Evaluation of the bvars Package for Multivariate Time Series Analysis
Over the past three days, I evaluated the bvars package (Liu, Ramirez Hassan, & Woźniak) for multivariate autoregressive modeling in R. The package is available on CRAN with DOI: 10.32614/CRAN.package.bvars.
Dataset Characteristics:
The evaluation employed 280 multivariate time series datasets, each comprising nine univariate series. Series length varied between 90 and 450 observations. The data consisted of operational metrics from 2022–2024, including purchase orders, schedule lines, material movements (inbound/outbound), sales orders (requested and confirmed quantities), deliveries, and inventory levels across multiple materials. Given their operational origin, these series exhibited substantial stochastic variation, making them appropriate for assessing time series analysis and forecasting methods.
Comparative Performance:
Within each dataset, the nine series exhibited significant lagged cross-correlations. The standard vars package failed to converge or produce valid estimates for approximately 67% of datasets. In contrast, bvars successfully converged for 279 of 280 datasets (99.6% convergence rate). All analyses employed the default configuration of 10,000 MCMC draws, which proved adequate for these datasets. Posterior diagnostics, model validation tests, and forecast assessments all indicated satisfactory model specification and estimation quality.
Conclusion:
These results suggest that bvars offers substantial advantages over traditional VAR methods for highly correlated multivariate time series with substantial noise. I recommend this package for practitioners working with similar operational or business datasets.
I gratefully acknowledge Rui Liu, Andrés Ramirez Hassan, and Tomasz Woźniak for developing this valuable tool.
#RStats #bvars #bsvars
