SPPMFN: Efficient Multimodal Financial Time-Series Prediction Network With Self-Supervised Learning

Financial time series exhibit high volatility and non-linearity, making analysis particularly challenging. Traditional statistical methods like ARIMA and GARCH struggle with non-linear data. At the same time, despite capturing complex price dynamics, machine learning and deep learning approaches often face the risk of overfitting. Additionally, the limited parameters of one-dimensional financial time series signals constrain feature representation. To tackle these challenges, we propose an efficient multimodal financial time series prediction network utilizing self-supervised learning based on the custom-designed SPPMFN network for stock trend forecasting. Our approach introduces a novel signal transformation strategy to extract richer multi-scale feature representations from financial time series signals. Specifically, we convert one-dimensional stock price time series into two-dimensional image sequences across varying time intervals using Gramian Angular Fields. These transformed data modalities are then simultaneously processed by the SPPMFN, enabling the extraction of features from multiple dimensions. Furthermore, we introduce a self-supervised learning framework that enhances the model’s ability to capture intrinsic relationships within the data, allowing it to detect underlying patterns while effectively mitigating overfitting. Experimental results on the CSI300E, CSI100E, and S&P500 datasets demonstrate the effectiveness of our method, showing superior performance in accurately predicting high-yield stocks and significantly outperforming industry benchmarks. Specifically, our model improves accuracy by over 3% and achieves a cumulative return of 21.62% on the CSI300E dataset. On the S&P500 dataset, accuracy increased by over 1.4% and cumulative return reached 18.66%. These findings highlight the robustness and practical applicability of our method across diverse financial markets.