Nonparametric method of structural break detection in stochastic time series regression model

Technical Summary: Nonparametric Method of Structural Break Detection in Stochastic Time Series Regression Model

Problem Statement
The paper addresses the challenge of detecting structural breaks in stochastic time series regression models, specifically focusing on changes in the conditional mean and/or conditional variance functions. Existing literature often relies on parametric assumptions regarding the data-generating process (e.g., linear regression with specific error structures like ARMA or GARCH) or assumes independence/mixing conditions that may not hold in complex real-world scenarios. Furthermore, many nonparametric approaches focus on marginal distributional changes or spectral features rather than the conditional regression functionals themselves. The authors aim to develop a flexible framework that does not impose specific parametric forms on the dependence structure of the regressor, the time series model, or the noise distribution, while accommodating both short-range and long-range dependence in the covariate process.

Methodology
The proposed methodology is built upon a stochastic regression model:
$$Y_t = \mu(X_t) + \sigma(X_t)\epsilon_t$$
where $\{Y_t\}$ is the response, $\{X_t\}$ is the covariate, $\mu(\cdot)$ is the conditional mean, $\sigma^2(\cdot)$ is the conditional variance, and $\{\epsilon_t\}$ is an i.i.d. noise process with unit variance.

1. Estimation Framework:
   The authors utilize Nadaraya-Watson type kernel estimators to estimate the difference between the conditional functions in two disjoint segments of the time series ($T^-$ and $T^+$). To handle bias inherent in kernel estimation, they employ a jackknife-type correction using a modified kernel function $K^*(u) = 2K(u) - K(u/\sqrt{2})/\sqrt{2}$. This allows for the construction of bias-corrected estimates for both the difference in means ($\hat{\mu}^*_{diff}$) and the difference in variances ($\hat{\sigma}^{*2}_{diff}$).

2. Dependence Handling:
   A key theoretical component is the treatment of temporal dependence in the covariate $\{X_t\}$. The authors assume $\{X_t\}$ admits a causal representation driven by i.i.d. innovations. They utilize martingale projection operators to quantify dependence, distinguishing between short-range dependent (SRD) and long-range dependent (LRD) processes via a quantity $\Xi_n$. The theoretical results hold for both cases, provided specific bandwidth conditions are met.

3. Hypothesis Testing:
   The core of the method involves testing the null hypothesis that the conditional function (mean or variance) remains constant across the time horizon against the alternative that a structural break exists.

   • Test Statistics: The test statistics are based on the supremum of the absolute difference between the estimated functions over a dense grid of points in the covariate space: $\sup_{x \in \Pi_n} |\hat{\mu}^*_{diff}(x)|$ for the mean and $\sup_{x \in \Pi_n} |\hat{\sigma}^{*2}_{diff}(x)|$ for the variance.
   • Asymptotic Distribution: Under the null hypothesis, the authors derive the asymptotic distribution of these supremum statistics. They show that, after appropriate normalization, the statistics converge to a Gumbel distribution. This allows for the construction of critical values for hypothesis testing without requiring parametric assumptions about the noise distribution.
   • Simultaneous Testing: To detect breaks in both mean and variance simultaneously, the authors propose a procedure using the maximum of the individual test statistics combined with a Holm-Bonferroni correction.

4. Localization Algorithm:
   While the primary theory assumes an "at-most-one-change" (AMOC) scenario, the paper extends the method to multiple changepoints using a binary segmentation algorithm. This involves recursively applying the AMOC test to sub-segments. To refine the localization and control false positives, a post-detection validation step re-tests detected breakpoints using local sub-data to confirm significance. The consistency of the estimated break locations is established theoretically.

Key Contributions and Results

• Theoretical Guarantees: The paper provides rigorous asymptotic theory showing that the proposed estimators are consistent and that the test statistics converge to a known distribution (Gumbel) under the null hypothesis. The consistency of the estimated changepoint locations is also proven.
• Flexibility: The method does not require the user to specify the functional form of $\mu(\cdot)$ or $\sigma^2(\cdot)$, nor does it assume a specific distribution for the noise $\epsilon_t$, other than moment conditions (finite second moments for mean testing, finite fourth moments for variance testing). It accommodates heavy-tailed distributions for the observed series $Y_t$ even if the standardized noise has lighter tails, as the conditional variance function absorbs the tail behavior.
• Simulation Performance: Extensive simulations across various data generating processes (including White Noise, ARMA-GARCH, and Threshold AR models) and noise distributions (Normal, Student's t, and Power Law) demonstrate that the test maintains good size control (Type I error rates close to nominal levels). The power of the test increases with sample size and is notably effective for detecting breaks in conditional variance, often outperforming mean-break detection in complex models. The method shows robustness to heavy-tailed innovations.
• Real-World Application: The method is applied to Bitcoin price data and Google News Interest Scores (GNIS). The analysis identifies two structural breaks in the conditional mean relationship between public attention and Bitcoin prices (March 2021 and July 2023), corresponding to periods of institutional interest and market stabilization, respectively. No structural breaks were detected in the conditional variance. The application illustrates the method's ability to identify shifts in the conditional relationship between market attention and price dynamics.

Significance and Claims
The authors claim that their approach offers a unified, nonparametric framework for detecting structural breaks in conditional regression functionals, distinguishing it from existing methods that rely on parametric assumptions or focus on marginal properties. By allowing for flexible temporal dependence (including long-range dependence) and avoiding strict distributional assumptions, the method is presented as particularly suitable for financial time series, which often exhibit heavy tails and complex dependence structures. The paper emphasizes that while the main theory is developed under the AMOC assumption, the sequential application of the test provides a practical and theoretically grounded solution for multiple changepoint detection. The authors note that the procedure is computationally efficient and superior in performance for heavy-tailed financial data compared to some existing alternatives. They conclude by suggesting that the theoretical framework could potentially be extended to higher-order conditional moments and high-dimensional settings in future work.