Parallel Heterogeneous Multi-classifier System for Decision Making in Algorithmic Trading
The most important factors of successful trading strategy are the decisions to sell or buy. We propose multi-classifier system for decision making in algorithmic trading, whose training is carried out in three stages. At the first stage, features set is calculated based on historical data. These can be oscillators and moments that used in technical analysis, other characteristics of time series, market indexes, etc. At the second stage, base classifiers are trained using genetic algorithms, and optimal feature set for each of them is selected. At the third stage, a voting ensemble is designed, weights of base classifiers are selected also using genetic algorithms. However, the usage of genetic algorithms requires considerable time for computing, so the proposed system is implemented in a parallel environment. Testing on real data confirmed that the proposed approach allows to build a decision-making system, the results of which significantly exceed the trading strategies based on indicators of technical analysis and other techniques of machine learning.