ABSTRACT Financial series are challenging to predict owing to their inherent nonlinearity and significant noise. Mainstream prediction typically employs single ensemble models, which suffer from low prediction accuracy and instability. Furthermore, the predictive performance of models with error correction networks has still been unsatisfactory because of the significant randomness and chaos in prediction error. Therefore, we propose a novel dual ensemble (DE) of artificial gorilla troops optimization (AGTO) and the Bayesian regression, combined with dynamic valid residual correction network (DVRCN) architecture for stock index prediction. This dual ensemble‐dynamic valid residual correction network (DE‐DVRCN) model leverages nine machine learning models and six distinct objectives‐based AGTO algorithm to construct the first‐layer ensembles. Next, the Bayesian regression is employed to construct the second‐layer ensemble with these single ensembles. We use variational mode decomposition‐sample entropy network to isolate valid residual information from prediction error, which is then combined with LASSO regression to predict error. Finally, the predicted error is merged with DE predictions to establish the DE‐DVRCN model. The new model combines the strengths of different ensemble strategies, and DVRCN eliminates interference from invalid residual information in predictions. Experiments are conducted on three major stock indices in the Chinese market; it turns out that our model markedly outperforms individual models and single ensembles in predictions.
Cheng et al. (Sun,) studied this question.