Man versus Machine Learning Revisited

• 成果类型: Article; Early Access
• 署名作者: Zhang, Yingguang; Zhu, Yandi; Linnainmaa, Juhani T.
• 署名单位: Peking University; Central University of Finance & Economics; Dartmouth College; National Bureau of Economic Research
• 刊物名称: REVIEW OF FINANCIAL STUDIES
• ISSN/ISSBN: 0893-9454
• DOI: 10.1093/rfs/hhaf066
• 发表日期: 2025
• 关键词: cross-section; heteroskedasticity; expectations; forecasts; returns

摘要

Binsbergen, Han, and Lopez-Lira (2023) predict analysts' forecast errors using a random forest model. A strategy that trades against this model's predictions earns a monthly alpha of 1.54% ($ t $-value = 5.84). This estimate represents a large improvement over studies using classical statistical methods. We attribute the difference to a look-ahead bias. Removing the bias erases the alpha. Linear models yield as accurate forecasts and superior trading profits. Neither alternative machine learning models nor combinations thereof resurrect the predictability. We discuss the state of research into the term structure of analysts' forecasts and its causal relationship with returns.