RF-LSTM carbon price prediction based on CEEMDAN decomposition and multiscale entropy reconstruction

Abstract This paper proposed an RF-LSTM hybrid prediction model based on CEEMDAN decomposition and multiscale entropy reconstruction. The new model solves the problem of insufficient prediction accuracy caused by the carbon price series’ nonlinearity, non-stationarity, and multifractal characteristics. The method first decomposes price series into IMFs using CEEMDAN, then reconstructs components through multiscale entropy analysis to reduce noise interference. A dual prediction framework combines Random Forest (capturing nonlinear patterns in high-frequency components) and LSTM (modeling long-term dependencies in low-frequency components). Empirical tests on Hubei and EU carbon market data show: (1) The model achieves superior accuracy over benchmarks (CEEMDAN-LSTM/LSTM) across multiple metrics; (2) Maintains high computational efficiency; (3) Demonstrates optimal comprehensive performance. Results verify that entropy-based reconstruction effectively mitigates mode mixing, while RF-LSTM synergy enhances cross-frequency prediction capability, providing methodological support for carbon market risk management.