The Investigation of LSTM-Random Search with Various Standardization and Normalization Technologies

Abstract

The stock market, being a complex and chaotic system, presents a formidable challenge for prediction. This paper explores the intricacies of stock prediction by focusing on data preprocessing, specifically investigating the impact of different standardization and normalization techniques on LSTM structures. To enhance predictive accuracy, this study employed a stochastic search technique in this study. It is tested using 1520 data points of the closing prices of Bitcoin and Amazon that were obtained from Kaggle between May 1, 2013, and May 14, 2019. It is discovered that the prediction performance under the LSTM framework is significantly impacted by various data preprocessing techniques. The normalization techniques employing MinMaxScaler and MaxAbsScaler performed better for AMZN, a stock with stable prices, and the RobustScaler normalization method yielded the best results for Bitcoin stock prediction, suggesting that it is more useful for datasets with higher volatility. These results provide a useful guide for future research on stock price prediction by highlighting the significance of selecting the right preprocessing technique based on data characteristics and highlighting the benefits of dynamically modifying the lstm structure for stochastic searches.