Improving Financial Forecasting With Hybrid ARIMA And LSTM Models

Abstract

The financial market is dynamic and subject to linear as well as nonlinear effects, which makes financial prediction and forecasting difficult for conventional approaches. Linear statistical models like the ARIMA model linearize effects as well as seasonality, but fail to address nonlinear dynamics, whereas deep learning models like the LSTM model can handle nonlinear dynamics. In this research, an innovative hybrid ARIMA-LSTM architecture has been introduced, which takes advantage of each of these models to develop better financial prediction capabilities. The historical financial data from January 2010 to December 2023 have been taken into account. Data were processed for normalization, missing value imputations, and stationarity checks. As for the prediction of the linear part, used the ARIMA model, and the LSTM model was applied for the prediction of the non-linear part, considering the residual of the linear part. The optimal prediction was achieved by the use of results obtained from the combination of the two models. The performance of the model was evaluated using RMSE, MAE, MAPE, and R². From the result, ARIMA-LSTM hybrid model performed better than other models considered, giving us RMSE = 9.87, MAE = 7.21, MAPE = 11.3%, and R² = 0.958. The findings indicate that the use of classical statistical models along with the neural network is a successful approach when solving real-life forecasting problems in finance. Potential areas for future research may include the addition of more market indices, multi-step prediction, and the use of attention networks.