The Optimal Use of AI in Financial Regulation

 

The Optimal Use of AI in Financial Regulation
By Christopher Clayton, Antonio Coppola

NBER Publication

Observations

Introduction

This paper explores how graph-based deep learning can revolutionize macroprudential financial regulation by analysing large-scale portfolio holdings data. The authors develop an innovative model tailored to security-level network structures among financial intermediaries to better anticipate systemic risks. By applying this framework to a massive dataset of non-bank financial institutions, they demonstrate how advanced AI can outperform traditional econometric approaches. Ultimately, the study bridges the gap between predictive machine learning and structural economic policy to maximize social welfare.

 

Key Points

Core Objective of the Research

• The study evaluates whether advanced AI methods applied to large-scale portfolio holdings can improve macroprudential financial regulation.
• It addresses the historical limitations regulators face when trying to predict systemic risk using fragmented or static data models.
• The authors aim to transform raw, complex financial data into actionable insights for stabilizing the broader financial system.
• By focusing on predictive accuracy, the research seeks to give regulators a proactive rather than reactive toolkit.

Graph-Based Deep Learning Architecture

• The authors construct a specialized graph-based deep learning model tailored to security-level data of financial intermediaries.
• This architecture treats the financial system as a dense network where both investors and assets act as interconnected nodes.
• It is uniquely designed to map the multi-layered relationships formed when multiple institutions hold identical or correlated assets.
• This structural approach allows the AI to capture complex, non-linear dependencies that standard linear models completely miss.

Integration of Economic Priors

• Unlike pure "black box" AI, this model explicitly incorporates fundamental economic priors into its deep learning architecture.
• It leverages established financial theories regarding investor behaviour, market liquidity constraints, and portfolio optimization strategies.
• Embedding these economic constraints ensures the model's predictions remain grounded in plausible real-world financial mechanisms.
• This hybrid approach combines the sheer predictive power of machine learning with the logical rigor of economic theory.

Learning Latent Representations

• The model successfully learns latent, hidden representations of both assets and investors from the portfolio network structure.
• These latent variables capture unobserved characteristics, such as an investor's true risk appetite or an asset's hidden vulnerabilities.
• By embedding these entities into a continuous mathematical space, the model quantifies similarities between seemingly unrelated institutions.
• This allows regulators to identify hidden clusters of systemic risk that do not appear in traditional sector classifications.

Massive Empirical Scale and Scope

• The empirical framework is applied to the global universe of non-bank financial intermediaries (NBFIs), such as hedge funds and asset managers.
• The dataset used in the study represents an astonishing scope, covering nearly $40 trillion in total institutional wealth.
• Managing a dataset of this magnitude requires the advanced computational scaling properties inherent to deep learning frameworks.
• The sheer scale proves that the model is robust enough to handle the immense data burdens of modern global regulators.

Outperforming Existing Forecast Models

• The graph-based model substantially outperforms existing econometric approaches in out-of-sample forecasts of intermediary trading behaviour.
• It accurately predicts how specific institutions will rebalance their portfolios when faced with shifting macroeconomic conditions.
• This superior predictive capacity holds true across different asset classes, validating the generalized nature of the network architecture.
• It demonstrates that tracking network connections is vastly superior to tracking isolated institutional balance sheets.

Superior Performance During Crisis Episodes

• Crucially, the model retains its high predictive accuracy during historical periods of intense financial market distress and systemic crises.
• Traditional models often break down during crises because correlations shift rapidly and market liquidity suddenly evaporates.
• This AI architecture successfully anticipates the frantic selling patterns and portfolio liquidations that characterize market panics.
• Providing reliable forecasts during stress events makes it an invaluable early-warning tool for real-world central banks.

Explaining Cross-Sectional Asset Returns

• The model boasts more than ten times the explanatory power of traditional approaches for cross-sectional asset return variations.
• It specifically isolates how asset prices diverge during severe market stress events based purely on who holds them.
• This reveals that an asset's risk during a crisis is heavily driven by the overlapping portfolio networks of its investors.
• Achieving a tenfold improvement over benchmark models represents a massive quantitative leap forward for empirical financial economics.

