Hybrid Models Using ARIMA and LSTM for Stock Price Prediction

Stock price forecasting remains one of the most challenging problems in financial analytics due to the presence of noise, structural breaks, volatility clustering, and nonlinear dependencies within market data. Traditional time-series models such as ARIMA (Autoregressive Integrated Moving Average) have been effective in extracting linear and autocorrelated patterns but often fail to capture nonlinear behaviors inherent in real-world financial series. Conversely, modern deep learning architectures like Long Short-Term Memory (LSTM) networks excel at modeling nonlinear and long-range dependencies but may underperform when the underlying series contains strong linear or seasonal components. To bridge this gap, this study proposes a comprehensive hybrid ARIMA–LSTM forecasting framework that systematically decomposes and models both linear and nonlinear structures in stock price movements. The hybrid pipeline first applies ARIMA to learn and predict the linear component of the OHLCV time series. The residual errors—representing the nonlinear portion of the dynamics— are then fed into an LSTM network trained to capture complex patterns unexplained by ARIMA. The final forecast is generated by combining ARIMA’s linear predictions with the LSTM’s nonlinear residual predictions. This integrated approach is implemented on multiple NSE and BSE equities using historical data obtained from publicly available sources. Model performance is evaluated using RMSE, MAE, MAPE, and directional accuracy, and is compared against standalone ARIMA and standalone LSTM baselines. Experimental findings demonstrate that the hybrid ARIMA–LSTM model consistently reduces forecasting error, enhances trend-direction prediction, and provides smoother and more robust forecasts in volatile market conditions. The results highlight the practical value of hybrid modeling for quantitative trading, algorithmic strategy development, risk assessment, and financial decision support systems. This study therefore reinforces the importance of combining statistical and deep learning methods to improve predictive reliability in complex financial environments.