A Hybrid CNN-LSTM Deep Learning Framework for Stock Price Prediction: Evidence from HDFC Bank on the National Stock Exchange of India

Abstract

Accurate forecasting of stock prices prediction has historically been considered one of the most complex issues in capital-related research. Markets are inherently messy they are nonlinear, constantly changing and full of noise. This study proposes and tests a hybrid CNN-LSTM model that blends Convolutional Neural Networks (CNN) which are great at spotting short-term local patterns with Long Short-Term Memory (LSTM) networks, which excel at remembering longer sequences. The combined model is applied to forecasting HDFC Bank's daily closing prices on the NSE. Data spanning used 247 trading days of real market data (9 February 2025 to 9 February 2026), enriched with two extra features intraday volatility and Volume-Weighted Average Price (VWAP) to give the model more context. A 20-day sliding window was used to generate 227 input-output pairs, which were then split chronologically (80/20) so the model never "peeked" at future data during training. Three architectures were benchmarked against a naïve random-walk baseline standalone CNN, standalone LSTM and the proposed CNN-LSTM hybrid. The hybrid model clearly outperformed both baselines across every metric achieving RMSE = 0.0228, MAE = 0.0190, and R² = 0.312, versus CNN (R² = 0.129) and LSTM (R² = 0.004). Peak validation accuracy hit roughly 87% for the hybrid versus 80% and 76% for the LSTM and CNN respectively.  Precision, Recall, and F1-Score are computed on direction-classification tasks. In August and September of 2025, a marked near about 50% price drop served as a real-world stress test that brought to light the difficulties that all data-driven algorithms encounter when market conditions shift significantly. The CNN-LSTM hybrid generates significantly superior accuracy in prediction along with quicker converging when contrasted against either a single architecture independently, which has important implications for short-horizon trading methodologies.