<p>With the advancements in deep learning architectures (DLAs), different application domains, including healthcare, communication, agriculture, finance, etc., get benefits in many terms. Its inclusion in stock markets in terms of prediction of closing prices can help the investors with better planning and other potential tasks such as portfolio management and decision process. However, the dynamic, unstable, and unknown patterns of stocks pose challenging situations for DLAs and increase the computational overhead. To deal with such challenges, this research work proposes an efficient stock market prediction system (ESMTS) with an ensemble DLA architecture. The convolutional neural network (CNN) in the ensemble model extracts meaningful information, such as trends and correlations in stocks. The recurrent neural network (RNN) focuses on capturing temporal dependencies and sequential patterns of stocks. This ensemble model can enhance generalization, versatility in data handling, and resilience to data shifts. Additionally, an enhanced krill herd optimization (EKH-Opt) is proposed to improve computational efficiency. The proposed model is evaluated using a small-scale and large-scale dataset of stocks from Yahoo Finances. The historical information is extracted from January 2014 to December 2023, and different technical indicators are also computed for better trend analysis and improved predictive power of the system. Different performance indicators such as mean squared error (MSE), <i>R</i>-squared (<i>R</i>²), mean squared logarithmic error (MSLE), explained variance score (EVS), and mean absolute error (MAE) are used for performance analysis and comparison. The proposed system demonstrates its effectiveness with an overall performance based on MSE of 0.0002, <i>R</i>² of 0.99, MSLE of 0.00009, EVS of 0.99, and MAE of 0.0078.</p>