Enhancing Predictive Accuracy with Hybrid Statistical–Deep Learning Models for Multivariate Time Series Forecasting: A Systematic Review with Taxonomy and Decision Framework
Precision Journal of Applied Mathematics and Statistics · 03.11.2026 · 1 · 1
DOI: 10.67887/pd2601208010 ↗Kurz gefasst
Systematische Übersicht in der Zeitschrift Precision Journal of Applied Mathematics and Statistics, online erschienen am 03.11.2026. Verfasst von Onyango Ronald. Die Arbeit nennt 100 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Introduction
Multivariate time series forecasting has become essential for decision-making in complex domains, including energy, healthcare, finance, and environmental monitoring. Traditional statistical models offer interpretability and efficiency but struggle with nonlinear dependencies, whereas deep learning models capture complex patterns yet often require large datasets and lack transparency. This systematic literature review addresses these limitations by examining the role of hybrid statistical–deep learning models in improving multivariate forecasting performance. Methodology: Following the PRISMA 2020 framework, studies published between 2020 and 2025 were identified from ScienceDirect, IEEE Xplore, Springer Nature Link, ACM Digital Library, and Taylor & Francis. From 2,526 identified records, 1,371 records were screened, 376 advanced to detailed eligibility review, 58 full-text reports were assessed, and 44 studies were included in the final synthesis. The selected studies were analysed according to hybrid architecture, forecasting horizon, dataset characteristics, performance outcomes, and methodological limitations.
Results
The review identified four major hybrid categories: residual/error-correction, feature-level, ensemble, and optimisation hybrids. Findings indicated that hybrid models generally outperformed standalone statistical and deep learning approaches, with reported improvements ranging from approximately 5% to over 40% depending on application context, forecasting horizon, and evaluation methodology. Residual-based approaches such as ARIMA–LSTM and VAR–LSTM were effective for combining linear pattern extraction with nonlinear learning, while decomposition and Transformer-based hybrids demonstrated advantages in complex long-term forecasting. However, challenges related to computational cost, interpretability, uncertainty modelling, and deployment feasibility remained.
Conclusion
This review contributes a structured taxonomy, decision-oriented framework, and analysis of failure conditions to support evidence-based selection of hybrid forecasting models.
Zitieren
APA
Ronald, O. (2026). Enhancing Predictive Accuracy with Hybrid Statistical–Deep Learning Models for Multivariate Time Series Forecasting: A Systematic Review with Taxonomy and Decision Framework. Precision Journal of Applied Mathematics and Statistics, 1, 1. https://doi.org/10.67887/pd2601208010
Vancouver
Ronald O. Enhancing Predictive Accuracy with Hybrid Statistical–Deep Learning Models for Multivariate Time Series Forecasting: A Systematic Review with Taxonomy and Decision Framework. Precision Journal of Applied Mathematics and Statistics. 2026;1:1. doi:10.67887/pd2601208010
BibTeX
@article{Ronald2026,
author = {Ronald, Onyango},
title = {Enhancing Predictive Accuracy with Hybrid Statistical–Deep Learning Models for Multivariate Time Series Forecasting: A Systematic Review with Taxonomy and Decision Framework},
journal = {Precision Journal of Applied Mathematics and Statistics},
year = {2026},
volume = {1},
pages = {1},
doi = {10.67887/pd2601208010}
}
EUVIMED