Enhancing Long-Term GDP Forecasting with Advanced Hybrid Models: A Comparative Study of ARIMA-LSTM and ARIMA-TCN with Dense Regression

Accurate long-term forecasting of Gross Domestic Product (GDP) is crucial for informed policy-making and strategic economic decisions. This research paper compares two hybrid forecasting models: ARIMA-LSTM and ARIMA-TCN. We also introduce an innovative methodology where linear and non-linear GDP components are fed into dense regression layers to enhance forecast accuracy. By combining the strengths of linear autoregressive integrated moving average (ARIMA) models with the memory-retaining capabilities of long short-term memory (LSTM) networks and temporal convolutional networks (TCN), we create hybrid architectures that capture diverse patterns in GDP time series. Additionally, dense regression is utilized to learn the optimal combination of components to improve accuracy further. Our empirical analysis involves extensive experimentation on real-world GDP datasets, assessing the models’ predictive capabilities in long-term forecasting through evaluation metrics such as MAE and RMSE. The investigation reveals that the ARIMA-LSTM hybrid model outperforms other models, demonstrating a superior ability to minimize significant errors in the presence of heteroskedastic innovations. These findings underscore the importance of hybridizing ARIMA and LSTM to enhance GDP predictive accuracy in volatile economies.