<p>Predicting stock market trends has long intrigued investors, who evaluate potential returns before making investment decisions. Traditional Long Short-Term Memory (LSTM) models face challenges in data mining and parameter selection due to inherent uncertainty and randomness. This study proposed a novel hybrid model, “STL-HES-ED-BiGRU”, using the Smooth Maximum unit (SMU) activation function to address these issues. The proposed model preprocesses data using Seasonal-Trend LOESS (STL) decomposition to extract informative features. Holt’s Exponential Smoothing (HES) is then applied to model the trend and seasonal components separately. The residual components are fed into an Encoder-Decoder based Bi-directional Gated Recurrent Unit (ED-BiGRU) network with SMU activation to capture complexity temporal dependencies and improve prediction accuracy. The SMU activation function dynamically adjusts its shape, enhancing flexibility in learning non-linear patterns. The model was evaluated on two stock price datasets: General Electric (GE) and Life Insurance Corporation of India (LIC). GE provides daily, weekly, and monthly stock price data, while LIC offers one-minute, three-minutes, and ten-minutes stock data. Experimental results consistently demonstrate high accuracy, Coefficient of Determination (R²) values, and lower error metrics, Mean Absolute Error (MAE), Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Theil’s U-Statistic (TUS). The STL-HES-ED-BiGRU-SMU model surpasses other established techniques, offering robust and reliable predictions for investors and financial experts through its bi-directional learning approach, which collects historical and future data points. This model’s adaptability and flexibility make it a powerful tool for forecasting stock price movements in volatile financial markets. Future work will focus on expanding the model’s applicability to other financial instruments and testing its robustness across market conditions.</p>