Thesis or dissertation
Date of this Version
The main goal of this study is to determine whether machine learning can outperform analysts in forecasting earnings. Using gradient boosted regression trees (a recursive regression tree-building method), this paper concludes that machine learning is unable to beat analysts’ predictions for earnings, when comparing median absolute percentage error. The model was trained on firms with Wall Street analyst coverage for earnings between years 2013 to 2016. Predictors from existing earnings forecasting literature were input for the model’s consideration. The model’s performance was compared to analysts’ forecasts on out-of-sample earnings for years 2017 to 2019. The results suggest that analysts hold some incremental information that is useful for forecasting earnings. This incremental information is either not contained in financial statements or has not been researched in existing literature.
machine learning, accounting, analysts, predictions, earnings, gradient boosted regression trees, GBRTs, forecast, performance
Date Posted: 19 May 2020