Impact of anomalous observations on Vector Autoregressive and Bayesian Vector Autoregressive Accuracy

Outliers pose significant challenges to statistical modelling by distorting inferences and forecasts. This study examines the robustness of Vector Autoregression (VAR) and Bayesian VAR (BVAR) models in the presence of outliers, employing simulation-based analysis across varying sample sizes (small: 16–32, medium: 50–100, large: 500–1000) and outlier magnitudes (small, medium, large). Using root mean squared error (RMSE) and mean absolute error (MAE) as criteria, the non-Bayesian VAR model (VAR2) was compared against four Bayesian variants (BVAR1– BVAR4) with Sims-Zha priors. Results demonstrate VAR2’s superior resilience to outliers, consistently achieving lower forecast errors across all scenarios. Bayesian models, while improving with larger samples, lagged due to excessive shrinkage of outlier-driven signals. VAR2’s parsimonious structure avoided over-reliance on prior assumptions, proving particularly advantageous in smaller datasets (n<100), where BVARs exhibited higher sensitivity. Conversely, BVAR4 showed moderate improvement in large samples but never surpassed VAR2. The study concludes that in outlier-prone environments, VAR2 is preferable for its robustness and simplicity. Practitioners should reserve BVARs for contexts requiring Bayesian uncertainty quantification, ideally with tailored priors to mitigate outlier effects. Recommending VAR2 for most scenarios, with BVAR4 considered only when domain-specific priors justify added complexity. These findings highlight the trade-offs between model flexibility and robustness, guiding empirical choices in macroeconomic forecasting and policy analysis.