Fourier SARIMA Model
Also known as: Fourier SARIMA, SARIMA with Fourier terms, Fourier-SARIMA, trigonometric SARIMA
The Fourier SARIMA model extends the classical Seasonal ARIMA framework by incorporating trigonometric (Fourier) terms as deterministic regressors. This allows the model to approximate smooth, complex, or multiple-frequency seasonal patterns without requiring a full seasonal ARIMA structure for every frequency, making it particularly useful for high-frequency data or series with non-integer or evolving seasonality.
Key highlights
- Handles large or non-integer seasonal periods (weekly, daily, hourly data) where classical SARIMA becomes over-parameterized.
- Parsimonious: a few Fourier harmonics can represent complex multi-frequency seasonal patterns with fewer parameters than full seasonal AR/MA terms.
- Allows multiple overlapping seasonal periods to be modeled simultaneously by stacking Fourier terms at different fundamental frequencies.
- Smooth seasonal approximation is statistically efficient and avoids over-fitting seasonal structure.
- Compatible with the full ARIMA machinery for trend, integration, and non-seasonal autocorrelation.
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 Fourier SARIMA when the seasonal period m is large (weekly, daily, or sub-daily data where m > 12) and a full seasonal ARIMA is impractical, or when the seasonal shape is smooth and complex rather than a single simple cycle. It is also appropriate when multiple seasonal periods overlap (e.g., both weekly and annual cycles in daily data) — in that case multiple sets of Fourier terms at different periods can be stacked. Do not use it as a drop-in replacement for standard SARIMA when m is small (e.g., m = 4 or 12) and the seasonal pattern is stable and well-captured by simple seasonal differencing, since Fourier SARIMA adds estimation complexity without benefit. Also avoid it when the seasonal pattern changes abruptly (structural breaks in seasonality), as trigonometric approximations assume smooth variation.
Strengths & limitations
- Handles large or non-integer seasonal periods (weekly, daily, hourly data) where classical SARIMA becomes over-parameterized.
- Parsimonious: a few Fourier harmonics can represent complex multi-frequency seasonal patterns with fewer parameters than full seasonal AR/MA terms.
- Allows multiple overlapping seasonal periods to be modeled simultaneously by stacking Fourier terms at different fundamental frequencies.
- Smooth seasonal approximation is statistically efficient and avoids over-fitting seasonal structure.
- Compatible with the full ARIMA machinery for trend, integration, and non-seasonal autocorrelation.
- Requires choosing K (number of harmonics), which adds a model-selection step; misspecification of K can leave residual seasonal autocorrelation or waste parameters.
- Fourier terms assume a smooth, approximately sinusoidal seasonal shape; sharp, irregular, or abruptly shifting seasonality is poorly approximated.
- With very many overlapping seasonal frequencies, the regressor matrix can become nearly collinear, complicating estimation.
- The method does not detect seasonal structural breaks — a Fourier-based unit root or break test should be run separately if breaks are suspected.
- Interpretation of individual Fourier coefficients is less intuitive than seasonal AR/MA parameters.
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 many Fourier terms K should I include?
Start with K = 1 and increase up to m/2 (or a practical maximum of 6–10), selecting the value that minimizes AIC or BIC. For most applied series 2–4 harmonics are sufficient to capture the dominant seasonal shape.
Can I use Fourier terms together with seasonal differencing (D = 1)?
Yes, but do so carefully. If the seasonal pattern has a stochastic (unit root) component, seasonal differencing is appropriate; Fourier terms then model the deterministic residual seasonality. However, combining seasonal differencing with a full set of Fourier terms can introduce near-collinearity, so use information criteria to guide the balance.
How does Fourier SARIMA differ from TBATS or Prophet?
TBATS (Trigonometric seasonality, Box-Cox, ARMA errors, Trend, Seasonal) uses a state-space form with exponential smoothing and embeds Fourier-like seasonal states; Prophet uses additive Fourier regressors with trend change points. Fourier SARIMA is more transparent, operates within the familiar ARIMA framework, and is easier to test and diagnose with standard tools, but requires more manual order selection.
Is Fourier SARIMA appropriate for weekly data?
Yes — weekly data with an annual cycle has m = 52, making a full seasonal ARIMA with 52 seasonal lags impractical. A small number of Fourier harmonics (typically K = 2–4) provides an efficient and estimable approximation.
What software implements Fourier SARIMA?
In R, the forecast package (Hyndman et al.) provides the fourier() function to generate Fourier regressors paired with auto.arima() or Arima() with an xreg argument. Python users can construct Fourier regressors manually and pass them to statsmodels SARIMAX as exogenous variables.
Sources
- 1.Harvey, A., & Scott, A. (1994). Seasonality in dynamic regression models. The Economic Journal, 104(427), 1324-1345.
- 2.Hyndman, R. J., & Athanasopoulos, G. (2018). Forecasting: Principles and Practice (2nd ed.). OTexts.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Fourier SARIMA model. ScholarGate. https://scholargate.app/econometrics/fourier-sarima-model