Machine Learning to Correct Analyst Earnings Forecast Optimism

Summary

This research tests whether machine learning can estimate and reduce the optimism bias in analyst earnings forecasts. It uses 27 variables drawn from research reports, analyst characteristics, company fundamentals, and market information. LASSO provides a linear benchmark, while gradient boosted regression trees model nonlinear relationships; each fiscal year is predicted using the prior year’s data.

In a 2010–2019 historical evaluation, the tree model had lower out-of-sample prediction error than LASSO except in 2011 and 2015. The researchers then adjust analyst forecasts for estimated optimism and combine recent revisions with weights based on predicted reliability. Their adjusted consensus forecasts are more accurate than the comparison data, though optimism remains because the model explains only part of it. Tests on China Securities 800 constituents found little change in a consensus earnings-to-price factor, while the earnings-change factor’s information coefficient rose from 0.02 to 0.03 and remained effective after 2018. The summary provides limited detail on validation design and does not establish performance outside this setting.

Key ideas

• The study predicts analyst forecast optimism using company, analyst, report, and market variables.
• It compares a linear LASSO model with nonlinear gradient boosted regression trees.
• The tree model generally had lower out-of-sample error over the historical evaluation, with two stated exceptions.
• Adjusting forecasts reduced but did not eliminate their optimism bias.
• The adjusted earnings-change factor showed stronger reported information content in the tested China A-share universe.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.