Market movement prediction using chart patterns and attention mechanism

Abstract

The new concept of forecasting market movements is introduced to fill the gap in stock price forecasting that uses a fixed time window. The candlestick data are grouped into waves before being input into the Transformer model to capture many different chart patterns and provide longer price movement patterns for better feature extraction. This work is divided into two main parts. The first part is wave extraction, for which we propose a new algorithm for grouping candlestick data into waves. The second part deals with Transformer modeling. The result of this work has shown that the Transformer model trained using wave series can help traders reduce their losses and choose a suitable trading strategy based on the prediction results. The model trading result of the test set has earned about 19.8% (using fixed 1 contract size per order) which is better than using MACD signal which can give earning of 12.5%.