یک مدل موازی برای بازار ارز

A numerical model for the foreign exchange (FX) market is developed and its implementation on a distributed memory parallel computer is discussed. The model considers a description of the market at the level of the real agents, such as traders and market makers. These actors are represented by interacting computerized agents. Parallelism allows the study of systems with many actors and realistic trading rules. In order to analyse the generic dynamical properties of the market, simulations are considered. The results agree with several observed features of the real market, such as non-Gaussian distribution and negative short-term autocorrelation of price changes.

The need for new methods to investigate and better understand the behaviour of economical and financial markets has brought in the past few years a lot of attention to simulations of economical systems [1], [2] and [3]. Microsimulations, which model on a computer the activity of most actors of a real system, are recognized as a very promising approach. Among the attempts made so far, focus is given on the evolution (neural nets, genetic algorithms) of the market agents rather than on the emergence of a complex and global behaviour resulting from the interaction between many agents [4] and [5]. On the other hand, several toy models have been proposed to capture this global behaviour [6]. Unfortunately, the underlying hypotheses in these models seem too simplistic with respect to the mechanisms in effect in real markets. Most of these models deal with the stock market and, in addition, have been run sequentially. The foreign exchange (FX) market, which is yet the most active financial area in terms of money amounts, has been neglected so far. Global turnover in all traditional foreign exchange market segments, such as spot transactions, outright forwards and foreign exchange swaps, reached an estimated daily average of 1490 billion US dollars in April 1998 [7]. An artificial FX market which aims at capturing the essential “microscopic” rules of its real counterpart is proposed. The interest does not reside in the predictive capabilities of our model, but rather in its capacity to identify the key ingredients and their effect on generic properties of the market. The validation of the model will be considered in terms of its ability to reproduce the statistical properties of the real market – which have been well studied [8], [9], [10] and [11]. For this purpose the parallel software environment includes a windows-based interface (Tcl-Tk) which allows an easy control of the parameters and provides tools to display the market evolution and monitor the behaviour of each agent. The paper is organized as follows. Section 2 presents the ingredients of the financial model under investigation. Simulation results are described in Section 3. The parallel implementation and numerical performance of the simulator are discussed in Section 4.

A model for the FX market has been developed. It is based on the interaction between agents who act according to realistic rules. The simulations we have performed show that our computerized market has many similarities with the real market, although some aspects may still require an adjustment of the parameters. Even though more ingredients of the real actors could be included in the model, the simplifying assumptions are already sufficient to capture several non-trivial collective behaviours observed in real FX markets. The results are promising and we plan more simulations to further validate the model through extensive comparisons with actual features of the FX market. One goal of this model is to analyse the market in terms of its sensitivity to the parameters and/or to study the consequences on the market of some possible scenarios. In the future work the analysis of these aspects will be considered in more detail. Due to the presence of many actors simultaneously active on the market, the model is designed as a parallel system. Its implementation is realized with message passing. Although the communication time increases with the number of market makers M and number of processors p, a parallel execution is beneficial when the number of traders is large N≫M. This is a reasonable assumption in comparison with the real market. In addition, communications do not increase as the traders and market maker strategies become more complicated. Thus, the communication to computation ratio is in favour of a parallelization and allows the investigation of large and realistic situations.