A Comparative Evaluation of Noise Reduction Versus Data Normalization Techniques in Stock Market Prediction Using Transformer Models

This study examines the impact of six normalization techniques and nine additional technical indicators on recursive Transformer models for stock market prediction across four sectors: S&P 500, Nasdaq, IBM, and Johnson & Johnson. The research compares custom-implemented recursive models against library-based implementations using MAPE, RMSE, and R-squared metrics. Two datasets are analyzed: raw data alone and raw data augmented with nine technical indicators. Results show custom models consistently outperform library-based ones across all scenarios. With raw data alone, normalization techniques have minimal impact on prediction accuracy. However, when incorporating additional indicators, sector-specific n ormalization s trategies e merge a s c rucial. Min-MaxScaler and MaxAbsScaler prove most effective for broad market indices (S&P 500) and technology stocks (Nasdaq, IBM), while StandardScaler and RobustScaler excel for healthcare stocks (Johnson&Johnson). The addition of technical indicators enhances prediction accuracy and reduces noise, particularly in volatile sectors. These findings underscore the importance of tailored preprocessing approaches in financial time series forecasting, emphasizing the need for sector-specific strategies to enhance prediction accuracy and model robustness across diverse market segments.