Advanced statistical modeling for bank asset analysis. Researchers developed a centered-innovation approach to Bayesian Dirichlet ARMA models for compositional time series, with direct application to Federal Reserve H.8 bank-asset share data spanning October 2015 through October 2025. The method replaces raw residuals with centered innovations using digamma identities, proving theoretical equivalence while delivering significant practical advantages. Testing across 104 rolling weekly origins shows predictive performance remains statistically indistinguishable, but computational stability improves dramatically. Hamiltonian Monte Carlo divergent transitions decreased by approximately an order of magnitude under the centered specification, eliminating catastrophic divergence spikes that previously occurred at isolated rolling origins.
Post from MarketNews_en
Log in to interact with content.