Decomposition-Based Transformer-LSTM Framework for Stock Index Forecasting

Stock index forecasting presents a complex challenge due to the nonlinear, nonstationary, and noisy nature of financial time series data. This paper introduces a novel hybrid framework that combines decomposition techniques with transformer- and LSTM-based models to address these challenges. The technique applies CEEMDAN to raw time series data, decomposing it into temporal IMFs and a residual component to reduce noise and highlight patterns of interest. Components that behave similar are grouped together based on their statistical and frequency traits, allowing for targeted modeling of the grouped IMFs. A hybrid model aims to leverage the transformer's capability to capture long-range dependencies with the sequential modeling power of LSTMs for each cluster, facilitating multiresolution forecasting. The generalized XGBoost meta-learning layer combines predictions from the models for each IMF while reducing the chance of overfitting. Four major stock indices—Bank Nifty, S&P 500, NASDAQ, and DJIA—are used to validate the framework. The proposed model performs better than baseline models using key metrics. The results demonstrate the proposed method's efficiency in real-world applications as a scalable and reliable method for stock market forecasting. Future research will evaluate the model under various market assumptions and study ensemble techniques for the fusion layer.