Koop-Potter (2006) Forecasting and Estimating Multiple Change-point Models with an Unknown Number of Change-points

Summary

The forecasting-focused companion to Koop-Potter (2004). It uses the same change-point model — unknown number of change-points, Poisson regime durations, approximately nesting the TVP and small-K change-point models — but develops two things more fully: a complete hierarchical prior for both regime durations and the regime-specific parameters, and, most distinctively, an efficient real-time forecasting scheme via sequential importance sampling (a particle filter) that updates predictive densities as new data arrive without re-running the MCMC. Applied to US GDP growth and inflation, it again finds more change-points than earlier work and a model resembling TVP-with-stochastic-volatility.

Key Claims

• Same model, fuller priors. Regime durations are Poisson; the paper builds "reasonable hierarchical priors both for regime durations and for the parameters which characterize each regime," so that durations and break sizes borrow strength across regimes rather than being estimated in isolation.
• Real-time forecasting without re-running MCMC. Writing the model in state-space form $p(y_t\mid z_t)$, $p(z_t\mid z_{t-1})$ with states $z_t=(\theta_t,S_t)$, a sequential importance sampler (SIS / particle filter) recursively updates the importance weights as each new observation arrives: $w_{T+1}=p(Z_T\mid Y_T)\,p(y_{T+1}\mid z_{T+1})\,p(z_{T+1}\mid z_T)/q(\cdot)$, and $w_{T+2}=w_{T+1}\times[\,p(y_{T+2}\mid z_{T+2})p(z_{T+2}\mid z_{T+1})/q(\cdot)\,]$ — avoiding a full re-estimation at each forecast origin.
• Negative-Binomial tail handling. When forecasting $h$ periods ahead the number of future regimes can exceed the in-sample maximum; the Negative-Binomial distribution implied by the duration hyperparameters supplies the extra regime-duration draws.
• Why the blend matters for forecasting. Small-K models assume no break after the last in-sample one; TVP models force a break every period of limited size. Only a model that admits infrequent-but-large future breaks of history-dependent size gives honest predictive densities — the whole point of the exercise.
• Empirics. On US GDP growth and inflation the techniques "work well"; the estimated number of change-points is larger than previously found and the implied model is close to a TVP model with stochastic volatility.

Concepts Introduced or Extended

• Change-Point Analysis

Entities Mentioned

• Gary Koop
• Simon M. Potter

Quotes

"We show how real time forecasting can be done in an efficient manner using sequential importance sampling."

"It is desirable to update forecasts without re-running the MCMC algorithm. ... sequential importance sampling methods [these are a popular type of particle filter methods] can be used with our model to achieve this goal."

My Take

Where the 2004 draft made the modeling case (Poisson vs. Geometric durations, unknown number of breaks), this version makes the operational case: a change-point model is only useful for policy if you can forecast with it in real time, and the sequential-importance-sampling recursion is the piece that turns an expensive one-shot MCMC posterior into something you can roll forward observation-by-observation. That framing — MCMC for the historical fit, a particle filter for the live forecast update — is the durable practical lesson and is why this pair belongs next to the state-space/particle-filter material in the wiki as much as next to the change-point material. The honest caveat is the usual SIS one (weight degeneracy over long horizons, here partly mitigated by the Negative-Binomial regime-count tail), which the paper handles at the level needed for macro forecast horizons but which is the natural place the method strains. Together with the 2004 draft this is the working-paper basis for the published Koop-Potter (2007, Review of Economic Studies), "Estimation and Forecasting in Models with Multiple Breaks."