Stock price trend forecasting based on multi-channel complementary network with CEEMDAN decomposition and transformer residual prediction

Abstract

Forecasting stock market movement provides important investment signals, but the presence of a large amount of noise in financial time series (FTS) data poses significant challenges to prediction accuracy. This paper proposes a novel multi-channel complementary network with data decomposition and a fusion strategy to effectively improve the prediction accuracy and robustness of stock price movement. The model first applies the algorithm of Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose the historical stock data sequence into a set of Intrinsic Mode Functions (IMFs), representing different frequency components of the original FTS data. Then, each IMF combined with six key indicators (i.e., open price, high price, low price, trading volume, price-earnings ratio and price-to-book ratio) are processed by an independent long short-term memory (LSTM) module to make a parallel prediction. To further enhance the forecasting accuracy, a transformer-based residual term prediction model is incorporated, serving as the complementary branch to the LSTM modules. Subsequently, the outputs from all network branches are fused together to obtain the final prediction result. A set of numerical experiments on different stock index datasets are conducted to verify the superiority of the proposed model in terms of average forecasting accuracy compared with other benchmark models. In addition, the effectiveness of different sub-modules in the proposed framework is proved by the ablation experiments.