Are option returns predictable? In our new working paper “Option Return Predictability with Machine Learning and Big Data” (authors: Turan G. Bali, Heiner Beckmeyer, Mathis Mörke, and Florian Weigert), we study how machine learning techniques can help to predict individual stock option returns.
For this purpose, we draw upon more than 12 million U.S. delta-hedged option return observations in the period from 1996 to 2020 and apply 265 option-based and stock-based characteristics as predictors. As machine learning models, we use penalized regression models (ridge, lasso, and elastic-net), dimensionality reduction regressions (principal component and partial least squares), and non-linear models (gradient-boosted regression trees with and without dropout, random forests, and fully connected feed-forward neural networks). Finally, we also compute equal-weighted ensembles of all linear and all nonlinear models to leverage the informational content of the individual models.
We find that:
- Option returns are more predictable than stock returns with monthly out-of-sample R2s reaching values of more than 2.5%.
- Nonlinear machine learning models (such as gradient-boosted regression trees and neural networks) outperform linear models (such as ridge and lasso regressions).
- Predictability of option returns leads to economically sizeable trading profits of more than 2% per month. The trading strategy remains profitable when we account for conservative estimates of transaction costs.
- Option-based characteristics (such as the implied volatility and moneyness) are more important than stock-based characteristics in the prediction exercise. However, stock-based characteristics deliver additional value when combined with option characteristics.
- Predictability of option returns is highest for stock with low institutional ownership and analyst coverage.
Have a deeper look at the research paper and download it on: