On a test for a parametric form of volatility in continuous time financial models (Q1776002)

Itô's diffusions are commonly used for representing asset prices. In general, the diffusion \(X_t\) is a solution to the stochastic differential equation \[ dX_t=b(t,X_t)\,dt +\sigma(t,X_t)\,dW_t, \] where \(W_t\) is a standard Brownian motion. An appropriate pricing of derivative assets requires a correct specification of the functional form of the drift and variance. Parametric models are attractive among practitioners because they admit a direct interpretation of the observed effects in terms of the parameters. However, the economic theory typically does not give much information about the drift and variance and misspecification of the model may lead to serious errors in the data analysis. If the assumption of a parametric model can not be justified, nonparametric estimates for the drift and variance of the diffusion should be used. The authors of the present paper develop a test for a specific parametric form of the variance function in a diffusion model of the above form. As a special case this procedure also gives a new simple test for homoscedasticity. Assume that we observe the diffusion on the interval \([0,1]\) at discrete points \(t_i=i/n,i=1,\dots,n,\) and we are interested in testing the hypothesis that the variance function in the stochastic differential equation is of the form \[ H_0:\sigma^2(t,x)=\sum_{j=1}^d\alpha_j\sigma^2_j(t),\quad t\in[0,1],\;x\in\mathbb R, \] for some constants \(\alpha_1,\dots,\alpha_d\in\mathbb R\), where \(\sigma^2_1(t),\dots,\sigma^2_d(t)\) are given nonnegative linearly independent functions. For the construction of appropriate test statistics let us define the pseudo residuals \[ \Delta_i=n(X_{(i+1)/n}-X_{i/n})^2,\quad i=1,\dots,n-1; \quad \Delta=(\Delta_1,\dots,\Delta_{n-1})^T, \quad \alpha=(\alpha_1,\dots,\alpha_d)^T, \] and the design matrix \[ X=(\sigma^2_j({i}/{n}))^{j=1,\dots,d}_{i=1,\dots,n-1}\in \mathbb R^{(n-1)\times d}. \] Consider the least squares problem \[ \widehat{\alpha}=\arg\min_{\alpha\in\mathbb R^d} (\Delta-X\alpha)^T(\Delta-X\alpha)= (X^TX)^{-1}X^T\Delta, \] and define a test statistics for the hypothesis \(H_0\) by \[ T_n= n^{-1} \{3^{-1}\Delta^T\Delta- \Delta^TX(X^TX)^{-1}X^T\Delta\}. \] The authors show that, under some conditions, \(T_n-M^2=O_p(n^{-1/2}\log n),\) where the random variable \(M^2\) is defined by \[ M^2=\min_{\alpha_1,\dots,\alpha_d\in\mathbb R} \int_0^1\left\{\sigma^2(t,X_t)-\sum_{j=1}^d\alpha_j\sigma^2_j(t)\right\}^2\,dt. \] The second main result states that the test statistic \(T_n\) has an asymptotic normal distribution under the null hypothesis \(H_0\) and diverges at an appropriate rate under the alternative. The results are illustrated by a small simulation study, and a data example is analyzed.