Mining Agent for Automated Stock Trading

v Acknowledgements I would like to express my deep gratitude to Dr. Peter Stone and Dr. Maytal Saar-Tsechansky for their guidance, advice, and encouragement. It has been a great privilege to conduct research under their supervision. I am very grateful for the confidence they had in me. I would also like to thank Dr. Suzanne Barber for her support. Stock market dynamics have drawn the attention of analysts from varied academic disciplines and commercial circles. The advent of online trading and real time facilities in the stock markets has fired a new field of interest in developing automatic trading agents that conduct trades in a relatively autonomous fashion under fixed strategies. A number of trading strategies have been implemented from the perspective of mathematical analysis, market making and artificial intelligence among other techniques. In this thesis, we examine a trading strategy based on analysis of external input in the form of online news. A news-based agent is designed to function within the framework of the Penn Lehman Automated Trading (PLAT) simulator vii [16]. A machine-learning model is built using the reaction of stock markets to news items spread over a period of time. The news-based agent uses this model in real time to predict the price movement of stocks, and place orders accordingly. The performance the agent is evaluated by conducting controlled experiments with three varied kinds of opponent strategies. Two of them base their decisions on statistical analysis of the market and its conditions, and the third one conducts trades in concurrence to suggestions from an online community of day traders and domain experts.