Modelling Structural Breaks, Long Memory and Stock Market Volatility: An Overview

Summary

Editorial overview for the Journal of Econometrics special issue vol. 129 (2005), arising from the December 2002 "Long Memory, Structural Breaks and Stock Market Volatility" conference at Cass Business School, London. Surveys three interconnected literatures: structural break testing (Perron 1989 through Bai-Perron 1998 and Hansen 2000), long memory and ARFIMA (autoregressive fractionally integrated moving average) processes (Granger-Joyeux 1980, Hosking 1981, GPH (Geweke-Porter-Hudak) estimator, local Whittle, R/S statistics), and their interface (how ignoring breaks generates spurious long memory / IGARCH (integrated GARCH), and how regime switching mimics genuine long memory).

Key Claims

• Nelson-Plosser (1982): cannot reject unit root in 13/14 US macro series under standard Dickey-Fuller (DF) tests.
• Perron (1989): allowing one structural break in the deterministic trend restores stationarity in 11/14 Nelson-Plosser series; break date treated as known exogenously.
• Zivot-Andrews (1992): break date must be treated as endogenous; use $\inf t(\lambda)$ statistic minimized over $\lambda \in \Lambda$.
• Bai-Perron (1998): multiple-break testing via $\sup F$ (fixed $k$ breaks), UDmax (unknown $k \leq M$, equal weights), WDmax (equalizes marginal p-values across k); break dates estimated by global sum-of-squared-residuals (SSR) minimization.
• Bai (1999) sequential likelihood-ratio (LR) test: null of $l$ breaks vs. $l+1$; limiting density is known in closed form.
• Hansen (2000): fixed-regressor bootstrap controls size even when regressors themselves undergo structural change; handles systems with non-stationary regressors.
• ARFIMA(p,d,q): $\phi(L)(1-L)^d(y_t-\mu) = \theta(L)\varepsilon_t$; $d = H - \tfrac{1}{2}$ (with $H$ the Hurst exponent); stationary for $|d| < \tfrac{1}{2}$.
• GPH estimator (Geweke-Porter-Hudak 1983): log-periodogram regression; $\hat{d}$ asymptotically normal; $m = \sqrt{T}$ rule of thumb for bandwidth.
• Robinson (1995b) local Whittle: frequency-domain Gaussian likelihood averaged near zero; more efficient than GPH.
• R/S statistic (Hurst 1951): range of partial sums / std dev; Lo (1991) modified R/S (MR/S) corrects for short memory and heteroscedasticity.
• Fractional cointegration (Granger 1980): $z_t = y_t - \beta x_t$ is $I(d-b)$ with $0 < b < d$.
• Hillebrand (2004, in vol.): unaccounted GARCH regime changes → $\alpha+\beta \to 1$; spurious IGARCH is the second-moment analogue of Perron's spurious unit root from ignored breaks.
• Diebold-Inoue (2001): stochastic regime switching (Markov, random-level-shift, permanent-break model) satisfies the variance-of-partial-sums definition of long memory — regimes mimic genuine fractional integration.

Concepts Introduced or Extended

• Structural Break Testing
• Long Memory and Fractional Integration
• GARCH and BEKK-GARCH
• Cointegration

Entities Mentioned

• Anindya Banerjee
• Giovanni Urga
• Clive Granger
• Robert Engle

Quotes

"Unaccounted-for parametric regime changes in GARCH models cause the sum of the autoregressive parameters to converge to one, leading to a finding of spurious persistence." (p. 11, summarizing Hillebrand 2004)

"Stochastic regime switching is easily confused with long memory." (p. 23, summarizing Diebold-Inoue 2001)

My Take

This is a useful map of two literatures and their intersection, not an original research contribution. The most durable insight is the symmetry: just as ignoring structural breaks in the mean inflates unit-root evidence, ignoring GARCH regime changes inflates variance-persistence evidence. The Bai-Perron multiple-break framework and the GPH/local Whittle semi-parametric estimators are the workhorses practitioners should know. The editorial does not distinguish clearly between editorial summaries and the authors' own views, so claims about individual papers should be verified in the originals.