The first chapter of my dissertation explores the roles “quants” or technically trained workers play at hedge funds. I quantify the impact of hiring “quants” on hedge fund strategy and risk taking by exploiting H-1B lottery results and a policy shock that significantly reduced the future supply of foreign high-skilled labor in the U.S. I find that the H-1B visa program allows hedge funds to pursue strategies that are more quantitative (e.g., hedging, systematic trading, etc.) as opposed to fundamentals-based (e.g., event driven). Although there is evidence of substitution between fund strategies, I also find that hedge funds overall become more diversified, both in terms of regional focus and fund investment style.
Traditional theories on corporate financial policy mostly focus on firm-specific determinants and assume that capital structure choices are made independently of the actions of peer firms. The second chapter of my dissertation introduces a theoretical model that embodies the dynamic of how peer behavior affects a firm’s optimal choice of debt level. Specifically, peer effect is captured by a measure of how much the liquidation value of assets are adversely impacted if peer firms choose to liquidate simultaneously, the effect of which is more prominent for industries with highly specialized assets – hence low asset redeployability. The theoretical model predicts that a firm’s optional debt level is lower when the liquidation prices of its assets are expected to be more severely impacted by the concurrent liquidation of peer firms. Moreover, in those situations where peer firms’ cash flows are more closely correlated, the optional debt level will be higher. Individual firms’ attempts to maximize their own utility lead to an industry-wide over leverage, further exacerbating the risk of general crises in an already highly correlated industry.
The third chapter of my dissertation explores the potential of using deep learning models to enhance equity trading strategies such as momentum and reversal trading. A deep feed forward neural network model (DFN) is fed with a training data set (1965-2000) that uses the rolling Z-scored cumulative returns of various horizons as predicative variables. With the model parameters generated by the training set, the model is applied to the validation set (2000-2016) to generate a signal for each stock predicting its likelihood of outperforming the cross-sectional medium. Based on this model-predicted signal, I have constructed an long-short investment portfolio out of the validation data set. The deep learning portfolios yield an annualized return of over 20 percent and generate significantly large alphas over commonly used factor models.
