Robust Autoregressive Model
Also known as: robust autoregression, outlier-robust AR, M-estimator AR, heavy-tail AR
The robust AR model fits an autoregressive time series specification using estimation methods — typically M-estimators or bounded-influence estimators — that resist distortion from outliers and heavy-tailed error distributions. Unlike OLS-based AR estimation, robust variants down-weight extreme observations so that a small number of contaminated data points cannot dominate the fitted dynamics.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use a robust AR model when a univariate time series is suspected to contain additive outliers (single anomalous observations), level shifts, or innovational outliers that cannot all be removed manually before estimation. It is appropriate when the error distribution is likely heavy-tailed (e.g., financial returns during crises) and when preserving the majority of the sample matters more than using a single-step estimator. Do not use it as a substitute for careful pre-processing when outliers have a known deterministic cause that can be modelled explicitly with a dummy variable. Also avoid when sample size is very small (n < 30), where robust estimators lose efficiency, or when the data are genuinely clean and Gaussian, since robust estimators are slightly less efficient than OLS in that ideal case.
Strengths & limitations
- Resistant to the influence of additive and innovational outliers that would severely bias OLS-based AR estimates.
- Delivers stable lag-coefficient estimates even when a modest fraction (up to the breakdown point, ~25–50%) of observations are contaminated.
- Produces bounded influence functions, preventing any single observation from having an arbitrarily large effect on the fit.
- Compatible with standard model-selection criteria (AIC/BIC adapted to robust loss) and can be extended to ARMA or ARIMA structures.
- Sandwich standard errors maintain valid inference under heavy-tailed or heteroscedastic innovations.
- Somewhat lower statistical efficiency than OLS-AR when the data are truly Gaussian and outlier-free.
- More computationally intensive due to iterative re-weighted least squares (IRLS) and sensitivity to initialization.
- Choice of tuning constant (Huber's c or Tukey's c) affects the efficiency-robustness trade-off and is not always obvious from data alone.
- Asymptotic theory is more complex; some software implementations use approximations that may be inaccurate in small samples.
- Does not automatically handle structural breaks or regime changes; a separate structural-break robust AR extension is needed for those settings.
Frequently asked
What is the difference between a robust AR model and simply removing outliers before OLS-AR?
Manual removal requires identifying outliers ex ante, which involves subjective judgement and risks discarding valid extreme events. Robust AR objectively down-weights unusual observations using a data-driven criterion, preserves all data points in the sample, and quantifies their influence through the weighting function.
Which robust loss function should I choose — Huber or Tukey bisquare?
Huber's loss is convex and gives a unique solution with moderate outlier resistance; it is a safe default. Tukey's bisquare completely down-weights very large residuals (yielding a local, not global, minimum) and offers higher breakdown resistance. For financial data with frequent extreme returns, Tukey's bisquare is often preferred; for macroeconomic series with rare data errors, Huber's is usually sufficient.
How do I choose the lag order p in a robust AR model?
Use the robust BIC or AIC computed from the robust log-likelihood, or apply information criteria on the robust residuals. Alternatively, start with a generous lag order suggested by the robust partial autocorrelation function and trim insignificant lags.
Can a robust AR model handle a unit root?
It can be applied to a unit-root process, but robust estimation does not resolve non-stationarity. Always apply a robust unit-root test (or standard ADF/PP test) first and difference the series if it is integrated before fitting a robust AR in levels.
Is a robust AR equivalent to a GARCH model for handling fat tails?
No. GARCH models time-varying conditional variance, which explains leptokurtosis through clustering of volatility. Robust AR addresses isolated large innovations by down-weighting them; it does not model the dynamics of variance. For financial series, both a robust conditional mean equation and a GARCH variance equation are sometimes combined.
Sources
- Martin, R. D., & Yohai, V. J. (1986). Influence functionals for time series. Annals of Statistics, 14(3), 781–818. DOI: 10.1214/aos/1176350027 ↗
- Francq, C., & Zakoian, J.-M. (2010). GARCH Models: Structure, Statistical Inference and Financial Applications. Wiley. ISBN: 978-0470683910
How to cite this page
ScholarGate. (2026, June 3). Robust Autoregressive Model. ScholarGate. https://scholargate.app/en/econometrics/robust-ar-model
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- ARIMA modelEconometrics↔ compare
- ARMA modelEconometrics↔ compare
- Autoregressive modelEconometrics↔ compare
- Robust GLSEconometrics↔ compare
- Robust OLSEconometrics↔ compare
- Robust VECMEconometrics↔ compare