Multivariable financial time series forecasting based on phase space reconstruction compensation

Accurate financial time series forecasting is an important challenge in the financial field due to varying levels of interaction among multiple financial time series, complicating the extraction of valid information from these variables. This study introduces an effective and efficient multivariable financial time series forecasting model based on phase space reconstruction compensation: Phase Space Reconstruction Compensation Long Short-Term Memory (PSR-LSTM). The PSR-LSTM model leverages the long short-term memory (LSTM) network to analyze short-term data behaviors, reconstructs multiple long-term variables by phase space reconstruction method, and utilizes multivariable trend attention to capture trend information in correlated variables. This trend information is finally used to correct the LSTM network’s predictive result. Experimental results demonstrate that the PSR-LSTM outperforms existing multivariable forecasting models by effectively mitigating noise interference while achieving optimal forecasting performance.