AI-Driven Analysis of Information Processing Capacity and Financial Stability in Delegated Asset

Abstract

The fast application of artificial intelligence in delegating asset management has radically changed the information environment in which fund managers conduct their business, but the consequences of this change on financial stability are not effectively comprehended on both the microeconomic and systemic scales. The paper formulates and testingly assesses a theoretical model that builds on the canonical principal-agent model by adding an informational constraint on the managerial capacity in the form of Shannon entropy.  We introduce a novel measure of AI-enhanced information processing capacity (IPC), defined as a composite index derived from transformer-based natural language processing features extracted from fund manager communications, trade-flow complexity metrics, and portfolio diversification indicators. We propose the notion of AI-enhanced information processing capacity (IPC) a composite measure derived by summing transformer-based natural language processing attributes derived based on fund manager communication, trade-flow complexity, and portfolio diversification indicators and test how it is nonlinearly related to fund-level financial stability on a sample of 4,217 delegated asset management funds in the years 20102023. Identification takes advantage of exogenous change in AI adoption due to the difference in exposure to a shock of supply of GPUs, instrumented using firm-level semiconductor procurement data. The fundamental findings are three. First, statistically and economically significant changes in fund-level Value-at-Risk and Conditional Expected Shortfall are positively related to moderate changes in IPC, which is also consistent with the hypothesis that AI augmentation will be able to extract more precise signals and optimise portfolios. Second, the critical IPC threshold is discovered to be at around the 72 nd percentile of the sample distribution, after which marginal capacity returns diminish with increasingly poor stability performance, which is again expected by the theoretical result of amplification of tail-risk by herding when AI-based managers all jump to the same signals. Third, the cross-sectional dispersion of IPC is very much a negative predictor of systemic risk in cases of macro-stress and this indicates that the heterogeneity in AI adoption is a buffer to contagion.  These findings have direct implications for macro-prudential oversight of AI deployment in asset management and highlight the need for disclosure standards that address algorithmic opacity and correlated technological adoption.