A Bayesian model for portfolio decisions based on debiased and regularized expert predictions

Abstract

Expert predictions of future returns are one source of information for educated stock portfolio decisions. Many models for the mathematical aggregation of expert predictions assume unbiased predictions, but in reality, human predictions tend to include biases, and experts’ competence may vary. We propose a Bayesian aggregation model that includes a regularization process to eliminate the influence of experts who have not yet shown competence. The model also includes a debiasing process that fits a linear model to predicted and realized returns. We applied the proposed model to real experts’ stock return predictions of 177 companies in the S&P500 index in 37 industries. We assumed that the decision-maker allocates capital between the industry index and the most promising stock within the industry with the Kelly criterion. We also conducted a simulation study to learn the model’s performance in different conditions and with larger data. With both the real and simulated data, the proposed model generated higher capital growth than a model that ignores differences between experts. These results indicate the usefulness of regularizing incompetent experts. Compared to an index investor, the capital growth was almost identical with real data but got higher when applied only to industries that were estimated to have multiple competent experts. The simulation study confirmed that more than two competent experts are necessary for the outstanding performance of the presented model.