A PERIODICITY EVIDENCE OF PRICE AND RETURN FORECASTING OF MAJOR TECH STOCKS USING LSTM MODELS

Abstract

This study investigates the predictive performance of long short-term memory (LSTM) neural networks in forecasting stock prices of four major technology companies, including Tesla, Google, Apple, and Amazon, using both daily and weekly datasets. The empirical analysis utilizes historical stock price data spanning from January 1, 2015 to October 31, 2023, covering both training and testing periods. The research examines multiple dimensions of forecasting accuracy, including comparisons between log price and log return data, as well as the influence of data periodicity on model performance. Results indicate that LSTM models are highly effective in capturing temporal dependencies, with daily datasets generally yielding superior predictive accuracy compared to weekly datasets. Log price-based models consistently outperformed log return-based approaches, highlighting the critical role of data representation in time-series forecasting. Furthermore, variations in predictive performance across stocks suggest that asset-specific characteristics, such as volatility and liquidity, significantly influence model efficacy. The findings provide valuable insights for investors, financial analysts, and policymakers regarding the adoption of deep learning techniques for informed decision-making, risk management, and portfolio optimization. In conclusion, the study confirms the robustness and applicability of LSTM models for stock price forecasting, particularly when high-frequency data and appropriate data representations are employed.] [NEW: It is recommended that practitioners and policymakers support the integration of advanced deep learning models and high-quality market data to enhance forecasting accuracy and financial decision-making.