The Learning Model Validation Based On The Human Subject Experiment

The core issue of building artificial stock market in the research of Agent-based Computational Finance is the design of human subject learning mechanism. As the emergence of large number research of learning models, it is gradually important to use more practical and high quality experimental data to validate the parameter of the model.Using subject experiment method to get the shares sequence of domestic investors and validate the forgotten parameter and analogy parameters in Roth-Erve learning model used in Agent-based Model. In a given interval, by comparing the fitness, you can find the parameter values which are more adaptive for domestic investors.This article first embarks from the height of whole research of Agent-based Computational Finance, discussed in detail the origin, the development and the significance of Roth-Erve learning model research in this field, along with the study of learning model validation. Through the literature and understanding of the problem, suggest a method to validate the learning model using human subject experiment which is more adaptive to the learning process and psychology of investors in our country.Because the essence of validating the learning model is survive from the perspective of psychology. Applying the western theory in the investors in China directly, it scientific nature and feasibility is questionable.Based on experimental economics software z- Tree, combining "Learning to Forecast Experiment"(Lt FEs) proposed by Hommoes, we design an human subjects experiment that can reflect our native investors' psychological characteristics better, through the experiment we obtained the relevant data such as stock price series. At the same time, we also design the Roth-Erve learning model calculating financial experiments based on the MATLAB software, simulate the investors' investment decision-making process, also collect data such as stock price sequence. By adjusting the parameter values in learning model, control the fitting effect of learning model, then discusses how to set up a learning model parameters can be better simulated the psychological characteristics and the learning process of investors in our country.Experiment result has shown that subject experiment concludes in convergence and damping vibration of two classes of shares sequence; When price series convergence, it instructs the heterogeneous of investors and adaptive learning process in the experiment is effective, it can achieve a similar state of rational expectations equilibrium; By computational financial experiments, using Roth- Erve learning model to simulate a sequence of two classes of shares, we can find the solution with high fitness, and combining with the practical significance of the four options, we analysis the learning process and the characteristics of shares sequence theory; Finally, in a fixed interval, by comparing the fitness, you can find the parameter values which are more adaptive for domestic investors. For the application of Agent-based Model, the research provides a more adaptive learning model of domestic investors.