Enhancing financial time series forecasting with hybrid Deep Learning: CEEMDAN-Informer-LSTM model

Abstract

Financial time series forecasting is fraught with challenges due to the significant noise and uncertainty in the financial market that can bias model prediction outcomes. However, deep learning, as an important branch of artificial intelligence, has demonstrated a strong ability in dealing with large-scale nonlinear data. Therefore, in order to solve this problem, this paper mainly focuses on the closing price of the CSI 300 index as the research object and proposes a new deep learning hybrid prediction model, CEEMDAN-Informer-LSTM. The model decomposes the signals using complete ensemble empirical mode decomposition of adaptive noise (CEEMDAN) and classifies the decomposed signals by the zero-mean T-hypothesis testing method into high-frequency (H-IMF) and low-frequency components (L-IMF). In order to utilize the prediction advantages of different models in different frequency ranges, as well as to more accurately capture the intrinsic patterns and features in the signals, the combination of Informer prediction of H-IMF and Long Short Memory Model (LSTM) prediction of L-IMF is used to form a hybrid model. In the empirical study, this paper is presented by comparing with BP, RNN, LSTM, Informer, Transformer, iTransformer, CEEMDAN-BP, CEEMDAN-RNN, CEEMDAN-LSTM, CEEMDAN-Informer, CEEMDAN-Transformer, CEEMDAN-iTransformer models are compared and four loss functions, MAE, RMSE, MAPE, and $R^2$, are selected to evaluate the model performance. The experimental results show that Informer has the best results in individual model prediction accuracy, followed by iTransformer, Transformer, LSTM, BP, and RNN. The prediction performance of the proposed CEEMDAN-Informer-LSTM hybrid prediction model is higher than all other models. In order to verify the robustness of the model, this paper has found out that the proposed hybrid model has good prediction accuracy and also exploited the application of the model in the stock market through multi-step prediction, MCS confidence test, dataset discussion, data leakage processing, replacing experimental data, and constructing a simple quantitative trading investment strategy from which the proposed hybrid model has good prediction accuracy.