Automated Recognition of Harmonic Patterns in Candlestick Chart: A Comparative Study of Machine Learning and Deep Learning Approaches

The stock market provides a platform where investors can easily buy and sell stocks of publicly traded companies. Successful stock market investments can help people accumulate wealth, which can increase consumer spending and support economic growth. As a standard practice many traders, investors and analysts use harmonic chart patterns as a part of technical analysis to make predictions about future price movements. In order to identify the presence of harmonic patterns, three different machine learning models and two variants of Multi-Layer Perceptron deep learning models are experimented in this paper. We have generated Four distinct technical indicators have been derived from the historical dataset of Dhaka Stock Exchange (DSE). In addition, we have analyzed 100-day-duration of candlestick chart to automatically identify the presence or absence of harmonic patterns using manually labeled dataset. According to our experimental studies, the MLP-2, which is a multi-layer perceptron with five dense layers, shows the best overall performance for this binary classification problem of determining whether a harmonic chart pattern is present or not. Using the models of Decision Tree, Random Forest, AdaBoost, MLP-1 and MLP-2, we have achieved accuracy scores of 76.15%, 80.30%, 82.20%, 87.50%, and 90.50% respectively, which is competitive with and often better than other existing models from past investigations.