Seth Gyamerah, Andrews Awuah
Abstract: Trend forecasting of financial time series is a complex task, many researchers have use statistical models, time series models and machine learning model in this filed of study. The advancement in computational technologies has made it easier, deep learning models is capable of handle time series data. In this research, our objective is to design a system architecture trading strategy algorithm to track the trends movement of prices and to verify if the principal component as input variables would increase the accuracy of Deep learning models in stock market predictions task. We used S&P500 stock exchange price data in our experiment conducted. In the first stage, we preprocess the dataset where we use wavelength transform to denoise the noise in the dataset and created a new factor differences as input variables, (the differences between Low price, High prices, etc). In the second stage, we use principal component as input variables to investigate if it will increase the model accuracy. We also selected technical indicators as input variables. The final stage, we trained and tested our design system Architecture by using LSTM model to forecasting exact future trend of the prices. We compare the result to a baseline machine learning model, Random Forest and Logistic regression model. Thorough empirical studies based on S&P500 dataset; our design system LSTM model demonstrate a high accuracy that outperform state of art- methods for trend forecasting of financial time series.
Keywords: LSTM, S&P500, Financial time Series, Technical indicators, Principal component analysis and wavelet transform