Machine Learning Predictions of International Stock Returns

This dissertation is broadly describing predictability of returns on individual stocks in international context. The first chapter covers required prerequisites for any study of fundamental anomalies outside the US. The second chapter studies the predictability of stock returns at an annual frequency. The last chapter then looks at possible profitability of the predictability on liquid universe of stocks at monthly frequency. In the first chapter, we study the role of the choice of a fundamental database on the portfolio returns of a set of 74 fundamental anomalies. We benchmark Compustat by comparing it to Datastream in the US and find systematic differences in the raw financial statements across the databases. These differences only have a small effect on the returns of anomalies when they are constructed on stock-months existing in both databases. Different stock coverage across the databases, however, leads to large statistically and economically significant disparities in the returns. Profitability anomalies yield negative returns on the Datastream universe. In the second chapter, we study statistical significance of 93 fundamental anomalies published in academic journals in a multiple hypothesis setting. We generate a universe of 48,387 data-mined fundamental strategies in order to overcome a problem of not being able to observe strategies that were tried but not published. The multiple hypothesis tests reveal that the number of significant anomalies heavily depends on the precise specification of the tests. We show that the adjustment of standard errors on portfolio returns for heteroskedasticity and autocorrelation is of first order importance and t-statistics on the portfolio returns may not have critical values of the normal distribution. In the third chapter, we study out-of-sample returns on 153 anomalies in equities documented in academic literature. We show that machine learning techniques that aggregates all the anomalies into one mispricing signal are 4 times more profitable than a strategy based on individual anomalies and survive on a liquid universe of stocks. The machine learning also leads to 2 times larger Sharpe ratios with respect to the corresponding standard finance methods. We next study the value of international evidence for selection of quantitative strategies that outperform out-of-sample. Past performance of quantitative strategies in the regions other than the US does not help to pick out-of-sample winning strategies in the US. Past evidence from the US, however, captures most of the predictability within the other regions. The value of international evidence in empirical asset pricing is therefore very limited.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2019

Author dc:creator

• Tobek, Ondrej

Advisor dc:contributor.advisor

• Linton, Oliver

• Machine Learning
• Empirical Asset Pricing
• International Stock Returns
• Anomalies
• Predictability of Stock Returns

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/2a0cce4d-9d02-449c-9cfc-1b8c70cc659e/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

en

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.44015

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/296974