TL;DR
This paper introduces a novel statistical test to detect structural changes in the autocorrelation of AR(1) time series models. By leveraging wavelet transformations for frequency decomposition, the new test offers improved power compared to traditional CUSUM tests, particularly in identifying shifts in serial dependence.
What Problem Does This Paper Address?
Identifying structural breaks, or sudden changes in the underlying data-generating process, is crucial in many fields, including economics and finance. While methods exist for detecting changes in means or variances, detecting shifts in the autocorrelation structure of time series, especially in autoregressive (AR) models, remains a challenge. Traditional methods often struggle with serially dependent data or require known change points. This paper aims to fill this gap by developing a robust test for detecting unknown structural changes in the autocorrelation coefficient of an AR(1) process.
Methods
The core of this paper's contribution lies in its innovative use of wavelet transformations. Specifically, the authors employ the Maximal Overlap Discrete Wavelet Transformation (MODWT) to decompose the variance of the time series into different frequency bands. The rationale is that the spectral properties (how variance is distributed across frequencies) of an AR(1) process are directly linked to its autocorrelation coefficient. By analyzing the variance contributions in specific frequency bands, particularly the low-frequency band, the test can detect changes in the autocorrelation.
The proposed test statistic is a cumulative sum of squares (CUSUM) of these wavelet-derived scaling coefficients. Under the null hypothesis of no structural change, this statistic is expected to oscillate around zero. However, a structural change in the autocorrelation will cause the statistic to deviate significantly from zero, signaling a break. The paper also details how to construct a feasible version of this test by estimating model parameters and using Monte Carlo simulations to derive critical values.
Datasets
The paper uses simulated AR(1) processes for Monte Carlo simulations to evaluate the test's size and power properties. For empirical application, it utilizes seasonally-adjusted monthly Consumer Price Index (CPI) data for the United States, obtained from the FRED Economic Data Website. This data spans from December 1964 to December 2010.
Key Insights and Ideas
- Wavelets as a Lens for Autocorrelation: The paper elegantly demonstrates how wavelets can decompose time series variance in a way that directly reveals information about the autocorrelation structure. Different autocorrelation coefficients manifest as distinct patterns in the frequency domain, which wavelets can effectively capture.
- Low-Frequency Focus: The test specifically targets changes in the low-frequency components of the variance, as these are most sensitive to shifts in the AR(1) coefficient. This targeted approach enhances the test's ability to detect subtle autocorrelation changes.
- CUSUM of Wavelet Coefficients: Adapting the well-established CUSUM framework to wavelet-derived coefficients provides a powerful new tool for structural change detection in time series with serial dependence.
- Empirical Superiority: The proposed wavelet-based test shows promising results, outperforming traditional CUSUM tests in terms of power, suggesting it's more adept at detecting actual structural breaks.
Key Results
The Monte Carlo simulations indicate that the proposed wavelet-based test exhibits reasonable empirical sizes (close to nominal levels) and superior power compared to CUSUM tests based on recursive and OLS residuals. This suggests the test is effective in detecting structural changes in the autocorrelation of AR(1) processes.
When applied to US inflation rates, the test found evidence of a structural change in the autocorrelation structure during the period of 1965-1979, potentially linked to the shift to flexible exchange rate regimes. For the more recent period of 2001-2010, the test did not reject the null hypothesis of no structural change, even during the 2007-2008 financial crisis, implying that while the crisis affected inflation levels, it did not alter the underlying autocorrelation pattern.
Limitations and Future Work
The authors acknowledge several avenues for future research:
- Asymptotic Distribution: Currently, critical values are derived through computationally intensive Monte Carlo simulations. Deriving the asymptotic distribution of the test statistic would offer a more efficient approach.
- Multiple Changes and Location: The current test is designed for detecting a single, unknown change point. Extending it to identify multiple changes and precisely locate them would be a valuable advancement.
- Directional Changes: The test currently detects any deviation from the null. Further work could explore distinguishing between increases and decreases in the autocorrelation coefficient.
- Generalization: The study is focused on AR(1) models. Extending the methodology to more general data-generating processes or regression settings would broaden its applicability.
- Wavelet Choice: The performance of the test can be influenced by the choice of wavelet filter and decomposition level. Investigating optimal choices for different scenarios is suggested.
Summary updated