Universal Reverser for the PXS Local Market Gurushyam Hariharan 1. Introduction In this paper I present an intuitive approach to automated trading tested in the PLAT simulation environment. The basis of the approach is to explore an arbitrage opportunity that exists because of the simultaneous presence of two separate, yet interdependent markets viz. 1) the Penn Exchange Simulator (PXS) market and 2) the Electronic Communication Network (ECN) (here Island). My strategy is to capture the difference in the demand and supply in the two markets and explore the gap between the two. The way it explores the gap is by playing against the cumulative mood of the local market. For this reason I call my agent the universal reverser for a local market. The remainder of the report is organized as follows. Section 2 of the paper discusses the motivation and hypothesis of the strategy. The strategy in its most basic form is presented in section 3. The subsequent section discusses some evaluations with experimental analyses and reasoning. This is followed by a note on related and future work. 2. Motivation and Hypothesis PXS aims for providing a test bed for simulating virtual orders and matching them seamlessly with the real world orders. In doing so, PXS creates a local market that is separate from that of the ECN. At every clock tick of the server, it tries to bridge a gap that is developed between the two markets. My hypothesis in the strategy is that the local market of PXS is very small as compared to the bigger market of the ECN. Hence, every local demand or supply that is created in the PXS market will be easily subsumed by the bigger ECN market. In the discussions to follow I will denote supply and demand of a market as the characteristics of the market. The abbreviation UMR (Universal Market Reverser) will be used to denote the agent. To understand the differences in the characteristics that exist in the two markets, I present an overview of the working of that part of the simulator that creates the local market and matches it to the ECN. At every tick, (virtual) orders are taken from the virtual clients. First an attempt is made to match them amongst themselves. If they are successfully matched, they are removed from the order book. If they are not, an attempt is made to match them with the real world order book. If they can be matched, the real world orders are removed from the PXS order book, else the virtual orders are inserted into the same. The PXS buy and sell books are processed in order. All matches between the buy and sell orders that cross are matched. The last price at which the sell and buy virtual orders are matched is the simulator last price for the present tick count. A similar last price also exists for the ECN. It is called the Island last price.The simulator last price (SLP) and the Island last price (ILP) represent the contemporary characteristics of the two interdependent markets. If these were independent markets operating separately, I could buy shares from one of these markets whose last price is the least and sell them in the other market. The profit that I would make is absolute difference: abs(ILP – SLP) per share. Since there is a real time matching of orders in these markets, the above strategy is not possible in this domain. One of the ways the difference in the two last prices could be exploited is if we hypothesize that the PXS market is very small as compared to that of the ECN. The hypothesis would mean that the two prices would tend to converge, and the hypothetical point of convergence would be the ECN price. Hence, difference in the two prices can be seen as an opportunity for an arbitrage. In plain words, if SLP is less than ILP we could see this as an opportunity to buy stocks in a local market at a price lower than the actual market price (read, the island price). The same stands for the opposite scenario. 3. The Basic Agent The basic agent strategy can be represented by the following algorithm While time permits { ILP = getIslandLastPrice() SLP = getSimulatorLastPrice() diff = ILP – SLP If (diff < 0) placeOrder(SELL, price, volume) else if (diff > 0) placeOrder(BUY, price, volume) } Strengths and weaknesses The strategy would work well if the local market (the PXS market) trades shares in volumes less than that in the ECN. In this case, the ECN market can be seen to be a huge market that subsumes the smaller PXS market. Hence the market characteristics of the smaller PXS market would be driven by the ECN market. There are a number of parameters that can be adjusted for a good performance of the strategy. Experiments that deals with different combinations of the same are be discussed in section [4]. The strategy would work well if there is an appreciable difference between the ILP and SLP. For example, if the SLP is $28, and the ILP is $30. If I buy one share at the SLP, I have made a profit of 30-28= $2. On the other hand, if the local market trades very close to the island price, no arbitrage opportunity emerges. Hence the agent doesn’t make profit.In a way, the agent tries to cumulatively play against the local market strategies. If the cumulative mood is to buy, a demand is created, and the agent sells shares. On the other hand if the cumulative mood is to sell, an excessive supply is created and the agent buys shares. In this way, it seamlessly replicates all individual market reversal strategies. E.g. the above discussed scenario is precisely what is achieved by the Reverse strategy discussed in [3]. The only difference being, the latter plays against the expected market mood, whereas, UMR plays against the actual market mood. On the same note, if the agents in the market are ambivalent, they do not create a deterministic demand or supply. In other words, they are unable to define a cumulative market mood effectively. The price difference could fluctuate in the positive and negative values with high frequencies. In such cases, UMR will not get useful information from the difference in prices and hence might not function effectively. 4. Experimental results and performance analysis As mentioned in earlier sections, there are a number of optimizations that can be achieved with proper parameter tuning. Here, the methodology being adopted in tuning the parameters is pure experimentation and reasoning. A number of experiments were conducted to achieve optimizations. The following are results of the some
View Full Document
Unlocking...