Maximizing Forecasting Precision: Empowering Multivariate Time Series Prediction with QPCA-LSTM

The increasing complexity and nonlinearity of features in Multivariate Time Series (MTS) data necessitate sophisticated Dimensionality Reduction (DR) techniques to extract meaningful insights and patterns for efficient analysis and interpretation. While applying Principal Component Analysis (PCA), Singular Value Decomposition, and other dimensionality reduction techniques to highly correlated datasets, such as stock datasets, often preserves a high percentage of information, certain MTS datasets with numerous features and higher dimensions, like those utilized in this study, do not retain as much information. Consequently, these datasets fail to provide reliable forecast analysis, underscoring the challenge in effectively analysing and interpreting high-dimensional MTS data using traditional DR methods. This study aims to retain a higher percentage of information from complex MTS datasets. To achieve this, introduce a novel method, Quartet Principal Component Analysis (QPCA), which enhances PCA for dimensionality reduction by maximizing explained variance. This QPCA method retains a higher percentage of information compared to traditional PCA, achieving a 10% increase in explained variance, thus minimizing information loss. Subsequently, Long Short-Term Memory (LSTM) networks are employed for prediction based on QPCA results (QPCA-LSTM). This approach yields superior results compared to baseline methods (LSTM and PCA-LSTM), evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Additionally, the Diebold-Mariano test at the 1% significance level confirms the superiority of the QPCA-LSTM model. This highlights the challenge of effectively analysing and interpreting high-dimensional MTS data using traditional DR methods, emphasizing the advantages of our proposed technique.