<p>This paper proposes a new rank-adjusted conditional autoregressive Wishart (RA-CAW) model designed to accommodate time-varying ranks in realized covariance measures. While conventional Wishart models typically assume a fixed or full-rank structure, the proposed framework allows the rank to be determined dynamically by the observed data. This flexibility effectively addresses rank-deficiency issues common in realized covariance measures under non-synchronous trading and range-based estimators. Due to the complexity of establishing strict stationarity in a process with time-varying ranks, a Bayesian approach utilizing the Sequential Monte Carlo method is employed for estimation. This methodology facilitates parallel computation and enhances efficiency compared to traditional Markov chain Monte Carlo methods regarding computational time. The empirical performance of the RA-CAW model is evaluated using exchange rate data for the EUR, GBP, and JPY. In-sample results indicate that the diagonal specification is generally preferred, whereas out-of-sample forecasting exercises demonstrate that RA-CAW models consistently outperform conventional multivariate volatility models in portfolio risk minimization. In particular, the rotated RA-CAW specification achieves superior performance within the global minimum variance portfolio framework. These findings underscore the importance of rank dynamics in multivariate financial volatility analysis.</p>