Fourier Moving Average (Fourier MA) Model
Also known as: Fourier MA, Fourier-augmented moving average, trigonometric MA model, harmonic moving average model
The Fourier MA model combines a Moving Average (MA) error structure with Fourier series terms — sine and cosine pairs — to capture complex or high-frequency seasonal patterns in time series data. It is particularly useful when the seasonal period is long or irregular, making classical seasonal ARIMA parameterisation infeasible.
Key highlights
- Handles long and high-frequency seasonal periods (daily, weekly, sub-hourly) that overwhelm seasonal ARIMA.
- Parsimonious: a small number of Fourier harmonics K can represent complex smooth seasonal shapes.
- Combines deterministic seasonality with a flexible stochastic residual structure (MA).
- Coefficients are directly interpretable as seasonal amplitude and phase contributions.
- Easy to extend: Fourier terms can be added to ARIMA, regression, or state-space frameworks.
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 the Fourier MA model when your time series has a long or non-integer seasonal period (e.g., m = 365 daily observations, or m = 52 weeks) that makes seasonal ARIMA impractical, and when you also see short-run autocorrelation in residuals best captured by an MA component. It suits regularly spaced, stationary (or differenced-to-stationarity) univariate series with smooth, repeating seasonal shapes. Avoid it when the seasonal pattern shifts substantially over time (prefer TBATS or STL-based approaches), when the series is non-stationary without clear periodicity, or when the sample is very short relative to the seasonal period.
Strengths & limitations
- Handles long and high-frequency seasonal periods (daily, weekly, sub-hourly) that overwhelm seasonal ARIMA.
- Parsimonious: a small number of Fourier harmonics K can represent complex smooth seasonal shapes.
- Combines deterministic seasonality with a flexible stochastic residual structure (MA).
- Coefficients are directly interpretable as seasonal amplitude and phase contributions.
- Easy to extend: Fourier terms can be added to ARIMA, regression, or state-space frameworks.
- Assumes the seasonal shape is fixed over time; evolving seasonality is not captured.
- Selecting K and q requires careful model comparison — over-specification leads to overfitting.
- The MA component assumes stationarity of the residuals; unit roots must be removed by differencing first.
- Performance degrades when the seasonal period is irregular or when multiple overlapping cycles interact strongly.
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 do I choose the number of Fourier harmonics K?
Fit models with K = 1, 2, … up to floor(m/2) and select K using AIC or BIC. Adding harmonics beyond the AIC-optimal K rarely improves out-of-sample accuracy and may overfit.
What is the difference between a Fourier MA and a seasonal ARIMA model?
Seasonal ARIMA uses seasonal lag operators and requires that the seasonal period m be an integer small enough to estimate seasonal AR/MA polynomials directly — feasible for m = 4 or 12 but impractical for m = 52 or 365. The Fourier MA model replaces seasonal AR/MA operators with a few trigonometric regressors, keeping the parameter count manageable for large m.
Can the Fourier MA model handle multiple seasonal periods?
Yes. You can include two sets of Fourier terms — one for each seasonal period (e.g., weekly and annual cycles in daily data) — and combine them with the MA error structure. This is more parsimonious than trying to build a doubly-seasonal ARIMA.
Does the series need to be stationary before fitting?
The MA component requires stationary residuals. Trend or unit-root non-stationarity should be removed by differencing or detrending before fitting. Fourier terms themselves address periodic patterns but not systematic drift.
How do I diagnose a good fit?
Examine ACF and PACF of residuals for remaining autocorrelation, check the Ljung-Box test for white-noise residuals, compare AIC/BIC across competing K and q values, and evaluate out-of-sample forecast accuracy (MAPE, RMSE) on a hold-out window.
Sources
- 1.Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). OTexts.
- 2.Harvey, A. C. (1993). Time Series Models (2nd ed.). MIT Press.ISBN 978-0262082242
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Fourier MA Model. ScholarGate. https://scholargate.app/econometrics/fourier-ma-model