Time Series Prediction Analysis of Long Short Term Memory Models

Stock market predictions are an active and broad field of research, and artificial intelligence techniques and algorithms, specifically deep learning algorithms, have helped improve the quality and accuracy of predicting future stock prices using the features available in the data. In this research, our goal is to enhance stock market predictions by using three algorithms: standard long short-term memory (LSTM), encrypted LSTM with decryption, and LSTM using Keras Tuner. These algorithms show promising results in terms of accuracy, in addition to a very important feature, which is identifying appropriate variables to obtain the best accuracy and identifying complex temporal patterns in financial time series. Historical data from SABIC in Saudi Arabia was used, where we used the Exponential Moving Average to get rid of the noise in the data as a common step between the three algorithms used in the research, as it is an obstacle to accuracy. Use the standard LSTM model to evaluate its ability to predict future prices. We then extended our research by applying encoder-decoder LSTM, allowing for sequence predictions and a deeper understanding of market dynamics. To improve prediction accuracy, we used the Keras Tuner library, as this algorithm selects the best variables through the tuning process to improve prediction accuracy. We found that standard LSTM, encoder-decoder LSTM, and LSTM with Keras Tuner, combined with optimized hyperparameters, achieved success rates of 7.062, 6.563, and 4.243, respectively. This study strongly supports the idea that LSTM-based models, especially those equipped with Keras Tuner, play a crucial role in financial risk management and investment decision making.