A U-MIDAS modeling framework for forecasting carbon dioxide emissions based on LSTM network and LASSO regression

Accurately predicting CO 2 emissions holds profound significance for the Chinese government as it shapes policies aimed at realizing the China’s “dual carbon” goals of carbon peaking and carbon neutrality. This research focuses on improving the dynamic forecasting accuracy of CO 2 emissions while including mixed-frequency independent variables. We propose a novel hybrid model, U-LSTM-LASSO, which combines the advantages of LSTM network for non-linear pattern recognition and LASSO regression for efficient feature selection within the unrestricted mixed data sampling (U-MIDAS) regression framework. Then U-LSTM-LASSO model is compared with two separate models, U-LSTM and U-LASSO, across all combinations of forecast horizons ( h ) ranging from 1 to 8 and maximum lag orders ( K ) set at 2, 5, and 8. The forecast performance is evaluated using root mean squared error (RMSE), the Diebold–Mariano (DM) test, and the cumulative sum of squared forecast error loss differential (CUMSFE). Empirical findings demonstrate that U-LSTM-LASSO exhibits lower testing RMSE values than U-LSTM and U-LASSO across all h and K combinations. The results of DM test indicate that U-LSTM-LASSO shows statistically more accurate forecasts than U-LSTM across all h and K combinations. Similarly, U-LSTM-LASSO statistically outperforms U-LASSO in all but two cases, when h = 3 , K = 8 and h = 6 , K = 8 . Further, the advancement of U-LSTM-LASSO model remains consistent throughout the entire testing period, as evidenced by the average CUMSFE values, all of which exceed 0 at every timestamp in the testing set. Additionally, U-LSTM generally outperforms U-LASSO in terms of RMSE under most situations, yet it does not exhibit a statistically significant difference in forecast accuracy across all h and K combinations. • A novel U-LSTM-LASSO model is proposed for forecasting CO 2 emissions. • Mixed-frequency data is incorporated without loss of information. • U-LSTM-LASSO is compared with two separate models, U-LSTM and U-LASSO. • U-LSTM-LASSO generally outperforms U-LSTM and U-LASSO numerically and statistically. • The superiority of U-LSTM-LASSO is persistent throughout the testing periods.