Skip to main content

Cookies on Macro by Mark.We use essential cookies for sign-in, locale, and security. With your permission, we also use optional cookies for preferences, for analytics, and for advertising measurement with the Meta (Facebook) Pixel. Read our cookie policy.

Macro by MarkMacro by Mark
  • Home
Loading…
Loading…
← Models
MacroConceptsPolicyModels

Smooth transition regression Model

Regression with gradual regime weights, allowing coefficients to transition smoothly as a threshold variable changes.

OverviewGraphMathReferences

Does the economy transition gradually between regimes, and can we estimate both the location and speed of that transition from the data?

Background

The Threshold VAR forces an abrupt switch at a single cutoff. In practice, monetary policy does not flip overnight, credit conditions do not jump from loose to tight at a single spread level, and consumer sentiment drifts rather than snaps. The Smooth Transition Regression (STR), formalized by Granger and Terasvirta (1993) and extended by Terasvirta (1994), replaces the indicator function with a bounded transition function (monotone logistic, or U-shaped exponential) that moves continuously between 0 and 1 as the transition variable moves relative to its location parameter.

The mechanism is a weighted average of two linear regressions. At every point in time, the model's effective coefficient vector is β(1)+G(st)β(2)\beta^{(1)} + G(s_t) \beta^{(2)}β(1)+G(st​)β(2) - equivalently the weighted average (1−G(st))β(1)+G(st)(β(1)+β(2))(1 - G(s_t)) \beta^{(1)} + G(s_t) (\beta^{(1)} + \beta^{(2)})(1−G(st​))β(1)+G(st​)(β(1)+β(2)) of the two regime coefficient vectors - where GGG is the transition function evaluated at the transition variable sts_tst​. When GGG is near 0 the model behaves like regime 1; when GGG is near 1 it behaves like regime 2; in between, both regimes contribute proportionally. Two parameters govern the transition: ccc (location - where the midpoint of the switch lies) and γ\gammaγ (speed - how quickly GGG moves from 0 to 1). The Threshold VAR is the limiting case as γ\gammaγ approaches ∞\infty∞.

Estimation is by nonlinear least squares (NLS) or conditional maximum likelihood. The transition speed γ\gammaγ is notoriously difficult to identify: the log-likelihood surface is often flat in the γ\gammaγ direction, and different starting values can produce very different estimates. Terasvirta (1994) proposed a specification cycle: (1) test linearity using a third-order Taylor expansion of the transition function; (2) choose between logistic (LSTAR) and exponential (ESTAR) forms based on sequential F-tests; (3) estimate by NLS with a grid of starting values; (4) evaluate with misspecification tests.

The STR family is widely used in exchange-rate modeling (purchasing-power-parity adjustment speeds depend on the size of the deviation), interest-rate dynamics (adjustment toward equilibrium accelerates when rates are far from target), and output-gap studies (Phillips-curve slopes differ in high-gap versus low-gap regimes). The Bank of Canada and the European Central Bank have used LSTAR models in their forecasting suites.

Connections to explore

Math ↗The structure and its conditions, step by step.References ↗Academic sources and guidance for using the model.

How the Parts Fit Together

G(st;γ,c)G(s_t; \gamma, c)G(st​;γ,c)

Transition function

Bounded [0,1][0, 1][0,1] function that determines the weight on regime 2. Logistic for asymmetric transitions, exponential for symmetric.

The STR takes the same inputs as a linear regression or VAR - a dependent variable (or vector), a set of regressors including lagged values, and a transition variable sts_tst​ that may be one of the regressors, a lagged endogenous variable, a time trend, or an exogenous indicator. The transition variable choice is the most consequential modeling decision, as it determines what drives the regime switch.

The core mechanical component is the transition function G(st;γ,c)G(s_t; \gamma, c)G(st​;γ,c). The logistic form (LSTAR) produces asymmetric dynamics: behavior differs depending on whether the economy is above or below ccc. The exponential form (ESTAR) produces symmetric dynamics: behavior at both extremes (far above and far below ccc) differs from behavior near ccc. These are qualitatively different stories - the LSTAR says booms and recessions obey different rules; the ESTAR says large deviations from equilibrium (in either direction) trigger different dynamics than small deviations.

Estimation requires choosing starting values for γ\gammaγ and ccc, then running NLS to minimize the residual sum of squares jointly over (β1,β2,γ,c)(\beta_1, \beta_2, \gamma, c)(β1​,β2​,γ,c). The transition speed γ\gammaγ is scale-dependent (its magnitude depends on the variance of sts_tst​), so practitioners often standardize sts_tst​ before estimation or divide γ\gammaγ by the sample standard deviation of sts_tst​. Output includes the estimated transition function G(st;γ^,c^)G(s_t; \hat{\gamma}, \hat{c})G(st​;γ^​,c^) evaluated at each observation, which provides a continuous regime-membership score between 0 and 1.

