| Abstract [eng] |
In this thesis, chart pattern influence on the GA-LSTM-CNN model is investigated. Chart patterns are a branch of technical analysis that uses stock price movement shapes or candlesticks to predict future price movement. The GA-LSTM-CNN model was selected due to its combination of several neural network algorithms that achieve self-structuring, temporal and spatial pattern recognition. As it is not clear which features should be chosen for stock trading due to numerous trading strategies, a genetic algorithm provides an automatic search for features that best fit the use case. LSTM and CNN combinations provide the model with temporal and spatial pattern recognition capabilities, respectively. That is needed in stock trading, as patterns exist between features and in time for the same feature. Before the preliminary tests were conducted, chart patterns were selected from financial data using analytical rules outlined in the analysed research paper. Searching of chart patterns was done backwards to not give the model any knowledge of financial events in the future. Preliminary tests showed that the model selected chart patterns as one of the features. The results were similar to those of the model without chart patterns. Final tests revealed that chart patterns do not have a significant impact on individual model performance, but could be useful in an ensemble scenario. This thesis is separated into seven sections. The first section is a brief introduction to the problem, object of experiments and experimental steps taken. The second section is state-of-the-art research, where 29 sources are analysed to get a comprehensive overview of automated stock market trading. The third section contains explanations of used methods, models and metrics. The fourth section has descriptions of the training and testing environments, data used for experiments. The fifth section has comparison of model testing results. The sixth section is a discussion on problems encountered in the experiments, successes and future work. Finally, conclusion is given on model performances in the last section. In total, this work contains 41 figures and 53 pages. |
