Combining market-guided patterns and mamba for stock price prediction

Stock prices prediction is a highly challenging task over many years, owing to the market’s high volatility. With the development of deep learning, various studies has focused on modeling temporal patterns for stock price prediction. Most existing approaches rely on a shared neural architecture that captures temporal patterns from individual stock series and then combines these temporal representations to form stock correlations. To overcome the above-mentioned problems, a novel market-embedding with Mamba (MEM) architecture, is proposed. Specifically, MEM consists of the following architectures, coarse-grained feature aggregation, fine-grained feature aggregation, temporal feature aggregation, which can extract the discriminative patterns for stock price prediction. The various experimental results show that the proposed method MEM surpasses previous approaches on two publicly available datasets, i.e., CSI300 and CSI800.