Chen-Choi-Zhou (2005) Nonparametric Estimation of Structural Change Points in Volatility

Summary

A frequentist nonparametric approach to detecting and estimating structural change points in the conditional volatility of a time series. The model is $Y_t = \mu(X_t) + \sigma(X_t)\varepsilon_t$ where both the mean $\mu(\cdot)$ and variance $\sigma^2(\cdot)$ functions are left unspecified; the change-point location $k_0$ is estimated without imposing any parametric form on marginal or transitional densities. The hybrid least-squares/kernel estimator achieves the parametric rate $O(T^{-1})$ for the change-point location while the variance function is estimated at standard nonparametric rates.

Key Claims

• Model. The data-generating process (DGP) is $Y_t = \mu(X_t) + \sigma_t(X_t)\varepsilon_t$ where $\sigma_t(\cdot) = \sigma_1(\cdot)$ for $t \leq k_0$ and $\sigma_t(\cdot) = \sigma_2(\cdot)$ for $t > k_0$; $k_0 = [T\theta_0]$ for some $\theta_0 \in (0,1)$ unknown; $\varepsilon_t$ i.i.d. with $E[\varepsilon_t|X_t] = 0$, $E[\varepsilon_t^2|X_t] = 1$. No parametric forms assumed for $\mu$, $\sigma_1$, $\sigma_2$.
• Hybrid estimator. (1) Estimate the mean function $\mu(\cdot)$ by kernel regression on the full sample; (2) form squared residuals $\hat{u}_t^2 = (Y_t - \hat{\mu}(X_t))^2$ as a proxy for $\sigma_t^2(X_t)$; (3) estimate $k_0$ by least-squares over $\hat{u}_t^2$, exploiting the jump in the variance function at the change point.
• Parametric rate for the change-point. Despite the fully nonparametric setting, the change-point estimator $\hat{k}$ satisfies $\hat{k} - k_0 = O_p(1)$, i.e. it converges at rate $O(T^{-1})$ — the same as maximum likelihood estimation (MLE) under a correctly specified parametric model. The nonparametric estimation of $\mu$ does not degrade the change-point convergence rate.
• Superiority over one-sided kernel tests. Existing one-sided kernel procedures have weak power and converge at nonparametric rates $O(T^{-2/5})$ for the change-point location. The least-squares component in the hybrid procedure restores the parametric rate.
• No distributional assumption needed. Unlike MLE-based change-point estimators (Bai 1994, Bai and Perron 1998), the procedure requires no explicit likelihood, making it robust to fat tails, skewness, or GARCH-type conditional variance dynamics that invalidate standard distributional assumptions.
• S&P 500 application. Applied to daily S&P 500 returns; detects significant change points in conditional variance in the early 1980s and around the 1987 crash, consistent with GARCH-based structural break findings.

Concepts Introduced or Extended

• Change-Point Analysis
• GARCH and BEKK-GARCH

Entities Mentioned

• (none with existing wiki pages)

Quotes

"The main advantage [of our approach] is that it does not require any specific form of marginal or transitional densities of the process."

My Take

A clean technical contribution to nonparametric econometrics. The key insight is that the least-squares step for the change-point location is semi-parametric: even though the variance function is nonparametric, the change-point itself acts like a finite-dimensional parameter and can be estimated at the parametric $O(T^{-1})$ rate. This is the nonparametric analogue of Bai-Perron for mean shifts. The S&P 500 application is illustrative rather than novel; the theoretical contribution (rates + asymptotics) is the core value. Less relevant to the Bayesian focus of this wiki, but provides a frequentist baseline for change-point detection in volatility alongside the Bayesian reversible-jump MCMC (RJMCMC) approach in Green (1995).