UC Irvine

Machine learning is increasingly gaining applications in Finance industry. In this dissertation, I use machine learning methods to predict mutual fund and hedge fund performances and address the issue whether mutual fund and hedge fund managers add value. Overall, machine learning methods tend to outperform OLS in terms of return prediction. From a machine learning point of view, mutual fund managers don’t add value while hedge funds do deliver risk-adjusted performance. Also, such outperformance of top hedge funds is persistent with three-year horizon. Furthermore, such outperformance provided by machine learning methods are not driven by fund characteristics. A regression of machine learning outperformance on macroeconomic variables show that machine learning models tend to perform better when the economy is in recession, when the market is bearish, and when the market and economic policy uncertainty are increased.

Recently, reinforcement learning (RL) is being explored by practitioners in trading. In particular, algorithmic trading provides an ideal setting for RL. In this dissertation, I trained an agent to place aggressive stock market orders with deep Q-network with multi-

step temporal difference. The back-test results show that, the aggressive order agent outperforms an agent with linear execution schedule by an average of 0.12 to 0.69 basis points on simulated orders in Asia Pacific stock markets. This shows that RL-based methods are capable of recognizing and utilizing market information and are promising to outperform traditional execution algorithms.

In summary, machine learning and reinforcement learning are promising to provide better performances in portfolio investment and trading.