Portfolio optimization focuses on risk and return prediction, yet implementation costs critically matter. Predicting trading costs is challenging because costs depend on trade size and trader identity, thus impeding a generic solution. We focus on a component of trading costs that applies universally – trading volume. Individual stock trading volume is highly predictable, especially with machine learning. We model the economic benefits of predicting volume through a portfolio framework that trades off tracking error versus net-of-cost performance – translating volume prediction into net-of-cost alpha. The economic benefits of predicting individual stock volume are as large as those from stock return predictability.
Authors
Ruslan Goyenko, Bryan T. Kelly, Tobias J. Moskowitz, Yinan Su, Chao Zhang
Related Organizations
- Acknowledgements & Disclosure
- We are grateful to the Columbia & RFS AI in Finance Conference participants and discussant Dmitriy Muravyev; seminar participants at Cornell, Syracuse, CityU Hong Kong, and George Mason; as well as Martin Lettau, Lu Lu, Andrew Patton, and Annette Vissing-Jorgensen for valuable comments and suggestions. We thank Zhongji Wei and Andy Yang for their excellent research assistance. AQR Capital Management is a global investment management firm, which may or may not apply similar investment techniques or methods of analysis as described herein. Moskowitz is a member of the NBER, has an academic consulting relationship with AQR Capital, and sits on the board of Commonfund. The views expressed here are those of the authors and not necessarily those of AQR or the National Bureau of Economic Research. Send correspondence to Bryan Kelly, bryan.kelly@yale.edu.
- DOI
- https://doi.org/10.3386/w33037
- Pages
- 53
- Published in
- United States of America
Table of Contents
- Introduction 3
- Preliminaries: Motivation and Data 9
- Motivation 9
- Data 10
- Prediction objects: daily stock trading volume 11
- Predictors 12
- Volume prediction from a statistical perspective 14
- Prediction methods 14
- Prediction results 15
- Volume alpha: the economic value of volume forecasting 18
- Tracking error optimization and its portfolio microfoundation 18
- Normalized tracking error and trading rate (z) 20
- The optimal policy ignoring forecast error (function s) 21
- Machine learning for the economic value of volume prediction 22
- The statistical and economic tasks of volume prediction 23
- Transfer learning via pre-training and fine-tuning 26
- Economic and statistical prediction results 28
- Investment performance in trading experiments 31
- Trading experiment design 31
- Implementing a simulated quantitative strategy 32
- Implementing factor zoo portfolios 36
- Conclusion 38
- Technical details 43
- Neural network implementation details 43
- Additional theoretical analysis 45
- Microfoundation of the tracking error part in the portfolio objective 45
- The economic task as predicting z"0365z 46
- Economic loss functions are not in Bregman class 47
- Further analysis on the loss functions 49
- Additional empirical results 51
- Prediction results in firm size groups and “mixture of experts” forecasts 51
- Additional results of implementing factor zoo portfolios 52
