Modeling Reserve-Currency Attrition: A Bayesian Neural Framework for U.S. Dollar Reserve Holdings

Abstract

This study develops a hybrid Bayesian modeling framework to estimate the probability of large-scale foreign divestment from U.S. Treasury securities and to assess the stability of the U.S. dollar’s reserve-currency status. The model integrates a Mixture Density Network (MDN) with Hamiltonian Monte Carlo (HMC) sampling to approximate the conditional distribution of reserve holdings given observed macroeconomic fundamentals and an inferred latent variable representing geopolitical alignment. Principal component analysis and hierarchical regularization are applied to ensure parsimony and mitigate overfitting in a small-sample environment. The empirical hazard rate is defined as the annual probability of a 10 percent contraction in foreign U.S. Treasury holdings—a threshold consistent with historical reserve reallocations during the sterling’s decline and central-bank portfolio adjustment behavior. Results indicate an estimated 6–12 percent annual probability at a 95 percent confidence interval, implying that a 10 percent drawdown could occur roughly once every 8–16 years under current macro-financial conditions. While Gaussian mixture components likely understate extreme tail risks, the findings highlight the resilience of U.S. reserve status and suggest that any future transition would be gradual, multi-decade, and contingent on structural geopolitical realignments rather than cyclical shocks.