Power transformation and threshold modeling for ARCH innovations with applications to tests for ARCH structure. (Q2574642): Difference between revisions

Let \(\{e_t\}\) be iid random variables with \(Ee_t=0\), \(var\,e_t=1\), and with a marginal density \(f\). The power-transformed threshold ARCH model for a process \(\{\varepsilon _t\}\) is defined by \(\varepsilon _t=\sqrt {v_{t-1}}e_t\), \[ v_{t-1}^{\delta } = \alpha _0 +\alpha _{11}[(\varepsilon _{t-1}^+)^2]^{\delta } + \alpha _{21}[(\varepsilon _{t-1}^-)^2]^{\delta } + \cdots + \alpha _{1m}[(\varepsilon _{t-m}^+)^2]^{\delta } + \alpha _{2m}[(\varepsilon _{t-m}^-)^2]^{\delta }, \] where \(\delta >0\), \(\alpha _0>0\), \(\alpha _{ij}\geq 0\), and \(\varepsilon _{t-i}^+ = \max (\varepsilon _{t-i},0)\), \(\varepsilon _{t-i}^- = \max (-\varepsilon _{t-i},0)\). For \(\delta =1/2\), \(\delta =1\), \(\delta =2\), and \(\delta \to 0\) this model reduces to special cases investigated earlier in the literature. Geometric ergodicity of the model and existence of stationary moments are studied. Using a local asymptotic normality approach, large sample tests for (i) symmetric ARCH models, (ii) symmetric ARCH models with \(\delta =0\), and (iii) ARCH models with \(\delta =1\) (no power transformation) are derived. Semiparametric tests are discussed for the case that the density \(f\) is not known.