Abstract

Financial time series are characterized by high volatility and non-linearity, presenting significant challenges for analysis. Traditional statistical methods, such as ARIMA and GARCH, struggle with non-linear data, while machine learning and deep learning techniques can capture intricate price transformations but are often susceptible to overfitting. In addition, the limited parameters of one-dimensional financial time series signals restrict feature representation. To address these challenges, we propose an efficient multimodal financial time-series prediction network with self-supervised Learning, employing the custom-designed SPPMFN network for stock trend forecasting. Firstly, we introduce a novel signal transformation strategy to capture and analyze richer multi-scale feature representations in financial time series signals. Specifically, we convert one-dimensional stock price time series data into two-dimensional image sequence representations spanning different time intervals through the Gramian Angular Fields. Then, both modalities of data are simultaneously input into the SPPMFN, enabling it to learn features from different dimensions. Moreover, we proposed a self-supervised learning framework, which is instrumental in strengthening the model's capacity to identify intrinsic data relationships, allowing the model to detect underlying patterns and structures while efficiently reducing overfitting. Experimental evaluations on the CSI300E and CSI100E datasets confirm the efficacy of our approach, which accurately predicts high-yield stocks and significantly outperforms industry benchmarks, providing robustness and exceptional performance in dynamic financial data environments. In particular, our method has a significant superiority in the performance of long-term prediction.