UC Irvine

Abstract: Huang and Merwade (2023), https://doi.org/10.1029/2023wr034947, hereafter conveniently referred to as HM23, wrongly claim improvement of their method for postprocessing multi‐model water stage predictions using Bayesian Model Averaging (BMA). Their results show all signs of a flawed implementation of the Metropolis algorithm. In this comment I will point out the many mistakes and shortcomings of the BMA methodology of HM23. Their method is deficient, inefficient and ineffective and wrongly quantifies BMA model parameter and predictive uncertainty. Furthermore, HM23 misrepresent BMA literature, articulate a poor understanding of Markov chain Monte Carlo methods and misuse the autocorrelation function for monitoring convergence of the sampled Markov chains. A proper implementation of the random walk Metropolis algorithm would have led HM23 to substantially different results and findings about their ensemble of water stage predictions. The MODELAVG toolbox of Vrugt (2018) https://www.researchgate.net/publication/299458373_MODELAVG_A_MATLAB_Toolbox_for_Postprocessing_of_Model_Ensembles satisfies all requirements of HM23 and provides robust estimates of BMA model parameter and prediction uncertainty for symmetric, skewed and truncated conditional forecast distributions of the ensemble members.