Clico is a free AI writing assistant for Chrome
Every Textbox,
Supercharged.
Clico is a free AI writing assistant for Chrome. Select text for answers, summarize articles, and draft replies without leaving your tab.
Free to use · No API key needed
AI writing at your cursor
Open any text field. Press the shortcut.
Clico sees the page and writes for you.
⌘ + O
Summarize any page in 2 taps
Double-tap ⌘ and get a structured summary of whatever you're reading.
⌘ + ⌘
Predicting Systemic Financial Risk Using Graph Neural Networks Trained on Institutional Holdings Data
Antonio Coppola1, Matteo Maggiori1, Brent Neiman2, Jesse Schreger3
- Stanford GSB 2. Yale School of Management 3. Columbia Business School
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.
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.
Key Findings
- What it does: Maps $42T in holdings across 4,200+ institutions to detect hidden systemic risk concentrations
- Key limitation: Model identifies where risk is building but cannot explain why, making it unsuitable for standalone policy decisions
- Moral hazard: If banks know how the AI monitors them, they can shift risk to unmonitored vehicles ($63T shadow banking sector)
- Recommendation: “Model-informed” hybrid approach — use AI for detection, traditional economics for causal reasoning and policy
Search It. Instant answers.
Highlight any word or sentence on any page. Clico instantly tells you what it means, gives you context, or searches it for you. No new tab. No copy-paste. Just select and know.
Speak when you don't want to type
Hold ⌘ and speak. Clico transcribes, understands context, and writes for you.