Kernel-Based Dimension Reduction Method for Time Series Nowcasting

• Published in: IEEE access Vol. 12; pp. 173223 - 173242
• Main Authors: Do Van, Thanh, Nguyen Minh, Hai
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Approximation; Big Data; Convolutional neural networks; Covariance matrices; Data sampling; Datasets; dimension reduction; Dimensionality reduction; dynamic factor model; Hyperplanes; Kernel tricks; Long short term memory; nowcasting; Prediction algorithms; Predictive models; Principal component analysis; Principal components analysis; Real time; Real variables; Sparse matrices; Time series analysis; Vectors
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3491499

Nowcasts are closely related to big data. Currently, the most popular nowcast model-building approach is to use the factor bridge equation (BE) model or factor mixed data sampling (MIDAS) model, where the factors are extracted from large datasets using the principal component analysis (PCA) method. However, the PCA method only suits datasets approximating a hyperplane, while real-world datasets are sometimes unsuitable. This study proposes a kernel-based dimension reduction method. It is a natural extension of the PCA method and is called KTPCA. The KTPCA method can reduce the dimensionality of datasets without approximating a hyperplane. The dimension reduction performance of the iterative KTPCA method was superior to that of the PCA, Sparse PCA (SPCA), Randomized SPCA (RSPCA), and Robust SPCA (ROBSPCA) methods. In addition, the PCA and SPCA methods are competitive regarding dimension reduction performance. This study also proposes a nowcasting procedure using the iterative KTPCA method. This procedure can be applied to build nowcast models and update the forecasts of a target variable at a low frequency under real-time data flow of the original variables at a higher frequency.