Time-Varying Parameter ARMA Model (TVP-ARMA)
Also known as: TVP-ARMA, time-varying ARMA, state-space ARMA, locally stationary ARMA
The time-varying parameter ARMA (TVP-ARMA) model extends the classical ARMA framework by allowing the autoregressive and moving-average coefficients to evolve over time. Embedded in a state-space representation and estimated via the Kalman filter, it captures structural change and parameter instability in time series without requiring an explicit breakpoint.
Key highlights
- Detects gradual and abrupt parameter change without imposing a known break date.
- Nested with constant-coefficient ARMA: testing Q = 0 gives a formal test of parameter stability.
- Produces a full trajectory of time-varying coefficients, enabling richer structural interpretation.
- Kalman filter estimation is computationally exact under Gaussian errors and scales well with sample size.
- Naturally accommodates missing observations within the state-space framework.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use TVP-ARMA when you suspect that the dynamic structure of a univariate time series has changed over the sample — for example, after policy regime shifts, structural breaks, or gradual economic transitions — but you do not want to impose a fixed break date. It is also appropriate when rolling-window ARMA fits show clear drift in estimated coefficients. Prefer a standard constant-parameter ARMA when the series is short (fewer than ~80 observations), when no evidence of instability exists (Nyblom test is insignificant), or when computational simplicity is paramount. TVP-ARMA is not designed for multivariate modelling — a TVP-VAR is more appropriate in that setting.
Strengths & limitations
- Detects gradual and abrupt parameter change without imposing a known break date.
- Nested with constant-coefficient ARMA: testing Q = 0 gives a formal test of parameter stability.
- Produces a full trajectory of time-varying coefficients, enabling richer structural interpretation.
- Kalman filter estimation is computationally exact under Gaussian errors and scales well with sample size.
- Naturally accommodates missing observations within the state-space framework.
- Estimation is sensitive to the initial conditions of the Kalman filter and the prior on Q, especially in small samples.
- Identification of the state-noise covariance Q is difficult when parameter variation is slow — the likelihood surface can be flat.
- The model can over-fit short series by attributing noise to spurious parameter drift.
- Forecasting performance deteriorates when the out-of-sample structure diverges sharply from the recent in-sample trajectory.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does TVP-ARMA differ from GARCH?
GARCH models time-varying conditional variance (volatility) while keeping the conditional mean structure fixed. TVP-ARMA instead allows the lag coefficients in the conditional mean equation to vary, making it suitable for modelling changes in persistence or impulse-response dynamics rather than volatility clustering.
Can I use TVP-ARMA for forecasting?
Yes, but cautiously. In-sample fit is generally better than constant-coefficient ARMA, but out-of-sample forecasts depend heavily on whether the parameter drift observed in-sample continues. In periods of structural stability, a simpler ARMA often forecasts comparably or better.
How do I choose the orders p and q in TVP-ARMA?
Start with the orders selected by AIC or BIC under a constant-coefficient ARMA, then test for parameter instability (Nyblom test) at those orders. The information criteria for state-space models (computed from the Kalman filter likelihood) can be used to compare alternative TVP-ARMA(p,q) specifications.
What software can estimate TVP-ARMA models?
The STAMP package (Koopman et al.) and the R packages dlm, KFAS, and bsts all support state-space ARMA estimation with time-varying parameters. Bayesian TVP-ARMA can be estimated with Stan or JAGS.
Is TVP-ARMA the same as a locally stationary process?
Related but not identical. Locally stationary processes (Dahlhaus 1997) provide a rigorous asymptotic theory for continuously varying spectra, while TVP-ARMA is a parametric state-space model. Both capture parameter non-stationarity, but TVP-ARMA is estimation-oriented and practical for applied work, whereas local stationarity theory underpins non-parametric spectral approaches.
Sources
- 1.Cooley, T. F., & Prescott, E. C. (1976). Estimation in the presence of stochastic parameter variation. Econometrica, 44(1), 167–184.
- 2.Harvey, A. C. (1989). Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge University Press.ISBN 9780521405737
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Time-varying parameter ARMA model. ScholarGate. https://scholargate.app/econometrics/time-varying-parameter-arma-model