Nonlinear SARIMA Model
Nonlinear Seasonal Autoregressive Integrated Moving Average Model · Also known as: NL-SARIMA, nonlinear seasonal ARIMA, threshold SARIMA, smooth transition SARIMA
The Nonlinear SARIMA model extends the classical Seasonal ARIMA framework by replacing the linear conditional mean function with a nonlinear specification — such as threshold switching or smooth transition — while retaining seasonal differencing and lag structure. It is used when seasonal time series exhibit regime-dependent dynamics, asymmetric adjustment, or other nonlinear patterns that a linear model cannot capture.
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 Nonlinear SARIMA when a linear SARIMA has been correctly identified and estimated but residual nonlinearity tests (LST or RESET) reject linearity, or when domain knowledge suggests regime-dependent seasonal dynamics (e.g., energy demand spikes only in extreme cold, tourism flows with asymmetric shocks). It is also appropriate when forecast errors from a linear SARIMA are systematically larger in certain regimes. Do not use it as a default specification: the added complexity is only justified when nonlinearity is evidenced by tests or theory. Avoid it with very short series (fewer than ~80 observations) where threshold parameters cannot be reliably estimated.
Strengths & limitations
- Captures regime-switching, asymmetric cycles, and amplitude-dependent mean reversion that linear SARIMA misses.
- Retains the interpretable seasonal differencing and lag structure of classical SARIMA, making it easier to communicate than fully nonparametric alternatives.
- Threshold and smooth-transition variants have well-developed testing and estimation theory, including the LST linearity test and grid-search threshold estimation.
- Can substantially improve multi-step seasonal forecasts when the data-generating process is genuinely nonlinear.
- Nested within the linear SARIMA as a special case, enabling formal likelihood-ratio or LM tests to justify the added complexity.
- Requires substantially larger samples than linear SARIMA; threshold estimation is unreliable below ~80 observations, and smooth-transition speed parameters are hard to identify with short series.
- Model selection (number of regimes, choice of transition variable, lag lengths for each regime) involves multiple specification searches, increasing the risk of overfitting.
- Forecasting from threshold models requires simulation or numerical integration because point forecasts from a piecewise-linear model are not a simple closed-form expression.
- Interpretation is more complex: each regime has its own autoregressive and moving-average coefficients, and the transition variable must be economically meaningful.
- Estimation is nonlinear and iterative; convergence failures or multiple local optima can occur, especially for smooth-transition models with poorly chosen starting values.
Frequently asked
How is a Nonlinear SARIMA different from a standard SARIMA?
A standard SARIMA constrains the conditional mean to be a fixed linear function of lagged values and lagged errors across all regimes. A Nonlinear SARIMA keeps the same seasonal differencing and lag skeleton but allows the parameters of that linear function to switch — abruptly in a threshold model or gradually in a smooth-transition model — depending on a state variable. The seasonal differencing order (D, s) is chosen the same way in both cases.
Which nonlinearity test should I use before fitting?
The Luukkonen-Saikkonen-Terasvirta (LST, 1988) LM test is the most widely used. It tests the null of a linear AR against the alternative of a logistic smooth-transition AR by regressing residuals from the linear fit on polynomials of the candidate transition variable. If the test rejects at 5%, there is evidence of smooth-transition nonlinearity; threshold nonlinearity can additionally be tested with Hansen's (1997) sup-LM bootstrap.
Can I combine a Nonlinear SARIMA mean with a GARCH error process?
Yes. It is common in financial and energy time series to model the conditional mean with a nonlinear (threshold or STAR) SARIMA and the conditional variance with GARCH or seasonal GARCH. Estimation typically proceeds in two stages: fit the nonlinear mean first, then fit a GARCH model to the squared residuals.
How many observations do I need for reliable threshold estimation?
As a practical guideline, each regime should contain at least 15–20% of the observations, and the total sample should be at least 80–100 observations to achieve reasonable precision on the threshold location and regime-specific coefficients. With fewer data points the threshold estimate has wide confidence intervals and the regime-specific parameters are highly uncertain.
Is the Nonlinear SARIMA nested inside a linear SARIMA?
Yes, but testing the restriction is non-standard. Under the null of linearity the threshold parameter r (or the transition speed γ) is unidentified, so classical chi-squared critical values are invalid. The Davies (1977) bound or a bootstrap p-value from Hansen's sup-LM test should be used instead of a simple likelihood-ratio test.
Sources
- Tong, H. (1990). Non-linear Time Series: A Dynamical System Approach. Oxford University Press. ISBN: 978-0198523000
- Franses, P. H., & van Dijk, D. (2000). Non-linear Time Series Models in Empirical Finance. Cambridge University Press. ISBN: 978-0521779654
How to cite this page
ScholarGate. (2026, June 3). Nonlinear Seasonal Autoregressive Integrated Moving Average Model. ScholarGate. https://scholargate.app/en/econometrics/nonlinear-sarima-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
- GARCH ModelEconometrics↔ compare
- SARIMA modelEconometrics↔ compare