Expectations Matter: When (Not) to Use Machine Learning Earnings Forecasts

We comprehensively examine the usefulness of machine learning technology to predict a firm’s earnings and offer three main findings. First, although prior literature suggests machine learning can offer better earnings forecasts than analysts, we show that this result is highly sensitive to machine learning model specification choices (i.e., 80% of evaluated machine forecasts fail to beat analysts). Second, we examine why the most accurate machine learning forecast consistently beats analysts, finding that they correct for predictable analyst biases that are both linear and nonlinear and largely relate to analysts’ prior forecast errors, forecasted earnings levels, and the firm’s stock price. Finally, we find that investors’ earnings expectations, as revealed through stock prices, largely—but do not fully—correct for these predictable analyst biases, with delayed price realization up to nine months. In additional analysis, we find that optimal machine learning specification choices remain stable over time and that, although the machine’s outperformance narrows in recent periods, it remains substantial among small-cap stocks. Overall, our study moves beyond the question of whether machine forecasts are superior to human forecasts and instead focuses on which machine forecast specifications matter, as well as when and why machine forecasts are most superior. In so doing, we provide code and estimates for the most accurate machine forecast specification and demonstrate that investors’ expectations appear to largely (but not fully) align with them. This paper was accepted by Suraj Srinivasan, accounting. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2024.05808 .