Approach to Testing Nonlinear Time Series Models
Yongmiao Hong,Yoon-Jin Lee
2014-01-01
Abstract:This paper proposes a uni
ed approach to testing adequacy of nonlinear time series models. The proposed test can be applied to various nonlinear time series models, including conditional probability distribution models, Markov chain regime-switching models, conditional duration models, conditional intensity models, continuous-time jump di¤usion models, continuous-time regression models, and conditional quantile and interval models. Our approach is built upon the fact that for many nonlinear time series models, model adequacy usually implies that a suitably transformed process is an independent and identically distributed (i.i.d.) sequence with some speci
ed marginal distribution. Examples include the probability integral transform of an autoregressive conditional distribution model, the integrated hazard function of a conditional duration or intensity model, the time-change transform of a continuous-time regression model, and the binary transformation of an autoregressive conditional quantile or interval model. These transforms are, respectively, i.i.d.U[0,1], i.i.d.EXP(1); i.i.d. N(0,1) and i.i.d. Bernoulli( ) for some known 2 (0; 1) when the time series models are correctly speci
ed. The transformed process may be called the generalized residuals of a time series model since they are generalizations of Cox and Snells (1968) concept of generalized residuals to a time series context. The proposed test checks the joint hypothesis of generalized residuals via a frequency domain approach and has omnibus power against a wide range of model misspeci
cations. It has a convenient null asymptotic N(0,1) distribution and is robust to dependent persistence in the underlying time series process. A Monte Carlo simulation study illustrates the merits of the approach.