Enhanced Systemic Risk Metrics

• At the individual institution level, the framework consistently outperforms existing regulatory systemic risk metrics like CoVaR or SRISK.
• It provides a more granular view of how a single firm's distress can cascade across the entire financial network.
• Regulators can use these metrics to pinpoint exactly which institutions are the most critical linchpins of systemic vulnerability.
• This prevents the misallocation of regulatory oversight by accurately ranking firms by their true potential to cause contagion.

Revealing Fire-Sale Vulnerabilities

• The model's learned representations demonstrate that the holdings network encodes rich, economically interpretable information regarding fire-sales.
• It maps out how forced liquidations by one distressed fund drop asset prices, triggering margin calls for neighbouring funds.
• This visualization allows regulators to see the precise pathways through which a localized shock turns into a systemic fire-sale.
• Understanding these pathways helps policymakers intervene early to break the vicious cycle of forced asset liquidations.

Fully Inductive Model Architecture

• The AI architecture is designed to be fully inductive, meaning it can generalize its findings to entirely unseen entities.
• It produces highly informative risk estimates even when entire asset classes or specific investors are completely withheld from training.
• This solves a massive machine learning hurdle by ensuring the model does not fail when encountering novel financial instruments.
• Regulators can confidently deploy the model in evolving markets where new types of funds or securities frequently emerge.

Embedding in an Optimal Policy Framework

• The authors explicitly embed their empirical AI approach directly into a macroprudential optimal policy and welfare framework.
• This theoretical integration formalizes exactly why these highly accurate predictive objects matter for high-level policymaking.
• It shifts the conversation from a purely statistical exercise to a concrete tool designed to maximize economic stability.
• The framework outlines how AI outputs can be mathematically translated into optimal regulatory actions like capital requirements.

Navigating the Lucas Critique

• The paper proves that the model improves economic welfare even in an equilibrium environment subject to the famous Lucas critique.
• The Lucas critique warns that historical parameters change when policies change, often rendering predictive models useless after an intervention.
• Because this model captures deep, underlying structural network positions, its predictive utility remains robust against changing policy regimes.
• This finding provides a powerful defence for using advanced predictive AI in deep macroeconomic policymaking.

Sharpening Policy Targeting

• The predictive information generated by the AI significantly improves welfare by sharpening the cross-sectional targeting of interventions.
• Instead of enforcing broad, inefficient regulations across the entire market, policymakers can surgically target specific vulnerable nodes.
• This minimizes the economic deadweight loss and compliance costs typically associated with sweeping, heavy-handed financial regulations.
• Precise targeting ensures that regulatory restrictions are only placed on the entities posing genuine systemic threats.

Complementarity of Prediction and Structure

• The research demonstrates a powerful structural complementarity between pure machine learning prediction and economic knowledge.
• It proves that predictive AI is most effective when guided by structural models, and structural models are enhanced by AI data.
• Neither data-driven algorithms nor abstract economic theory are sufficient on their own to manage modern financial systems safely.
• Blending both disciplines creates a superior regulatory framework capable of safeguarding trillions of dollars in global wealth.

Key lessons for financial regulators

Map the Network, Not Just Individual Balance Sheets

Traditional regulation heavily relies on analysing individual institutional balance sheets in isolation. This study proves that systemic risk is inherently structural and relational. Regulators can transition to mapping the dense, multi-layered networks of overlapping portfolio holdings. The true vulnerability of an asset or institution is dictated by who else is connected to it through shared exposures.

Embrace Graph-Based AI for Massive Datasets

When dealing with massive, complex sectors—such as the $40 trillion non-bank financial intermediary (NBFI) universe—traditional econometric tools fall short. Regulators need to adopt graph-based deep learning architectures. These advanced AI models excel at processing security-level big data and capturing the complex, non-linear dependencies that standard linear models completely miss.

