Investors have always been keen on predicting stock market movements, often calculating their potential return on investment beforehand. This study evaluates stock price predictability using statistical and Deep Learning (DL) models on General Electric (GE) daily, weekly, and monthly closing stock prices. A hybrid model is proposed to address both short-term and long-term dependencies, combining an Encoder-Decoder based Bi-directional Gated Recurrent Unit (Encoder-Decoder Bi-GRU) network with the Double Exponential Smoothing (DES) technique. The Encoder-Decoder Bi-GRU forms a nonlinear prediction model using original data, nonlinear residuals, and linear prediction results, while DES develops a linear model for forecasting linear components. The Smooth Maximum Unit (SMU) Activation Function is mainly incorporated to boost model performance. Applied to GE’s dataset, the DES-ED-Bi-GRU hybrid model shows consistently high R2 values (0.998) with high accuracy (98.7%), and low error values MAE (1.415), MSE (3.579), RMSE (1.891), MAPE (1.449), and Theil’s U-Statistic (TUS) (0.007). Results show that the DES-ED-Bi-GRU model, enhanced with the SMU activation function, surpasses other well-known techniques. Its bi-directional learning captures historical and future data points, providing reliable predictions for investors and financial experts. Future work will involves applying the DES-ED-Bi-GRU hybrid model to a broader range of stocks and sectors to validate its generalization.