This paper focuses on finding starting-values for maximum likelihood estimation of Vector STAR models. Based on a Monte Carlo exercise, different procedures are evaluated. Their performance is assessed w.r.t. model fit and computational effort. I employ i) grid search algorithms, and ii) heuristic optimization procedures, namely, differential evolution, threshold accepting, and simulated annealing. In the equation-by-equation starting-value search approach the procedures achieve equally good results. Unless the errors are cross-correlated, equation-by-equation search followed by a derivative-based algorithm can handle such an optimization problem sufficiently well. This result holds also for higher-dimensional VSTAR models with a slight edge for the heuristic methods. Being faced with more complex Vector STAR models for which a multivariate search approach is required, simulated annealing and differential evolution outperform threshold accepting and the grid with a zoom.
Vector STAR model starting-values optimization heuristics grid search estimation non-linearieties