Specification Tests for the Variance of a Diffusion (Q4939809)

The authors consider a diffusion process \(X(t)\) observed at the points \(t=j/n\), \(j=1\),\dots,\(n\). The process can be generated by different kinds of data generating processes (DGP): (DGP2.1) -- homoscedastic model: \(dX_t=b(X_t)dt+\sigma_0 dW_t\); nonparametric heteroscedastic models: (DGP2.2) -- \(dX_t=b(X_t)dt+\sigma(t) dW_t\), (DGP2.3) -- \(dX_t=b(X_t)dt+\sigma(X_t) dW_t\), where \(W_t\) is a standard Brownian motion and \(b\) is an unknown function. In the nonparametric models the function \(\sigma\) is completely unknown. The parametric heteroscedastic models DGP3.2 and DGP3.3 are equivalent to DGP2.2 and DGP2.3, respectively, but it is supposed that \(\sigma(z)=c(z,\vartheta)\) where \(c\) is some known function and \(\vartheta\) is an unknown parameter. The paper is devoted to hypotheses testing: DGP2.1 vs DGP2.2 or DGP2.3 and DGP2.i vs DGP3.i. Nonparametric estimates for \(\sigma^2\) are used in the first case: \[ \hat s_n(\tau)=(2nh_n)^{-1}\sum_{i=1}^{n-1}1_{|\tau-i/n|<h_n} n(X_{(i+1)/n}-X_{i/n})^2 \] for DGP2.2 and \[ {\hat S}_n^2(x) = \sum_{i=1}^{n-1}1_{|X_{i/n}-x|<h_n} n(X_{(i+1)/n}-X_{i/n})^2\left(\sum_{i=1}^{n-1}1_{|X_{i/n}-x|<h_n}\right)^{-1} \] for DGP2.3. Here \(h_n\to 0\) is a smoothing parameter (bandwidth). Quasi maximum likelihood minimum contrast estimates \(\hat\vartheta_n\) for \(\vartheta\) are used for DGP3.2 and DGP3.3. The test statistics are based on Mahalonobis-like distances between the estimates of the tested models. E.g., for testing DGP2.3 vs DGP3.3 the authors consider \[ M_n=nh_n\sum_{k=1}^m(2c^2(\tau_k,\hat\vartheta_n))^{-1}(\hat s_n^2(\tau_k)-c^2(\tau_k,\hat\vartheta_n))^2. \] It is shown that under some conditions \(M_n\to_{w}\chi_m^2\) if DGP3.3 holds, and \(\Pr(M_n>C_n)\to 1\) as \(n\to \infty\) if \(C_n/(nh_n)^{1-\alpha}\to\infty\) if DGP2.2 holds and \(\sigma^2(\tau_k)\not=c^2(\tau_k,\vartheta^*)\) for some \(k=1\),\dots \(m\), \(\vartheta^*\) is the limit of \(\hat\vartheta_n\) as \(n\to\infty\). Analogous results are derived for the other variants of hypotheses testing.