Zhora D. Analysis of random subspace classifier and its application for stock market forecasting

This work provides the mathematical study of random subspace classifier characteristics. Methods for improving it's performance are suggested and comparison to other classification algorithms is provided as well. This neural network model shows that it's quite effective when applied for technical forecasting of the stock market. The following approaches for improving classifier characteristics are suggested: optimization of classifier structure configuration parameters in order to maximize the Hamming distance between binary images of two different input vectors; adaptation of threshold density distribution to the actual probability distribution of input vectors; local averaging of synapse matrix coefficients for tasks with overlapping probability distributions. The functioning of this neural network is evaluated on well known classification database. It is proved, that random subspace classifier is the universal classifier. The information theory approach is suggested to estimate the forecasting effectiveness. Besides, the software agent algorithm is developed in order to test and improve stock trading strategies. The classifier architecture allows effective parallelization of corresponding basic algorithms for applications on multicore and/or multiprocessor computing systems.