Forecasting carbon prices under diversified attention: A dynamic model averaging approach with common factors

Abstract

To improve the predictability of carbon prices under diversified attention, this paper develops a dynamic model averaging approach with common factors (DMA-CF) which uses dimension-reduction techniques to extract factors from all models including the subsets of attention predictors and allows time-varying coefficients and model switching. The in-sample results using univariate models reveal the strong predictive power of the attention measured by the Google search volume index (AGSVI). The out-of-sample results confirm it as the strongest attention predictor, and show the superior performance of DMA-CFs relative to the original DMA and the benchmark. We further investigate how common factors improve the forecasting performance of DMA. The empirical evidence indicates that common factors efficiently aggregate the information and reduce the estimation errors in complicated models which are assigned with higher probability in DMA-CFs than in DMA, thereby digging out more predictive information for forecasting carbon prices. Moreover, DMA primarily depends on AGSVI, with the highest weights, while DMA-CFs slightly downweight the AGSVI and allocate more weight to the attention proxy of abnormal trading volume (AAVol) which provides complementary information. Finally, our DMA-CF methods can improve the economic gains in the portfolio exercise with carbon futures.