Intelligent Analysis of Stock Trading Patterns Using Adaptive Learning Models for Market Dynamics and Predictive Indicator Discovery
TL;DR
The results highlight the significance of using alternative information on sentiment along with traditional market indicators and show that the machine learning techniques outperform other techniques especially LSTM.
Abstract
Stock market forecasting is difficult due to its nonlinear nature, volatility, time dependencies, and fast-changing sentiments of investors. Conventional models of statistical nature offer an important benchmark but might be insufficient in capturing the complexity of market dynamics. The objective of this research is to examine whether financial, technical, social media, and HFT data can help forecast stock prices better, and whether models of machine learning will yield superior results compared to conventional statistical methods. The approach in the paper is empirical and based on the use of historical stock market data as well as alternative market data sources. Among the variables used for forecasting are moving averages (MA10, MA50, and MA200), relative strength index (RSI), moving average convergence divergence (MACD), trading volume, and social media sentiment. Models’ Multiple regression, ARIMA, and GARCH are considered traditional models while Random Forest, gradient boosting machine (GBM), and long short-term memory (LSTM) are considered machine learning models. From the findings, it is clear that the regression coefficient (β) for social media sentiments is highest at 0.42, followed by moving average at β=0.35 and MACD at β=0.28, whereas RSI has a negative regression coefficient (β=-0.15). LSTM is the most accurate machine learning technique, recording the lowest error rates of 1.65 for MAE and 2.50 for RMSE, with an accuracy of 0.91. GBM has 1.75 MAE and 2.60 RMSE, whereas Random Forest has 1.85 MAE and 2.80 RMSE. ARIMA and GARCH have higher errors. The results highlight the significance of using alternative information on sentiment along with traditional market indicators and show that the machine learning techniques outperform other techniques especially LSTM. Future studies should use macroeconomic variables, financial news, up-to-date data, transaction costs, and risk-adjusted profit.