Genetically Optimised Artificial Neural Network for Financial Time Series Data Mining

This paper is a step towards the econometric foundation of computational intelligence in finance. Financial time series modeling and forecasting are addressed with an artificial neural network, examining issues of its topology dependency. Structural dependency of results is viewed not as a model’s weakness, but as a (current) limitation to explain existent relationships. Simulations of price forecasts and trading strategies development reveal optimal network settings, considered further as nonlinear generalizations of the ARMA processes. Optimal settings examination demonstrates weak relationships between statistical and economic criteria. The search for a statistical foundation of computational intelligence in Finance calls for a parallel search for its economic basis. The choice of evaluation criteria combining statistical and economic qualities is viewed as essential for an adequate analysis of economic systems. Our research has demonstrated that fine-tuning the artificial neural network settings is an important stage in the computational model set-up for results’ improvement and mechanism understanding. Genetic algorithm is proposed to be used for model discovery, making technical decisions less arbitrary and adding additional explanatory power to the analysis of economic systems with computational intelligence. Considering a stock market as a complex socio-economic system, this paper examines the evolutionary artificial neural network settings for price forecasting and trading strategies’ development. After learning financial time series dynamics, economic agents search for optimal predictions, exploiting existing temporal correlations of the data.