Investigation of the role of data conversion on prediction results: Evidence from USA's energy-related emissions and source-based energy use data

Abstract

Considering data conversion practices in empirical research, this research investigates the role of data conversion on prediction results in the United States (USA), where yearly and monthly data on energy-related carbon dioxide (CO₂) emissions and source-based energy consumption is available, which makes the USA an appropriate case for empirical analysis. In this context, this study considers CO₂ emissions as the dependent variable, uses source-based energy use indicators as the explanatory drivers, and performs cointegration regression (CR) approaches on monthly datasets between 1989/1 and 2023/12, which consist of monthly original series (MOS), monthly converted series by quadratic-average-approach (MCSQA), and monthly converted series by quadratic-average-sum (MCSQS). The empirical results reveal that (i) data conversion increases R² values and improves the goodness of fit criteria of the prediction models, where training and testing results are above 96%; (ii) data conversion causes a change in the coefficients of the explanatory variables. While the direction of the variables changes from MOS to MCSQA and MCSQS, it is the same across between MCSQA and MCSQS, but coefficients and p-values slightly differentiate; (iii) dynamic OLS approach has the highest prediction performance among approaches applied; (iv) the importance of source-based energy use indicators on CO₂ emissions differentiate. Overall, the study empirically demonstrates the increasing but varying impact of data conversion on prediction results. Accordingly, the study discusses to benefit of the use of converted data series in empirical predictions, where policymakers can benefit from increasing the impact of data conversion on prediction capacity and prevent incorrect prediction results.