Every Textbox,
Supercharged.

Predicting Systemic Financial Risk Using Graph Neural Networks Trained on Institutional Holdings Data

Antonio Coppola¹, Matteo Maggiori¹, Brent Neiman², Jesse Schreger³

1. Stanford GSB   2. Yale School of Management   3. Columbia Business School

Abstract

We develop a graph transformer architecture trained on quarterly SEC 13F filings to predict systemic risk events across 2,847 institutional portfolios representing $42 trillion in disclosed holdings. The model achieves 91.4% accuracy on out-of-sample stress events (2007-2008, 2020) but exhibits fundamental limitations in causal inference that render it unsuitable for standalone policy decisions.

1. Introduction

Since the 2008 global financial crisis, central banks and regulatory bodies have invested heavily in macroprudential surveillance frameworks designed to detect systemic risk before it materializes. The Federal Reserve's stress testing regime (CCAR/DFAST), the European Central Bank's SRISK methodology, and the Financial Stability Board's G-SIB assessment framework each attempt to quantify interconnectedness risk across the financial system. Despite these efforts, the shadow banking sector—estimated at $63 trillion globally by the FSB (2025)—remains largely opaque to traditional regulatory oversight.

Our approach differs from prior work (Battiston et al. 2012; Acemoglu et al. 2015) in three respects. First, we model the full bipartite network of institutions and securities rather than aggregating to sector-level exposures. Second, our graph transformer architecture captures higher-order interactions between portfolios that linear models miss. Third, we demonstrate that prediction accuracy alone is insufficient for regulatory application—a finding with implications for the growing literature on AI-assisted financial regulation.

2. Data and Network Construction

We construct quarterly bipartite networks from SEC 13F filings spanning Q1 2003 through Q4 2025 (92 quarters). Each network links institutional investors (nodes) to securities (nodes) via weighted edges representing position sizes. The resulting graphs contain 4,200+ institutional nodes connected to 12,000+ security nodes with approximately 2.3 million edges per quarter. We supplement holdings data with CDS spreads, repo rates, and options-implied volatility surfaces to capture funding liquidity dimensions absent from equity holdings alone.

Table 1 presents summary statistics. Mean portfolio size is $9.8B (median $2.1B), reflecting the heavy right tail of institutional AUM. Network density increases monotonically from 0.034 in 2003 to 0.089 in 2025, consistent with the secular trend toward portfolio diversification and the proliferation of multi-strategy funds. Critically, 34% of edges in our 2025 network involve at least one counterparty in the shadow banking sector (hedge funds, private credit vehicles, CLO warehouses), compared to just 12% in 2003.