A next-generation genetic programming environment for prediction: a sample financial market data application

摘要

The prediction of futures prices from current data is a hard job. Whereas in inferring scientific laws from measured data we may have knowledge of 'measurement error' and some basic deterministic laws, it is not so in the markets. We expect fluctuations in financial data to be near random- the accepted models, such as CAPM, tell us so. Measurement error is important because it provides the basis for statistical decision making as probability theory, classical or Bayesian, would have it. There is, however, an incidental convergence of various mathematical disciplines which promises to tackle this problem. In this paper we present a method of predicting financial data which improves upon current methods in the following important criteria: 1) Algorithmic Search. It searches the space of applicable models (programs) for the most reliable one, like genetic programming (GP). It improves upon the GP paradigm by more efficiently encoding and altering programs. 2) Parameter Search. Unlike GP, we use efficient methods for adjusting continuous valued parameters within a model. The global optimization search used is of the clustering type, which organizes simultaneous downhill (Levenberg- Marquardt) searches over parameter space. 3) Fit Criteria and Validation. One of the problems in comparing different models and in applying them once found is gauging their reliability and performance. This is accomplished by using Kolmogorov complexity theory and probability theory to yield the following simple fit criteria as an extension of the traditional chi-squared error on a data set of N points (see text). 4) Scaleable implementation. This issue deals with the ability of a system identification technique to deal with large sets of data and a wide search over model space. In modem computing terms, this means distributed processing, which both neural network and GP paradigms are eminently capable of utilizing. The current program uses Java to distribute fitting routines across the network, which execute on client computers inside applets.