Automated PLAT Trading Agent Using Order Imbalance in Volume Harish

Volume of trades and order book volume imbalances have long been established as important criteria in evaluating portfolios and long term investment strategies. The hypothesis being evaluated in this project is that it is an essential component of intraday trading strategies – important enough to be effectively used exclusively as an indicator of the market behavior. The design of a trading strategy based on order imbalance in volume with a genetic algorithm used to tune order volume as a parameter is discussed. The performance of this agent in various environments and some preliminary data from a live competition with other agents is studied. Introduction and Early Agent Design A multitude of day trading strategies from ‘resistance and support’ to the ‘market making with volume control’ [3] strategy discuss volume as a parameter, in the former case, to aid the decision process and in the latter, as a control mechanism. Most studies however, have considered this a secondary factor and hence, literature on studies of its exclusive effect on intraday trading is scarce. Volume effects have long been studied as a factor in long term investment decisions and portfolio management [4], [5]. In this project, the initial hypothesis was that an agent might be expected to decide trading activity based solely on volume. However, this hypothesis was quickly revised, since any decision to trade stocks brings with it two questions – when to trade and how much to trade. Of course, if the ‘when’ is answered by a volume measure (order imbalance) [6], then the ‘how much’ needs to be a measure of price – and this led to the conclusion that this decision is dependent on price and hence price can be adapted as a factor in tuning the order volume – thus answering the ‘how much’ question. For this, we use a simple genetic algorithm. This pushes one of the parameters to be decided into an independent search space and hence the monitoring of these two parameters can be performed simultaneously. Early work in parameter tuning for trading agents paved the way for this development [7], [8]. Before it was decided that the decision was to be made on volume alone, a number of vistas were explored, most prominently, the effect of ‘human’ factors such as loss aversion [9] (a concept stemming from Behavioral Economics). Specifically, it was hypothesized that if loss aversion could be modeled into the utility function of an agent, then the agent would decide, at each point, what his utility would be from the trade, and decide based on this. Of course, a loss aversion term would mean this decision would be one of minimizing losses rather than increasing profit, a concept consistent with loss aversion, but redundant in the day trading scenario. A major hurdle was modeling the utility function adequately. And if that was a big problem, a more fundamental flaw in this reasoning came to the fore. While the behavioral effects are pronounced in investment strategies and portfolio balancing (where endowment effect plays a huge