IMPACT OF THE ANOMALOUS OBSERVATIONS ON THE FORECAST ACCURACY OF BAYESIAN VECTOR AUTOREGRESSION MODELS

Abstract

This study examines the impact of outliers on forecasting performance in Bayesian VAR (BVAR) models, with a focus on identifying robust methodologies under varying outlier magnitudes and sample sizes. Through an extensive simulation framework, we evaluate four BVAR models (BVAR1–BVAR4) under small, medium, and large outlier conditions across sample sizes ranging from n = 16 to n = 1000, using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) as accuracy criteria. Results reveal that outlier severity significantly degrades forecasting precision, with errors escalating as outliers grow larger. However, larger datasets (n ≥ 500) consistently mitigate these effects across all models. BVAR4 emerges as the most resilient to outliers in large-sample regimes, achieving the lowest errors (e.g., RMSE = 1.46 for small outliers, 1.48 for medium, and 1.54 for large at n = 1000), attributable to its structural complexity and adaptive priors. Simpler models like BVAR2 perform competitively for moderate samples (n = 50–500), while small datasets (n ≤ 50) favor fewer complex specifications to avoid overfitting. Notably, outlier magnitude imposes a persistent accuracy "floor," even with ample data. The findings underscore the critical role of aligning model complexity with data availability and outlier characteristics. We recommend BVAR4 for large-scale applications with severe outliers, hybrid approaches for moderate data, and outlier-aware preprocessing to enhance robustness. This study provides actionable insights for econometricians and practitioners in selecting models that balance accuracy, computational efficiency, and resilience to data anomalies.