Comparative Performance Analysis of Deep Learning, Classical, and Hybrid Time Series Models in Ecological Footprint Forecasting

Abstract

In a globalized world, factors such as increasing population, rising production rates, changing consumption habits, and continuous economic growth contribute significantly to climate change. Therefore, successfully forecasting the Ecological Footprint (EF) effectively indicates global sustainable development. Despite the significant role of the EF as one of the indicators of sustainable development, there is a gap in the literature regarding time series methods and forward-looking predictions. To address this gap, Ecological Footprint (EF) forecasting was performed using deep learning methods such as LSTMs, classical time series methods like ARIMA and Holt–Winters, and the developed hybrid ARIMA-SVR model. In the scope of the study, first, a spreadsheet was created using the total Ecological Footprint (EF) worldwide between 1961 and 2022, obtained from the Global Footprint Network database. Second, the forecasting performances of the ARIMA, Holt–Winters, LSTM, and the hybrid ARIMA-SVR models were compared using MAPE and RMSE metrics. Finally, the forecasting performances of the time series models were statistically validated through Wilcoxon Signed-Rank and Friedman tests. The study findings indicate that the proposed ARIMA (1,1,0) model demonstrated better performance with an average MAPE of 2.12%, compared to Holt–Winters (MAPE of 2.27%), LSTM (MAPE of 3.19%), and ARIMA-SVR (MAPE of 2.68%) methods in the test dataset. Additionally, it was observed that the ARIMA model forecasted the EF, which experienced a sudden decrease due to the COVID-19 lockdown, with a lower error compared to other models. These findings highlight the adaptability of the ARIMA model to variable and uncertain conditions.