Analyst Versus Machine Learning Predictions of Risk

The use of artificial intelligence has rapidly evolved over the last decade to become a common component of financial analysis and we provide an early evaluation of its effectiveness. We confirm analysts’ ability to predict investment risk, defined as future share price volatility, and contrast their effectiveness with predictions based on “machine learning”. Analysts marginally dominate the machine learning approach for the relatively homogenous US sample, but marginally under-perform for a diverse international sample. Both analysts and machine learning approaches appear inefficient as they ignore relevant available information; each assessment would also be improved by incorporating information from the other. Although our analysts are not employed by a sell-side institution, and thereby avoid some of the most obvious incentives to reveal bias, we find evidence of analysts’ under-estimating risk for stocks classified as “buy”. We found no evidence of equivalent bias in the machine learning predictions. Our results suggest that machine learning forecasts of risk are valuable, most obviously in the absence of analysts’ forecasts, or when analysts will find the information environment complex, and are likely to become increasingly influential. As yet, head-to-head, analysts’ predictions of risk still compete with those based on machine learning.