The Learning of Machine Learning: The Effect of Human Capital on the Performance of Algorithms

Abstract

Strategy formulation hinges on managers’ ability to formulate accurate predictions about developments in external environments. As managers face various cognitive limitations that bias their predictions, strategy scholars suggested artificial intelligence, in the form of machine learning, as a solution to improve firms’ predictive ability. Notwithstanding these insights, we still know little about machines’ predictive ability in uncertain external environments, where information about future trends is limited. We maintain that machines’ prediction accuracy can be improved by leveraging specific dimensions of human capital as input in their training sets. To test our theory, we created a unique machine learning algorithm that forecasts the earnings of publicly listed firms by leveraging financial data and financial analysts’ past earnings predictions. Contrary to prevailing views, we demonstrate that machine learning algorithms that are trained on past predictions by high human capital analysts – who are likely to be biased in various dimensions – yield more precise predictions than similar machines that are trained on predictions by low human capital – less biased – analysts, as well as those that are trained only on financial data. Our theory and results provide fresh insights on how firms could leverage their human capital to improve machine learning’s forecast accuracy