Assumptions

Testable

Observable transition variable

The variable driving the transition (st)(s_t)(st​) is observable and predetermined at time ttt.

If violated:If the transition driver is latent, the STR cannot be estimated. Markov-switching is more appropriate.

Maintained

Smooth, monotone transition

The transition function GGG is smooth, bounded between 0 and 1, and monotone in sts_tst​ (logistic) or symmetric around ccc (exponential).

If violated:Misspecification of the transition function shape (e.g., using logistic when the true form is exponential) biases γ\gammaγ and ccc and distorts regime weights.

Testable

Correct transition variable selection

The chosen sts_tst​ is the true driver of regime changes, or at least a good proxy for it.

If violated:If sts_tst​ is incorrect, the linearity test may fail to reject, or the estimated transition function will be flat (small γ\gammaγ), producing results close to the linear model.

Testable

Ergodicity and mixing

The process {yt,st}\{y_t, s_t\}{yt​,st​} is strictly stationary and β\betaβ-mixing with geometric mixing rate.

If violated:Non-stationarity invalidates the NLS asymptotics. Unit roots in yty_tyt​ or sts_tst​ require cointegration-based STR extensions.

Testable

Innovation regularity

Innovations are i.i.d. with zero mean, finite fourth moments, and constant variance (or specified heteroskedasticity structure).

If violated:Heavy tails inflate NLS standard errors. Unmodeled ARCH effects can be confused with transition effects.

Testable

Single-transition specification

One transition function is sufficient. No additional nonlinearity remains after accounting for the STR component.

If violated:Remaining nonlinearity shows up in Eitrheim-Terasvirta (1996) tests and suggests a second transition function or a different model class.

Applications

The European Central Bank uses LSTAR models to study asymmetric Phillips-curve dynamics: inflation responds more strongly to the output gap when the gap is positive (overheating) than when it is negative (slack). Terasvirta and Anderson (1992) documented this asymmetry for several OECD economies, and the finding has been replicated in more recent samples with time-varying natural rates.

Exchange-rate economists use ESTAR models to study purchasing-power-parity (PPP) adjustment. Michael, Nobay, and Peel (1997) showed that real exchange rates revert to PPP faster when deviations are large (in either direction), which is exactly the symmetric nonlinearity that ESTAR captures. This transaction-cost interpretation - small deviations are not worth arbitraging, large deviations trigger trade flows - motivates the exponential transition form.

The STR cannot handle latent regimes, multiple simultaneous transitions, or high-dimensional systems without severe curse-of-dimensionality problems. When the transition speed γ\gammaγ is weakly identified (flat likelihood surface), the confidence interval for γ\gammaγ can span the entire feasible range, making the nonlinearity essentially untestable. The model also assumes a single transition variable; if regime changes are driven by a combination of variables (e.g., output gap and credit spread jointly), the STR requires ad hoc dimension reduction.

Recent extensions include panel STAR models (Gonzalez, Terasvirta, and van Dijk 2005), smooth-transition GARCH for volatility dynamics, and multivariate STVAR systems (Hubrich and Terasvirta 2013). Bayesian estimation with priors on γ\gammaγ helps regularize the transition speed estimate when the likelihood surface is flat.

Extensions and open questions

Variants

  • LSTAR (logistic) - asymmetric transition, most common form
  • ESTAR (exponential) - symmetric transition around location c
  • LSTVAR / STVAR - multivariate smooth-transition VAR
  • TV-STAR - time-varying parameters within each regime
  • Panel STAR - cross-section extension with heterogeneous transition
  • ST-GARCH - smooth-transition volatility models

Open questions

  • Identification of the transition speed γ\gammaγ remains the central estimation challenge. When γ\gammaγ is weakly identified, inference on ccc and the regime coefficients is also unreliable, but corrected standard errors for this case are not widely available.
  • Selecting between LSTAR and ESTAR using sequential F-tests has low power when the sample is short or the true transition form does not match either template.
  • Extension to high-dimensional systems (large-nnn STVAR) faces the curse of dimensionality: the number of parameters scales with n2n^2n2 per regime plus the transition parameters.

Keep going

Read next

Continue with a related concept, model, or example.

Theory model

IS-LM Model

a well-connected theoretical model two steps from here

New angle: theoretical model →
Macro by Mark
Help
Privacy PolicyTerms of UseEthics and Compliance

© 2026 Mark Jayson Nation LLC