Look Past Industry Labels to Find "Latent" Risks

Traditional regulatory buckets (e.g., grouping firms strictly as "hedge funds," "asset managers," or "insurance companies") can obscure actual market behaviour. The study shows that AI can learn "latent representations"—hidden, unobserved characteristics encoded in trading patterns. Regulators should use these AI-generated embeddings to identify clusters of institutions that share the same underlying risk appetites and vulnerabilities, regardless of their official sector labels.

Prioritize Early-Warning Capability for Crisis Episodes

Many predictive models perform well during calm periods but break down completely during market panics because historical correlations shift overnight. This study demonstrates that a network-informed AI model can successfully maintain high out-of-sample predictive power during severe stress events. Regulators could leverage this capability to forecast frantic portfolio rebalancing and liquidation patterns before they trigger a systemic collapse.

Decode the Specific Pathways of Fire-Sale Contagion

The model reveals that portfolio networks hold rich, economically interpretable data regarding fire-sale vulnerabilities. It shows exactly how a forced liquidation by one distressed fund depresses asset prices, triggering margin calls and cascading distress for neighbouring funds. Regulators can use these network maps to trace the precise pathways of contagion, allowing them to intercept a localized shock before it spirals into a market-wide fire-sale.

Combine Machine Learning with Economic Structure

Regulators do not have to choose between "black-box" data algorithms and rigid economic theory. The study highlights a powerful complementarity: AI is most effective when it incorporates economic priors, and structural policy models are vastly improved by AI's predictive precision. Blending both disciplines allows regulators to build frameworks that are both highly accurate and economically logical.

Use Predictive AI for Surgical, Welfare-Optimizing Interventions

A common fear in macroeconomics is the Lucas critique, which warns that changing a policy can alter human behaviour and make previous data predictions useless. The authors prove that because graph-based AI captures deep, underlying structural network positions, its insights remain robust against changing policy regimes. Instead of enacting broad, costly regulations across the entire market, policymakers can use AI to surgically target specific vulnerable nodes. This high-precision targeting minimizes economic deadweight loss and significantly improves overall social welfare.

Summary for Policymakers

Advanced AI shouldn't just be viewed as a tool for commercial finance; it is a critical asset for public policy. By embedding graph-based deep learning into optimal policy frameworks, regulators can shift from a reactive stance to a proactive, surgical, and mathematically grounded approach to safeguarding global financial stability.

Conclusion

In conclusion, Clayton and Coppola deliver a compelling defence for integrating graph-based machine learning into the foundation of macroprudential policy. By effectively mapping the $40 trillion non-bank financial sector, their model proves that network structures hold the key to predicting systemic crises and fire-sales. The framework successfully overcomes traditional modelling barriers, including the Lucas critique, by blending economic theory with deep learning. Ultimately, this research provides global regulators with a precise, scalable, and mathematically sound toolkit to protect modern financial ecosystems.

Abstract:

We study whether AI methods applied to large-scale portfolio holdings data can improve macroprudential financial regulation. We build a graph-based deep learning model tailored to security-level data on the holdings of financial intermediaries. The architecture incorporates economic priors and learns latent representations of both assets and investors from the network structure of portfolio positions. Applied to the universe of non-bank financial intermediaries, covering nearly $40 trillion in wealth, the model substantially outperforms existing approaches in out-of-sample forecasts of intermediary trading behavior, including in crisis episodes. The model has more than ten times the explanatory power for the cross-sectional variation in asset returns during stress events compared to traditional approaches, and it outperforms existing systemic risk metrics at the institution level. Its learned representations show that the holdings network encodes rich, economically interpretable information about fire- sale vulnerability. The architecture is fully inductive, producing informative estimates even when entire asset cl-asses or investors are withheld from training. We em-bed our empirical approach into a macroprudential optimal policy framework to formalize why these ob-jects matter for policy and welfare. We show that even in an equilibrium environment subject to the Lucas critique, the predictive information from the model improves welfare by sharpening the cross-sectional targeting of policy interventions, and we demonstrate a complementarity between prediction and structural knowledge.