Abstract

Bayesian Model Averaging (BMA) is a popular statistical device that permits computing model parameters or similar model outcomes by weighting different models according to their evidence and the prior beliefs of the modeler. This, together with the robustness provided by model averaging, has granted BMA considerable popularity as a risk assessment tool. In fact, BMA has been extensively used in application domains as diverse as geosciences, medicine, and finance. However, at the same time, some works have called for attention to the observation that BMA can perform poorly, even worse than very simple methods. It turns out, BMA can be affected by overfitting and by model selection bias (epistemic uncertainty). This work examines the sources of bias that affect BMA and calls to use this framework more carefully. In particular, it suggests to pay more attention to model selection. To illustrate these ideas, this work utilizes the inverse problem that arises from interpreting geophysical data for inferring source properties of tsunamis and earthquakes. It is shown that risk assessments based on BMA can exhibit significant sensitivity to modeler assumptions such as particular parameterization, while, in contrast, properly carried Bayesian model selection can lead to intrinsically parsimonious solutions.