McVCsB: A new hybrid deep learning network for stock index prediction

Abstract

Forecasting the stock composite index is a challenge on account of the abundant noise-induced high degree of non-linearity and non-stationarity. Numerous predictive models based on statistical and intelligent methods have been proposed but using a single model or method cannot afford noise reduction so as to achieve credible results in facilitating regulation and investment. This study presents a novel hybrid model of multi-channel VMD-CBAM-BiLSTM named McVCsB based on an improved decomposition-reconstruction predicting framework for stock index prediction. The variational modal decomposition algorithm (VMD) and convolutional block attention module (CBAM) are adopted to achieve noise filtration and deep-level feature extraction. The denoised and concentrated feature will feed the predicting mechanism of bi-directional long-short-term networks (BiLSTM) to produce a stock index for the next day. The multi-channel input structure is innovated to resolve the inherent cumulative error problem of the decomposition-reconstruction framework by changing the parallel processing manner. We design three levels of model comparison by selecting and constructing a variety of benchmark models. In both statistical and practical assessments, the proposed model outperforms others with high reliability and robustness. According to the empirical results on four stock composite indexes, the McVCsB gives the best performance on China’s SCI out-of-sample prediction with the mean absolute error of 0.0016, root mean square error of 0.0019, symmetric mean absolute percentage error of 0.3862, R2 of 0.9989 and Sharpe ratio of 0.6648. Thus, this new proposed model provides an effective predicting tool and can be extensively applied to stock markets in a more widen range.