Yavor Kovachev
ESSAYS IN MACHINE LEARNING APPLICATIONS FOR ASSET PRICING
Yavor Kovachev ESSAYS IN MACHINE LEARNING APPLICATIONS FOR ASSET PRICING
ISBN 978-91-7731-202-4
DOCTORAL DISSERTATION IN FINANCE
STOCKHOLM SCHOOL OF ECONOMICS, SWEDEN 2021
Yavor Kovachev
ESSAYS IN MACHINE LEARNING APPLICATIONS FOR ASSET PRICING
Yavor Kovachev ESSAYS IN MACHINE LEARNING APPLICATIONS FOR ASSET PRICING
ISBN 978-91-7731-202-4
DOCTORAL DISSERTATION IN FINANCE
STOCKHOLM SCHOOL OF ECONOMICS, SWEDEN 2021
Essays in Machine Learning Applications for Asset Pricing
Yavor Kovachev
Akademisk avhandling
som för avläggande av ekonomie doktorsexamen vid Handelshögskolan i Stockholm framläggs för offentlig granskning onsdagen den 2 juni 2021, kl 10.00,
Swedish House of Finance, Drottninggatan 98, Stockholm
Essays in Machine Learning Applications for Asset Pricing
Yavor Kovachev
Dissertation for the Degree of Doctor of Philosophy, Ph.D., in Finance
Stockholm School of Economics, 2021
Essays in Machine Learning Applications for Asset Pricing
© SSE and Yavor Kovachev, 2021 ISBN 978-91-7731-202-4 (printed) ISBN 978-91-7731-203-1 (pdf)
This book was typeset by the author using LATEX.
Front cover illustration: © Konstantin Faraktinov/Shutterstock.com Printed by: BrandFactory, Gothenburg, 2021
Keywords: Machine learning, deep neural networks, empirical asset pricing
To my family
Foreword
This volume is the result of a research project carried out at the Department of Economics at the Stockholm School of Economics (SSE).
This volume is submitted as a doctoral thesis at SSE. In keeping with the policies of SSE, the author has been entirely free to conduct and present his research in the manner of his choosing as an expression of his own ideas.
SSE is grateful for the financial support provided by the Jan Wallander and Tom Hedelius Foundation which has made it possible to carry out the project.
Göran Lindqvist Per Strömberg
Director of Research Professor and Head of the Stockholm School of Economics Department of Finance
Stockholm School of Economics
Acknowledgements
I thank the entire faculty and staff at the Swedish House of Finance. I am deeply grateful and indebted to my advisor Michael Halling for his advice and support during my PhD studies. Special thanks goes out to Irina Zviadadze who went above and beyond to help me navigate the job market process and Jungsuk Han for sharing research ideas and his willingness to collaborate with me. I am thankful to all my fellow PhD colleagues:
Alberto, Lazslo and Ivika who started this journey with me on day one, Xingyu for stimulating discussions on research, life and academia, Erik for sharing an interest and passion for all things related to machine learning and artificial intelligence. Further thanks to Valentin, Hendro, Yue, Andreas, Katarina and Mahamadi. All staff at the Swedish House of Finance were exceptional at handling and assisting with administrative tasks and helping me understand and navigate the regulatory aspects of the program. I offer sincere gratitude to Anneli, Jenny, Hedvig, Elisabeth and Anki. Finally, I gratefully acknowledge the financial support of The Swedish Bank Research Foundation.
Uppsala, April 30, 2021 Yavor Kovachev
Contents
Contents ix
Introduction 1
1 Predicting stock price movements with news implied information sen-
timent 5
1.1 Introduction . . . 6
1.2 Deep Learning Heuristics . . . 10
1.3 Model . . . 15
1.4 Data . . . 19
1.5 Results . . . 22
1.6 Future work . . . 42
1.7 Conclusion . . . 43
1.8 Appendix . . . 45
1.9 References . . . 79
2 Calibration of stochastic volatility option pricing models 83 2.1 Introduction . . . 84
2.2 Models . . . 87
2.3 Methodology . . . 93
2.4 Data . . . 99
2.5 Results . . . 100
2.6 Summary and future work . . . 107
2.7 Appendix . . . 109
2.8 References . . . 121
3 On the numerical solution of a reformulated version of the Kyle-85 model 123 3.1 Introduction . . . 124
3.2 Model overview . . . 124
3.3 Numerical solution process . . . 127
3.4 Conclusion and future work . . . 133
x
ESSAYS IN MACHINE LEARNING APPLICATIONS FOR ASSET PRICING
3.5 Appendix . . . 135 3.6 References . . . 145
Introduction
The result of my research efforts as a Finance PhD student at the Stockholm School of Economics this doctoral thesis is a collection of three independent essays. The unifying theme is the application and development of machine learning techniques and method- ologies to address (and potentially solve) research questions in the field of empirical asset pricing.
