The research studies the impact of economic recessions on India’s Nifty 50 stock index, using advanced deep learning models to predict closing prices during recession periods. The proposed work analyzed the efficiency of four models: Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Convolutional LSTM (CNN-LSTM), and Transformer models, to determine which is most effective at predicting stock prices. To train each model, the model applied historical Nifty 50 data from some important recessionary events, i.e., the 2008 Global Financial Crisis and the COVID-19 pandemic. To evaluate their accuracy, we used key metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R2. Among the models tested, the Gated Recurrent Unit (GRU) model consistently delivered the most accurate predictions, achieving the lowest error rates. These studies concentrate on improving Gated Recurrent Unit (GRU) networks and developing a new GRU-Transformer model. The GRU model was tuned to increase the effectiveness of its forecasting through hyperparameter tuning. At the same time, the hybrid model uses the sequential dependency management strength of GRU, while the long-range correlations, attention mechanisms, and multi-head self-attention of the Transformer are used for the management of the long-range correlations. Experimental results confirm that the GRU model has the highest forecasting accuracy, while the hybrid model boasts overall better accuracy. The paper also shows that these models have useful applications in finance based on back-testing and direction accuracy. It has been established that hybrid architectures can greatly improve stock market analysis, which is advantageous in financial analytics.