The Battle of the Models: Modern Takes on Traditional and Machine Learning Techniques in Empirical Finance
- Publication Type:
- Thesis
- Issue Date:
- 2023
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Consensus views in finance must be continuously challenged and re-evaluated. This thesis uses new techniques and modern perspectives to challenge commonly held beliefs, both new and old, in financial markets. Across three chapters, this thesis addresses the presence of asynchronicity in financial markets, the purported death of the Standard and Poor's (S&P) index effect, and the presence of biases and overfitting in machine learning models used in asset pricing. First, this thesis establishes dynamic time warping (DTW) as a measure of dynamic asynchronicity and applies DTW to improve asset pricing and price discovery models. Recent research on the S&P index effect, a phenomenon where stocks added to or deleted from the S&P 500 index experience abnormal price responses, argues that it has disappeared. Second, this thesis finds that the S&P index effect has not disappeared. Stocks added into the S&P 500 from outside the broader S&P 1500 universe still experience positive abnormal price responses. However, stocks that move between the S&P 500, S&P 400, and S&P 600 no longer exhibit abnormal price responses to index change announcements. The results connect stock price reactions to announcements of changes to the three main S&P U.S. domestic equity indexes with the impact of relative passive ownership and informed trading on the informational content of these events. Finally, this thesis explores the performance of machine learning models trained on size-specific groups of stocks. Contrary to expectations, grouping stocks by market capitalization improves the performance of machine learning return predictions compared with models trained on the full cross-section of stocks. The superior performance of size-specific models is attributable to a lack of regularization of the target stock returns in the standard machine learning return prediction framework. The results underscore the importance of data selection and prediction target design when training machine learning models for return prediction and serve as a cautionary reminder that machine learning requires careful guidance to reduce biases and overfitting. The findings in this thesis underscore the necessity for inventive approaches and reassessing long-standing and newly emerging beliefs.
Please use this identifier to cite or link to this item: