A Multimodal Temporal Feature Fusion Method Based on Time-Lagged Tensor CP Decomposition and Its Application to Stock Market Indices

Multimodal time-series data are extensively utilized in weather forecasting, traffic flow modeling, and energy management, frequently exhibiting intricate hysteresis interaction structures among their diverse, heterogeneous properties. It is challenging to capture the dynamic linkages across high-dimensional, diverse, and non-synchronous modes simultaneously, and existing approaches have major drawbacks when it comes to modeling cross-modal interdependence and time-lag coupling. This research therefore suggests a feature fusion technique based on time-lag cross tensor and CP decomposition. The approach uses CP decomposition to extract low-rank potential features, reorganizes and fuses the features to suit the prediction job, and builds a tensor structure that fuses time delay relationships from several sources. This study compares the approach with ARIMA, single-feature GAN, multi-feature GAN, and other models after validating it on actual multimodal time series data. The usefulness of the suggested approach in multimodal time series modeling is confirmed by the experimental findings, which demonstrate that it outperforms the current models in terms of prediction accuracy and stability.