Forecasting Extreme Weather Events Using Statistical Methods Up to 6 months in Vietnam

Abstract

This study evaluates the forecasting performance of three statistical methods Canonical Correlation Analysis (CCA), Principal Component Regression (PCR), and Multivariate Linear Regression (MLR) in predicting extreme weather events in Vietnam up to six months in advance. Among the methods assessed, PCR consistently outperforms both CCA and MLR across various climatic regions and event durations (i.e., 2–4 days and over 5 days), particularly in forecasting cold waves and moderate to heavy rainfall during the October–December period. By transforming correlated predictors into orthogonal components, PCR effectively addresses issues of multicollinearity, facilitates dimensionality reduction, and captures dominant climate signals with greater precision. Notably, PCR demonstrates superior adaptability to multiple climate models (e.g., CanCM4i, GFDL-SPEAR, CanSIPSv2) and maintains stable predictive accuracy without heavy reliance on post-processing techniques such as quantile mapping an advantage not observed with CCA and MLR. A key contribution of this research is the integration of PCR into a long-range climate forecasting framework tailored for Vietnam, enabling more robust and reliable predictions of extreme events across diverse climatic zones. Furthermore, this study introduces a seasonal drought forecasting system that produces probabilistic maps of average conditions, extreme climate indices, and drought occurrences up to six months in advance. By addressing the limitations of conventional statistical approaches and enhancing prediction accuracy, this research offers critical insights for disaster risk management and climate adaptation planning in Vietnam.