This paper studies default outcomes and fraud classification in UK COVID-19 government-guaranteed lending using loan-level data on 1,107,934 limited-company facilities under the Bounce Back Loan Scheme (BBLS) and the Coronavirus Business Interruption Loan Scheme (CBILS). Combining logit models, Cox proportional hazards, and accelerated failure time specifications with propensity-score matching, entropy balancing, and external validation against Companies House records, it characterises both the incidence and timing of default under contrasting scheme designs and isolates the contribution of lender classification practice to measured fraud.
Three findings emerge. First, observed fraud reflects lender detection and classification as well as underlying borrower behaviour. Conditional on default and observable borrower characteristics, challenger and fintech lenders classify fraud at 37 and 46 per cent of the major-bank rate respectively. A propensity-score benchmark calibrated on major-bank practices implies a substantial volume of under-classified fraud, corroborated by external validation on 2,266 objectively ineligible loans, of which major banks classify 45.6 per cent as fraud against 24.5 per cent at challenger banks and 21.4 per cent at fintech lenders. Second, scheme design is a primary determinant of risk: relative to CBILS, BBLS loans exhibit substantially higher default and fraud-classification rates and reach default more quickly. Third, conventional credit-risk indicators — pre-loan probability of default, leverage, repayment extensions, and governance characteristics — remain informative despite the pandemic setting.
The results imply that administrative measures of fraud in delegated guarantee schemes are not directly comparable across lenders and may understate true exposure when investigative capacity differs. More broadly, the findings highlight how crisis lending design affects not only the level and timing of default but also the measurement of adverse outcomes.
The paper is available at SSRN : https://dx.doi.org/10.2139/ssrn.6587579