Beyond CAMELS ratios: A hybrid AI framework for counterparty credit risk monitoring

Abstract

Between 2002 and 2025, 556 US insured depositories failed; more than 90 per cent held assets below US$10bn at the date of failure, locating the historical failure record in the small regional and community-bank population.1 The dominant screening tools, probability-of-default (PD) models, internal credit scores, and external ratings model screens2,3 are supervised classifiers fit on a curated set of accounting ratios and a fixed historical failure cohort, and they generalise poorly across distinct distress regimes (the 2008–2012 credit-loss wave, the 2013–2017 agriculture-and-energy stress wave, and the 2023 interest-rate-driven wave that culminated in the Silicon Valley Bank, Signature Bank, and First Republic Bank failures). This paper proposes a three-layer hybrid artificial intelligence framework. Layer 1 is an unsupervised, peer-relative anomaly engine over the Federal Financial Institutions Examination Council call report data.4 Layer 2 augments it with structured market signals for publicly listed bank holding companies and unstructured signals from regulatory disclosures, news, and social media. Layer 3 produces an analyst-facing triage brief. A population-scale proof of concept covers 4,978 institutions and 610,873 bank-quarter observations over Q1 2002–Q4 2025. On the 21 financial distress failures of 2016–2025 (five fraud-driven cases excluded), the Layer 1 engine flags 90 per cent at a top 10 per cent queue threshold with at least one year of advance warning, and 100 per cent at top 15 per cent. The 2023 Silicon Valley Bank cohort is detected at multiple pre-failure quarters by an engine trained only on data through Q4 2018, demonstrating regime portability without hindsight. The contribution is twofold: an unsupervised triage filter that narrows analyst attention to the right banks, and an analyst-facing screen that narrows attention, on each flagged bank, to the right questions.This article is also included in The Business & Management Collection which can be accessed at https://hstalks.com/business/.