Forecasting Carbon Dioxide Emissions of China, the United States, and the European Union: Comparing ARIMA and Naïve Predictions

Background

• Earth climate changes become more apparent across the world, with scientists attributing the rising temperatures primarily to carbon dioxide emissions.
• The top three emitters by annual CO2 emissions in 2017 are China (27% of global emissions), the United States (US; 15%), and the EU-27 (10%).
• There is a strong need of policymakers to monitor the current status of the CO2 emissions as well as to understand their trajectories to design & implement effective policy responses.

Objective

• Use forecasting methods to forecast the CO2 emissions of the top 3 emitting countries or unions until 2030.
• Demonstrate the use and the limitations of ARIMA models and share concrete steps to improve the predictions.

Approach

• Apply ARIMA methods following the Box-Jenkins-Approach to generate univariate one-step ahead and 2030 forecasts.
• Benchmark the method versus a simple naïve forecast and evaluate it thoroughly using RMSE, MAE, MAPE, and MASE measures.

Key Results

• (0, 2, 1), (1, 1, 0), and (0, 1, 1) ARIMA models were used as final models for EU, China, and US, respectively. The EU and China model passed the residual checks.
• The naïve forecast outperformed the ARIMA forecast for Europe, whereas the ARIMA model performed better across measures for China (due to heteroscedasticity, the U.S. model was not used for forecasting; a GARCH model would be required):

• The 2030 forecast for the EU and China is shown below:

• The Box-Jenkins-Approach appears to not be well-suited for applied long-term predictions. Future work should utilize more sophisticated methods (ensembles of statistical models, deep learning models that can model dependencies such as (bidirectional) LSTMs), utilize cross-learning or multivariate models, and include relevant covariates to improve the predictive performance of the model.

Installation Instructions

• Install R (4.2.1) and RStudio
• Change the data loading path to your local working directory
• Run the script and consider reading the written report for in-depth insights