The highly intelligent virtual agents for modeling financial markets G. Yang 1, Y. Chen 2 and J. P. Huang 1 1 Department of Physics, Fudan University. 2 Department of Human and Engineered Environmental Studies, Graduate School of Frontier Sciences, The University of Tokyo. Introduction Agent-based modeling is proven to be a very promising method to deal with the complexity of economic or social systems. The bottom-up approach is a most significant feature, which resembles the methods in statistical physics. The difference is that the micro-units in agent- based models (ABMs) are artificial agents, while those in physics systems are entities with no observing, learning or adapting abilities. Among complex human systems, modeling of financial markets is the most attracting project for the researchers. Three roles of ABMs for modeling financial markets: A microscopy for underlying dynamics: reproduce the stylized facts, and explain the micro-dynamics behind the phenomena that appear in the macro-states. This can give us a better understanding of human behaviors in real markets. An evaluator for system properties: find how the statistics of financial markets evolves under the change of macroscopic or microscopic environments. This can offer suggestions to policy-makers. A forecaster for future states: predict the future market movements. This can be used to make trading strategies. Three Principles A high information processing ability: agents can easily observe, collect and organize multiple information from the environment. A high learning ability: agents can learn from trial and error to optimize the internal states of their own decision-making models. A high adaptation ability: agents can replace their strategies to improve their adaptability, by using self evolutionary algorithms. Table 2. Comparison of agents' mean wealth from index trading of S&P at the end of the optimization and the test periods respectively. Evaluation of the agents iAgents observe three sign series of information flows: the price change flow, the volume change flow and the volatility change flow. Impact factor of a certain sign series on an iAgent is expressed as:. The position-changing level of an iAgent is defined as:, and his/her position is set as:. iAgents’ strategy form:. Dynamic Genetic Algorithm (DGA) is adopted to update iAgents’ strategies. Random traders: change positions randomly. WG agents: a type of agents from wealth game (a modified minority game to study financial markets). WG-DGA agents: WG agents equipped with DGA process. Building Agents WG-DGA agents have slightly more intelligence than WG agents. But both the two types of agents are more like passive fund managers whose main goal is just to follow the market. However, the mean wealth of iAgents on the test period is 8.56 times the market performance. In addition, the mean position of iAgents shows a degree of clustering. This implies that trading strategies may be constructed on the iAgents’ decision-making model. Conclusions We have designed a kind of highly intelligent virtual agents called iAgents based on three principles. Intelligence of iAgents have been tested through virtual index trading on S&P (also on NKY in the paper), along with random traders, WG agents and WG-DGA agents. It has been shown that trading behaviors of WG agents and WG-DGA agents are more like those of passive fund managers. However, iAgents can in average outperform the market greatly, which means that the three principles give iAgents a higher intelligence. G. Yang, Y. Chen, and J. P. Huang, The highly intelligent virtual agents for modeling financial markets, Physica A: Statistical Mechanics and its Applications 443, 98 (2016). iAgents: WG agentsWG-DGA agents iAgents