Koop-Potter (2004) Estimation of Multiple Change-point Models with an Unknown Number of Change-points

Summary

Koop and Potter develop a Bayesian change-point model in which the number of change-points is unknown and, crucially, the distribution of regime durations is not restricted to be Geometric. Building on Chib's (1998) reframing of change-points as durations of a hidden Markov chain, they replace the constant-hazard (Geometric duration) assumption with a Poisson duration distribution embedded in a hierarchical prior, so the model approximately nests the two dominant approaches — the time-varying-parameter (TVP) model (a break every period) and the small-fixed-number-of-breaks change-point model. Applied to US GDP and PCE-deflator inflation, it finds more change-points than earlier studies and behaves like a TVP-with-stochastic-volatility model with heterogeneous transition innovations.

Key Claims

• Change-points as hidden-Markov durations (Chib 1998). Locating change-points is equivalent to determining the durations of a Markov regime-indicator chain; Chib's (1996) forward-filtering/backward-sampling algorithm — valid even for non-time-homogeneous chains — draws the regime path.
• The Geometric-duration defect. Under Chib (1998) / Chernoff-Zacks, within-regime exit probability is constant, so regime duration is Geometric — a change is most likely in the regime's first period and monotonically less likely thereafter, an empirically unattractive restriction.
• Poisson-duration alternative. Koop-Potter model each regime's duration as a Poisson draw within a hierarchical structure, giving a non-monotone, more flexible duration distribution while keeping conditionally-conjugate priors to limit computational cost.
• Unknown, unbounded number of breaks. Not fixing the number of change-points avoids the awkward prior a fixed count requires — imposing exactly $K$ breaks forces a non-time-homogeneous Markov chain near the end of the sample, a prior with "unusual and potentially undesirable properties."
• End-of-sample forecasting matters. Fixed-break models assume no break after the last in-sample one; TVP models assume a break every period with size limited by the every-period assumption. Neither handles "infrequent but large" future breaks well — the motivation for a model where a future break's size can depend on the previous regime and on the full break history.
• Empirics. On US real GDP and PCE inflation the number of change-points is larger than previously estimated; the implied model resembles a TVP model with stochastic volatility and heterogeneous innovations in the transition equation.

Concepts Introduced or Extended

• Change-Point Analysis
• Time-Varying Parameter Model

Entities Mentioned

• Gary Koop
• Simon M. Potter

Quotes

"In the approach of Chib (1998), the problem of locating the change-points is converted into the problem of determining the duration of a Markov regime."

"Since (within a given regime) the probability of a change is constant ... the distribution of the regime duration is Geometric. Thus, the regime change is most likely to happen in the first period of the regime, then next most likely in the second period and so on."

My Take

The paper's leverage is a single well-chosen relaxation: Chib's hidden-Markov reframing of change-points is elegant but silently commits you to a Geometric duration (constant hazard), and Koop-Potter show that this is doing real empirical work — swap in a Poisson duration inside a hierarchical prior and both the number and the timing of breaks change. The framing that a change-point model and a TVP model are two ends of one spectrum (breaks every period vs. a handful) is the conceptual payoff, and it makes precise why the choice matters most for forecasting: only a model that admits future breaks of history-dependent size gives honest predictive uncertainty. The subtle technical contribution — that imposing a fixed number of breaks forces a non-time-homogeneous chain at the sample's end, requiring a strange prior — is the kind of thing that is easy to miss and genuinely useful to have flagged. This is the "estimation" half of the project; the forecasting emphasis is developed in the 2006 companion, and the two were later published together as Koop-Potter (2007, Review of Economic Studies).