirect discrimination Discuss institutional design, notably • Data and model quality checks for financial regulators • Establishing a central office for financial data" (cf. US Office of Financial Research)?