Determining the functional structure of financial time series by means of genetic learning

Forecasting of security prices based on historical data has been a very popular and equally demanding application area. Genetic programming (GP) is one of the such data-driven techniques which has been employed in predicting security prices based on historical data. GP has been frequently used in searching the mathematical form of the process that generates financial time series. Besides; GP has been also reported to achieve very promising results as a pattern recognition technique in developing rule based systems which produce signals for entering and leaving the markets. When searching for the predictor mathematical form or for a rule pattern by the machine, an ill-defined optimization problem with a very large search space has to be faced with. Thus; instead of a deterministic one, a satisfying approximate solution within accepted time by limited computing power can only be found by metaheuristics such as GP. To this aim, three systems have been developed throughout this work. The first two of these systems are symbolic regression applications and the other one is characterized as a rule-based system. The first two systems strive to generate predictor mathematical models by the help of symbolic regression from simple terminal and function sets. Then strong and weak sides of this approach is presented. The developed system within the scope of third application, searches whether in/out timing decisions for marketscan solely be determined depending on historical data without "à priori" knowledge. System speculation performance has been tested for different learning and test periods in the frame of a specific experiment design. As a result, it is statistically put forth that, this rule based system developed by the help of GP, could provide superior strategies compared to buy-and-wait strategy for testing periods of 6 months (120 trading days) and 12 months (240 trading days). In contrast, similar findings could not be reached for relatively shorter test periods like 1 week (5 trading days), 1 month (20 trading days) and 3 months (60 trading days). Experiments clearly show that; the system can provide profitable buy-sellstrategies without "à priori" knowledge but only discovering rule combinations of some technical analysis indicators derived fromdaily historical prices. Keywords: evolutionary computation, genetic programming, symbolic regression,rule-based systems, ?nancial time series, technical analysis, ef?cient market hypothesis