Abstract
Stock market forecasting using technical indicators (TIs) is widely applied by investors and researchers. Using a minimal number of input features is crucial for successful prediction. However, there is no consensus about what constitutes a suitable collection of TIs. The choice of TIs suitable for a given forecasting model remains an area of active research. This study presents a detailed investigation of the selection of a minimal number of relevant TIs with the aim of increasing accuracy, reducing misclassification cost, and improving investment return. Fifty widely used TIs were ranked using five different feature selection methods. Experiments were conducted using nine classifiers, with several feature selection methods and various alternatives for the number of TIs. A proposed cost-sensitive fine-tuned naìˆve Bayes classifier managed to achieve better overall investment performance than other classifiers. Experiments were conducted on datasets consisting of daily time series of 99 stocks and the TASI market index.