TimesFNP: Contrastive Learning for Financial Domain with Noise-Resilient Prediction

Long-term time series forecasting is a long-standing research topic in financial scenarios. Recent studies have shown that Transformer models have great potential in time series forecasting. However, time series in financial scenarios contain complex noise disturbances. To address this issue, we propose a new framework called TimesFNP, which aims to improve the noise resistance of long-term time series forecasting in financial scenarios based on contrastive learning methods. Specifically, first, we introduce a noise-based data augmentation method to simulate the noise components in financial scenarios. Second, we fully consider the potential correlations between different companies in long-term time series forecasting tasks and represent the time series features at the company level. Remarkably, our framework demonstrates significant improvements over state-of-the-art algorithms in time series forecasting, achieving performance enhancements of at least $\mathbf{8. 5 7 \%}$ in MSE and $\mathbf{4. 4 2 \%}$ in MAE