This paper is intended to test the robustness of the fitness of nested GARCH models.
Both Monte‐Carlo simulation data and real‐world data are used in the paper. Likelihood‐family tests are used to test in‐sample fitness, while mean‐squared prediction error is employed for out‐sample prediction tests.
The paper finds that, generally, the parsimonious principle is found to work well for both criteria. However, it is found that conflict exists between the two criteria: in‐sample likelihood‐family tests pay more attention to conditional distributions or are more sensitive to fat tail effects; while the out‐sample criteria focus more on the accuracy of parameter estimation.
The paper shows that complexity does not necessarily mean good fitness; sometimes, the simpler model can fit better, especially for real‐world data.
