TIME SERIES CLUSTERING IN MULTIVARIATE PRICES MODELING

Time series data are crucial in agricultural price analysis, with the Vector Auto-Regressive (VAR) and Vector Error Correction Model (VECM) being essential tools. VECM is necessary for cointegrated series to capture short-term and long-term dynamics. However, the increasing availability of disaggregated agricultural commodity price data over the past three decades has resulted in high-dimensional datasets, challenging the efficacy of VECM and Johansen’s maximum likelihood test. This thesis tackles this issue by using time series clustering to reduce data dimensionality. Clusters are dynamically formed based on price similarity, allowing Johansen’s framework to estimate cointegration relationships. Applied to the Chinese hog market before and after the 2018 African Swine Fever outbreak, this method reveals clusters aligned with industry patterns. Using hierarchical clustering with dynamic time warping, this approach reduces dimensionality, recovers cointegrating relationships, and offers insights into potential trade patterns, showing the benefits over traditional geographical clustering.