Three Automated Stock-Trading Agents

Three Automated Stock-Trading Agents: A Comparative Study Alexander A. Sherstov and Peter Stone The University of Texas at Austin Department of Computer Sciences Austin, TX 78712 USA {sherstov, pstone}@cs.utexas.edu Abstract. This paper documents the development of three autonomous stock- trading agents within the framework of the Penn Exchange Simulator (PXS), a novel stock-trading simulator that takes advantage of electronic crossing net- works to realistically mix agent bids with bids from the real stock market [1]. The three approaches presented take inspiration from reinforcement learning, myopic trading using regression-based price prediction, and market making. These ap- proaches are fully implemented and tested with results reported here, including individual evaluations using a fixed opponent strategy and a comparative analysis of the strategies

in a joint simulation. The market-making strategy described in this paper was the winner in the fall 2003 PLAT live competition and the runner- up in the spring 2004 live competition, exhibiting consistent profitability. The strategy’s performance in the live competitions is presented and analyzed. 1 Introduction Automated stock trading is a burgeoning research area with important practical applica- tions. The advent of the Internet has radically transformed the nature of stock trading in most stock exchanges. Traders can now readily purchase and sell stock from a remote site using Internet-based order submission protocols. Additionally, traders can moni- tor the contents of buy and sell order books in real time using a Web-based interface. The electronic nature of the transactions and the availability of up-to-date order-book data make autonomous stock-trading applications a promising alternative to immediate human involvement. The work reported here was conducted in the Penn Exchange Simulator (PXS), a novel stock-trading simulator that takes advantage of electronic crossing networks to realistically mix agent bids with bids from the real stock market [1]. In preparation for an open live competition, we developed three parameterizable trading agents and defined several instantiations of each strategy. We optimized each agent independently, and then conducted detailed controlled experiments to select the strongest of the three for entry in the live competition. It is important to realize from the outset that this research is primarily an agent study pertaining to the interactions of particular agents in a fixed economy. Although PXS makes a strong and reasonable claim to implementing a realistic simulation of the stock market, the results and conclusions in this paper pertain to test economies includ- ing specific other stock-trading agents. In particular, we do not aim to create strategies that are ready for profitable deployment in the real stock market (otherwise we would likely not be writing this paper!). Rather, this paper makes three main contributions. First, it contributes an empirical methodology for studying and comparing stock-trading agents—individually as well as jointly in a shared economy—in a controlled empiri- cal setting. Second, it implements this methodology to compare three specific trading agents based on reinforcement learning, myopic trading using regression-based price prediction, and market making. Third, this paper contributes detailed specifications of promising strategy designs, one of which vastly outperformed competitor strategies in an open stock-trading competition and exhibited consistent profitability under a variety of market conditions. The remainder of the...

Website: www.cs.utexas.edu | Filesize: 166kb
No of Page(s): 14
Download Three Automated Stock-Trading Agents - Department of Computer ....pdf