Multi-modal and Optimised LSTM-CNN Attention based Framework for Prognosis of Dynamics in Stock Prices

Abstract

However, the problem of equity price prediction is still complicated due to the interaction of the factors involved, which are saliently nonlinear, highly volatile and sentiment driven. This is just further complicated by the fact that investor sentiment is less or more unpredictable because it's not quantifiable, and it contains emotions. The research aims to present a hybrid deep learning structure of LSTM-CNN-Attention mechanism and technical, fundamental and sentiment-based features to enhance the accuracy of the forecasts. The model features three indicators sets, price trend, market volatility and sentiment indicators, the latter on the basis of the social media, online financial news and online newspapers. A method is used to eliminate the redundant inputs and to keep the best inputs of the described inputs, namely Principal Component Analysis (PCA) and Sequential Feature Selection (SFS). These are the feature selection techniques which enhance the capacity of the model to capture important market trends and behavioural indications. The comparison is carried out using the set of 10 industrial domains, involving five different deep learning architectures including LSTM, CNN, and LSTM-CNN hybrid, LSTM with Attention and LSTM-CNN-Attention. It is seen that the LSTMCNN Attention configuration performs best among the different configurations as it has the lowest Mean Absolute Error (MAE) in most industries. The results show the significance of feature optimization and sentiment integration to enhance predictive performance. The suggested framework provides a scalable, interpretable, and data-driven platform to real-time sentiment-based price forecasting of equity and intelligent investment decision support.