Last update: March 10th, 1998
| 11. | Name: | Sommer Joerg | |
| Organiz.: | University of Augsburg | ||
| Co-Authors: | Joerg Sommer, Horst Hanusch | ||
| JELC: | D84; G11; C45 | ||
| Keys: | Expectations, Speculations; Portfolio Choice; Neural Networks | ||
| Title: | A Neural Net Model of Portfolio Choice | ||
| Artificial neural nets are considered to be suitable describing complex social systems. Due to the natural model, the human brain, the main point of this technique is the ability to learn. For this reason the increasing relevance in economics becomes obvious, whereby two main applications can be distinguished: A statistical/econometric one and a theoretic/experimental one. Former mainly concern with time series forecasting of speculative markets. However, in comparison with conventional methods the superiority is controversial discussed. Therefore, it is recommended to proof the quality of such applications in the setting of theoretical models. In this context, the basic elements are seen in the description of the decision process and the overlapping of individual decisions. The focus is set on the examination of the circumstances and the information which are responsible for the supply and demand decisions. Exogenous macro information is dissembled in individual decisions or, in other words, in micro information. The other way round, this decisions on the microeconomic level determine the macroeconomic development, the market price respectively. This perspective enables the connection between the micro- and macroeconomic level. This paper integrates this point of view in a traditional model of the capital market, namely Arrow’s State Preference Model (SPM). In every period the agents make a decision about their portfolio of security holdings. For simplicity, there are only two securities to select: the risky security M and the risk-free security O. Without any loss of generality, the price of the risk-free security doesn’t change while the price of the risky security depends on the market activities. In contrast to the SPM the agents form expectations about the next periods price of security M. Therefore, the expectation building process consists of two components. The first component reflects the more rational part of the investors behavior. The information processing of alternative scenarios of the economic development leads to anticipated bottoming and topping prices. The difference between the current price and the individual limit expectations then leads to a more optimistic or pessimistic attitude. The second component reflects the individuals’ belief that their limits will prevail on the market and is highly influenced by interaction. The agents anticipate the expectations and reactions of the other individuals in their own decision. Therefore, this component is the more speculative element in the decision process. By distinguishing this, we follow to a certain extent the uncertainty concept of Knight that the business man himself not merely forms the best estimate he can of the outcome of his action, but he is likely also to estimate the probability that his estimate is correct. The individual information processing and the mutual influence upon one another determine the final price expectation and investment decision. The aggregation of the individual decisions then leads to the market price of the next period. Therefore, every agent is described by a single net, the Multilayer Perceptron, but all individuals are connected with each other representing the macroeconomic level. The aim of the agents is to learn the market structure in order to make forecasts of probable yield. The learning process is based on the difference between the expected and actual price. By adjusting the connection weights according to the backpropagation algorithm, this error leads to a change of the magnitude of impact of the macro information as well as the anticipated price limits of the other market participants. Because of the formalism of this learning algorithm each investor considers the past decisions as "training examples". Mentally, at the end of each period with the actual connection weights they look upon all well known decisions of the former periods. Therefore, the opinion is taken into account that precedents have an important influence on later action. The simulation studies of this model show, that the agents adapt to each other generating a decline in the total market error. This result holds for constant information as well as for periodical fluctuations. Market entries of extreme optimistic or pessimistic agents can disturb this structure and induce erroneous forecasts of the remaining investors. On the microeconomic level it can be seen that similar characters can profit from each other. Precondition therefore is an income related dominant market position of some of them. | |||