A Novel Continuous Peephole LSTM with Neural Controlled Differential Equations for Timely Financial Risk Prediction

Time series forecasting is a critical technique in various domains, including finance, meteorology, and beyond. Traditional forecasting models often rely on metrics like mean squared error (MSE), which fail to adequately address the issue of timely event detection. Predictions that follow actual events-particularly in scenarios such as financial risk-reduce the practical relevance of forecasting, where early and accurate warnings are essential. In this paper, we propose a novel solution to mitigate prediction delays, especially in the context of time series forecasting for financial risk. Specifically, we introduce a Continuous Peephole LSTM framework, integrated with a continuous Neural Controlled Differential Equations (NCDE) approach, to capture intricate temporal dependencies. Instead of relying solely on MSE, we incorporate new evaluation metrics, such as Dynamic Time Warping (DTW) and Temporal Distortion Index (TDI), to quantitatively analyze the timeliness of the model's predictions. These metrics are integrated into the training process, enabling us to evaluate performance from multiple perspectives, including both predictive accuracy and dynamic trend capture. Empirical results across five financial time series datasets demonstrate that our approach outperforms state-of-the-art models. On average, it achieves improvements of 14% in MSE, 6.52% in DTW, and 14.63% in TDI. These results highlight the effectiveness of our model in improving both the accuracy of the prediction and the timeliness of detecting key events in the prediction of financial risks. The source code is publicly accessible at: https://anonymous.4open.science/r/CPLSTM-AC6C/.