Abstract
Stock market prediction remains a challenging endeavour, given the volatile and complex nature of financial markets. In this research, we investigate the effectiveness of different Long Short-Term Memory (LSTM)-based neural network architectures, including LSTM, Bidirectional LSTM (BiLSTM), and CNN-LSTM, to forecast stock prices. We specifically focus on evaluating these models using various time frames of high-frequency stock data to understand which configuration yields the best predictive performance. Our analysis utilizes Apple Inc. stock data collected over different periods and frequencies: 7days of every 1-minute data, 30 days of every 5-minute data, and 10 days of both 1-minuteand 5-minute data. Through this study, we demonstrate that the 7-day 1-minute data configuration yields the most accurate predictions among all the considered setups. Performance metrics, including Mean Squared Error (MSE), will be used to evaluate and compare the predictive power of the models. This research highlights the importance of selecting appropriate model architectures and timeframes when developing stock prediction systems. Our findings contribute to the understanding of how high-frequency data can be leveraged in financial forecasting, providing insights into the relative strengths and limitations of different LSTM variants. We leave room for further exploration of optimization techniques and additional data configurations to enhance model accuracy and stability..