The first paperPredicting stock price movements with news implied information senti- ment: a machine learning approach contributes to the field of financial textual analysis by demonstrating how advanced deep learning techniques can be used to gain insights on the interactions between information and stock prices at higher frequencies than previously achieved in the asset pricing literature. To this end, I design and train a recurrent deep neu- ral network (RDNN) model which predicts sentiment i.e. positive or negative opinions expressed in text. Using a sample of over 240,000 news articles covering Google, Apple, IBM and Microsoft and published between 2005 to 2018 I show that RDNN sentiment predictions correlate with contemporaneous stock price changes and the correlation is stronger when price changes are high. Furthermore, I demonstrate that sentiment ob- tained with the RDNN can be used to make statistically significant predictions of the direction of future stock price changes at short time horizons of 8 to 24 hours. To test the economic significance of these predictions I construct sentiment-based momentum strategies and show they have positive cumulative returns with Sharpe ratios ranging from 1 to slightly higher than 2. I show that extremely positive news implied sentiment persists over longer time horizons - 16 to 24 hours, while extremely negative sentiment is priced in much quicker - within 8 hours. Econometric tests indicate there is no co-integration between cumulative returns series of sentiment-based momentum strategies generated from Wall Street Journal (WSJ) and Dow Jones New Services (DJNS) articles and cu- mulative return series from non WSJ&DJNS articles. Additionally, non WSJ&DJNS cumulative return series are much closer to random walks and are, in almost all cases, non-stationary suggesting that the RDNN is potentially picking up on noise trading defined by Black (1986) as individual investor’s irrational behavior which is uncorrelated with any fundamental factors.
The second paperCalibration of stochastic volatility option pricing models: risk vs
2
ESSAYS IN MACHINE LEARNING APPLICATIONS FOR ASSET PRICING
performance focuses on model risk in derivative valuation. Due to their complexity, calibrating stochastic volatility models to market data involves highly non-linear, non- convex search spaces. Using particle swarm optimization (PSO) (Kennedy and Eberhart, 1995), a non-gradient based global optimization technique developed in the machine learning literature, I study the calibration performance of seven well known stochastic volatility models: Heston (1993), Bates (1996), Barndorff-Neilsen and Shephard (2001) and four stochastic time change models including the Normal Inverse Gaussian – Cox- Ingersoll-Ross of Geman et al. (2003) over three years of option chain data on the S&P 500. I find that the parsimonious Heston model and the significantly more advanced stochastic time changed Normal Inverse Gaussian – Gamma Ornstein–Uhlenbeck model display the best overall calibration performance.
The third chapterOn the numerical solution of a reformulated version of the Kyle- 85 model outlines a numerical technique for solving a modified version of the Kyle-85 model developed by Han (2013). In this reformulated version the informed trader has additional sources of information beyond the fundamental value of one risky asset as in the standard setting. Additionally, the model features a feedback mechanism through which the informed trader takes into account his or her impact on equilibrium prices which leads to significant technical complications in the solution process. The main finding is that a hybrid optimization technique combining particle swarm optimization (PSO) (Kennedy and Eberhart, 1995) and standard gradient-based techniques overcomes these technical difficulties.
INTRODUCTION 3
References
Barndorff-Neilsen, O. E. and N. Shephard (2001). “Non-Gaussian Ornstein-Uhlenbeck- based models and some of their uses in financial economics.” In:Journal of the Royal Statistical Society, Series B 63, pp. 167–241.
Bates, D. S. (1996). “Jumps and stochastic volatility: exchange rate processes implicit in deutsche mark options.” In:Review of Financial Studies 9.1, pp. 69–107.
Geman, H., P. Carr, D. B. Madan, and M. Yor (2003).Stochastic Volatility for Levy Processes.
Economics Papers from University Paris Dauphine 123456789/1392. Paris Dauphine University.
Han, J. (2013). “Dual trading and liquidity.” In:Working paper.
Heston, S. L. (1993). “A closed-form solution for options with stochastic volatility with applications to bond and currency options.” In:Review of Financial Studies 6, pp. 327–
343.
Kennedy, J. and R. Eberhart (1995). “Particle swarm optimization.” In:Neural Networks, 1995. Proceedings., IEEE International Conference on. Vol. 4, 1942–1948 vol.4.
Kyle, A. S. (1985). “Continuous auctions and insider trading.” In:Econometrica: Journal of the Econometric Society, pp. 1315–1335.
