All models are wrong. Some might be OK!

Laura Ryan, Baoqing Gan and Geoffrey J. Warren

Abstract

Overlooking model uncertainty can lead to flawed decisions. This paper examines the key sources of model uncertainty – that is, aleatoric uncertainty (inherent randomness), epistemic uncertainty (variable selection and coefficient uncertainty) and functional form uncertainty – and the challenges involved in addressing them and their sources. Using equity–bond correlation models as a case study, we show that even sophisticated techniques such as bootstrapping, multimodel averaging, density-based spatial clustering of applications with noise (DBSCAN) and Gaussian mixture model (GMM) clustering fall short of identifying the true data-generating process. Functional form uncertainty proves to be particularly problematic, often persisting despite large data sets and advanced modeling tools. We show that these challenges are amplified when models are used for prediction rather than explanation because of structural change, overfitting and unstable data relationships. Our findings suggest that, rather than ignoring model uncertainty, financial practitioners should embrace it by adopting flexible, ensemble-based approaches and maintaining transparency about model limitations. This study offers practical insights for navigating uncertainty in financial modeling and portfolio construction.