A Framework for Enhancing Quarterly Earnings-Per-Share Predictions Using Bayesian Averaging and Transfer Learning

This paper proposes a novel approach to modelling quarterly Earnings-per-Share (EPS).

We overcome the problem of training on sparse data by formulating the problem as a multi-model regression task, leveraging Machine Learning (ML), Statistical Estimators (SM), and Bayesian Model Averaging for Transfer Learning. The proposed method treats a prediction as a weighted average over ML and SM estimators; the weights reflect evidence of an estimator's performance based on external data sets, thus serving as the transfer learning component in our approach.

We test our methodology on 100 randomly selected US publicly traded companies. Results indicate better performance across all measures compared to using single estimators.

Based on joint work with Michael Kampouridis and Tasos Papastylianou.