IMF Study Merges Machine Learning with Monetary Policy to Strengthen Risk Frameworks

The IMF, University of Tokyo, and BIS study by Yuji Sakurai presents a data-driven framework that uses econometric and machine learning models to help central banks set collateral haircuts and manage counterparty risks more transparently and consistently. It combines risk metrics, cross-country modeling, and Variational Autoencoders to ensure haircut policies remain conservative, non-procyclical, and resilient under financial stress.

The International Monetary Fund’s Research Department, in collaboration with the University of Tokyo and the Bank for International Settlements, has released a pioneering working paper titled “A Quantitative Approach to Central Bank Haircuts and Counterparty Risk Management” (IMF Working Paper WP/25/225) by Yuji Sakurai. The study offers a data-driven, model-based framework for determining collateral haircuts and managing counterparty risk, issues that sit at the core of financial stability. Since the global financial crisis of 2008, central banks have been rethinking collateral policies to prevent liquidity operations from amplifying market stress. Sakurai argues that haircuts, the margins applied to collateral to protect against value declines, should be set through robust quantitative models that ensure transparency, consistency, and resilience under stress.

Rethinking Haircuts: From Judgment to Quantification

Haircuts act as a cushion, protecting central banks against collateral devaluation during crises. Sakurai’s framework redefines their calculation using tail-risk measures such as Value-at-Risk (VaR) and Expected Shortfall (ES), showing that these approaches can be equivalent when calibrated correctly. The model treats asset prices as stochastic processes and reveals that haircut levels depend primarily on volatility, time-to-liquidation, and the central bank’s confidence level. Crucially, he warns that raising haircuts during downturns can trigger a “liquidity spiral” by forcing asset sales. Instead, haircuts should be pre-stressed, calibrated conservatively during normal times, to avoid procyclicality.

The paper also models residual exposures, the uncollateralized risks that remain after applying haircuts, as financial options. This approach reveals how such exposures grow nonlinearly with volatility and liquidation periods, underscoring the need for dynamic calibration rather than relying on static historical averages.

Four Models for a Complex Financial World

Sakurai introduces four distinct models suited to different asset types and data environments, enabling central banks across varying levels of sophistication to apply the framework. For marketable assets in data-rich settings, the DASV model (Duration Approximation with Stressed Volatility) estimates haircuts by combining bond duration with volatility derived from an EGARCH model, capturing both clustering and asymmetry in yield movements. It also includes adjustments for liquidity and credit quality, ensuring realistic and conservative risk assessments.

For countries with limited financial data, the DACR model (Duration Approximation with Cross-Country Regression) fills data gaps through regression-based estimation of sovereign spreads using macroeconomic indicators such as debt-to-GDP, inflation, exchange rates, and trade openness. This model is particularly valuable for developing economies that lack deep bond markets. For non-marketable assets like loans, Sakurai applies a modified ASRFA model based on the Basel II Asymptotic Single Risk Factor approach, linking stressed probabilities of default to haircut levels. For data-scarce environments, he proposes using simplified structural or cross-country extrapolation models that remain conservative and transparent.

Machine Learning Joins the Monetary Toolbox

A breakthrough feature of the paper is the application of Variational Autoencoders (VAEs), a class of neural networks, to validate haircut models. By training VAEs on historical U.S. Treasury data, Sakurai generates realistic yield curve scenarios that capture nonlinear patterns traditional models often miss. These simulations reveal how level, slope, and curvature factors interact under extreme conditions, allowing central banks to test haircut robustness in synthetic but plausible stress environments. The results show that the DASV model’s haircuts are slightly conservative, providing adequate protection without stifling liquidity. Sakurai suggests that integrating machine learning into central bank modeling can transform risk management into a continuously adaptive, intelligence-driven process.

Toward Transparent and AI-Enabled Central Banking

The study concludes that central banks need consistent, transparent, and empirically grounded haircut frameworks that extend beyond traditional judgment-based practices. Quantitative methods can harmonize treatment across asset classes, protect balance sheets under stress, and strengthen accountability. Sakurai emphasizes ongoing monitoring of uncollateralized exposures as an essential part of counterparty risk control. Looking forward, he envisions an AI-driven “Super Economist”, a system capable of analyzing IMF Article IV reports to extract macroeconomic insights and feed them into regression-based models, particularly for data-poor nations. Such tools, he argues, could enable even smaller central banks to match the sophistication of major institutions.

In essence, Sakurai’s work transforms haircut policy from an administrative rulebook into a scientific discipline that bridges finance, econometrics, and artificial intelligence. By prioritizing data integrity, non-procyclicality, and conservative calibration, the framework sets a new standard for central bank risk management. The paper serves both as a technical manual and a visionary roadmap for a financial world where quantitative reasoning and AI-driven foresight underpin the stability of global monetary systems.