Documente Academic
Documente Profesional
Documente Cultură
Yanhua Shao1,2, Jianshi Li1 Feng Tian1, Tianjian Chen1, Jinrong wang1
1. College of Computer Science & technology 2. College of Economics & Management
Guizhou University Guizhou University for Nationalities
Guiyang, Guizhou Province, China Guiyang, Guizhou Province, China
online-8j@163.com sh_yh2002@163.com
I. INTRODUCTION
a1
a w
s =0
s ks
0
an wkn
Because of their ability to deal with uncertain, fuzzy, or
insufficient data which fluctuate rapidly in very short periods
of time, neural networks have become very important method Fig. 1 a node or artificial neural
for predictions [Ref. 1]. These include business forecasting, The processing element (PE) can be considered as an over
credit scoring, bond rating, business failure prediction, simplified sketch of a natural neuron, where the parameters wkj
medicine, pattern recognition and image processing. A large represent the strength of the synapses. The aj values
number of studies have been reported in literature with (activations) are the inputs of the PE or neuron, which
reference to the use of ANN in modeling stock prices. calculates their weighted sum (this operation explains the use
However, not much work along the approach to neural network of the conventional name weight to designate the parameters
based on agent has been reported. of a function in the ANN jargon). The same operation, in
Stock market is a special example of complex economics matrix algebra, is done by the product Ax, where the x vector
system [Ref. 2, 3, 4]. Conventional stock market model bases contains the aj activations and the matrix A the wkj parameters.
on the assumption that the trader takes action rationally in The PE produces an output applying a non linear
accordance with market. Stock market seems to be a dead transformation to the weighted sum, as that shown in equation
physical system on the premise of the model assumption. (1), where k is the parameter that determines the closeness of
However, the assumption is far away from the realistic stock the function to a crude step function.
market, people often characterize stock market by the language
that express human psychology, such as stock market durative 1
f (z) =
downturn, crazy going up. Furthermore, phenomena like 1 + e ikz (1)
bubbles and crashes can emerge as collective phenomena. We
will put emphasis on how to supersede agent with neural Each row of A and B contains the weights of a PE. To find
network and how to use a neural network approach to the self- the A and B values (i.e. the ANN parameters) we have to use a
development of consistency in agents' behavior in this paper. numerical approximation with a simplified gradient method,
the socalled back propagation method [Ref. 5, 6] (which
comes from propagating back the error measures from the
output layer to the hidden one), starting with random
parameters and adapting the performances of the ANN function
Random
Environment Environment
Agent Action Effects (Actual) Targets
Requested Action to Match
Random
Imitating Rule-Master
Action Effects (Guess) Action (Guess)
Agent
E1 E2 Em A1 A2 An Output
Fig. 2 model of stock market
The main value of the ERA scheme is that it keeps both the Hidden
environment, which models the context by means of rules and
general data, and the agents, with their private data, at different
conceptual levels. Another advantage of using the ERA
structure is its modularity. I1 I2 Ik Input
REFERENCES
[1] Schoeneburg, E., Stock Price Prediction Using Neural Networks: A
Project Report, Neuro-computing, vol. 2, 1990, pp. 17-27.
[2] John H. Holland. Hidden Order: How Adaptation Builds Complexity.
Addison Wesley Publishing Company, 1995.
[3] John H. Holland. Emergence: From Chaos to Order. Addison-Wesley
Publishing Company, 1998.
[4] Arthur W B, Holland J, Lebaron B, et al. Asset Pricing under
Endogenous Expectations in an Artificial Stock Market [C]//Arthur W
B, Durlauf S, Lane D. The Economy as an Evolving Complex System II.
Boston: Addison-Wesley, 1997, pp: 15-44.
[5] Anderson, J.A., An Introduction to Neural Networks, Cambridge, MA:
Fig. 5 the time sequence of the orders The MIT Press, 1995.
With the interaction of heterogeneous agents and structured
[6] Zekic, M., Neural Network Applications in Stock Market PredictionsA
environments or markets, we use the cross targets technique to Methodology analysis, University of Josip Juraj Strossmanyer in Osijek,
train artificial neural networks to develop minimal behavioral 1998.
rules. Moreover, generated sequences of prices linked to the [7] Pietro Terna (2000a). Self-Development of Consistency in Agents'
agents actions in nonlinear ways. Therefore, the experiment Behavior, in F. Luna and B. Stefansson (eds.), Economic Simulations in
satisfies to understand the dynamic specialty of stock market Swarm: Agent-Based Modeling and Object Oriented Programming.
fully. Dordrecht and London, Kluwer Academic, 2000.
[8] Gilbert, N., Terna, P., How to build and use agent-based models in
social science, Mind & Society, no. 1, 2000..