Nonstandard Errors in Capital Shortfall Estimation
Type
conference paper
Date Issued
2026-08-04
Author(s)
Abstract
Market-based systemic risk measures such as MES and SRISK are increasingly used in macroprudential surveillance and capital adequacy analysis for banks and insurers. However, these measures are highly sensitive to empirically defensible research-design choices, including stress-index selection, volatility and dependence models, and tail thresholds. Building on the concept of nonstandard errors (NSEs), we quantify the uncertainty arising from modeling discretion rather than sampling variation (Menkveld et al., 2024). Using a scalable many-analysts framework implemented with large language models (LLMs), simulated research teams select designs from a bounded, literature-consistent specification space that is propagated through a common estimation pipeline. We find that design-induced uncertainty is comparable to, and often exceeds, sampling uncertainty: SRISK exhibits a typical firm-month cross-design standard deviation of about USD 1 billion, rising to several billion for the largest institutions. Although firm rankings are broadly stable, membership of the most systemically important firms varies substantially, driven primarily by the prudential capital ratio rather than the choice of stress index or volatility model. The results provide a transparent framework for reporting specification-driven uncertainty alongside conventional sampling risk, with direct implications for solvency monitoring, capital management, and supervisory stress testing across the financial system.
Language
English (United States)
HSG Classification
None
Refereed
No
Event Title
ARIA 2026 Annual Conference
Event Location
Orlando, FL
Event Date
August 2-5, 2026
Subject(s)
Division(s)