Seasonal ARIMA Process

• A Seasonal ARIMA process is a time series model that combines nonseasonal and seasonal differencing with AR and MA components to transform nonstationary data into a stationary series.
• It applies high-pass and comb filtering techniques to effectively remove polynomial trends and seasonal periodicities, ensuring reliable analysis.
• Applications span finance, environmental science, and engineering, with robust estimation methods like MLE, ECF, and Whittle likelihood enhancing its forecasting precision.

A Seasonal ARIMA (SARIMA) process is a generalized class of time series models specifically designed to accommodate both nonstationary trends and periodic seasonal structures in observed data. SARIMA models leverage autoregressive (AR), moving-average (MA), and differencing operators—applied both on the nonseasonal and seasonal time scales—to transform original nonstationary series into a stationary series amenable to classical time series analysis and forecasting. Processes with fractional integration and stable innovations further extend SARIMA’s modeling flexibility, allowing for heavy tails and long-memory behavior, as encountered in finance, environmental, and engineering contexts.

1. Mathematical Formulation and Model Structure

The canonical SARIMA(p, d, q)×(P, D, Q)ₛ model applies both nonseasonal and seasonal differencing to a univariate time series $\{X_t\}$, with $B$ the back-shift operator ($BX_t = X_{t-1}$) and $s$ the seasonal period. The formal parametric model is

$$\Phi(B^s)\,\phi(B)\,(1 - B)^d\,(1 - B^s)^D\,X_t \;=\;\Theta(B^s)\,\theta(B)\,\varepsilon_t$$

where

• $\phi(B) = 1- \phi_1 B - \dotsb - \phi_p B^p$ \hspace{1em}(nonseasonal AR)
• $$\Phi(B^s) = 1 - \Phi_1 B^s - \dotsb - \Phi_P B^{Ps}$$ \hspace{1em}(seasonal AR)
• $\theta(B) = 1 + \theta_1 B + \dotsb + \theta_q B^q$ \hspace{1em}(nonseasonal MA)
• $$\Theta(B^s) = 1 + \Theta_1 B^s + \dotsb + \Theta_Q B^{Qs}$$ \hspace{1em}(seasonal MA)
• $(1 - B)^d$, $B$0 \hspace{1em}(nonseasonal and seasonal differencing of order $B$1 and $B$2).

This generalizes ARIMA(p, d, q) to include multiplicative seasonal AR and MA polynomials and seasonal differencing (Ndongo et al., 2012, Chai, 2021, Sak et al., 2012, Tewari, 2020).

2. Principles of Seasonal and Nonseasonal Differencing

Nonseasonal differencing of order $B$3 and seasonal differencing of order $B$4 are implemented via the operators $B$5 and $B$6. In spectral terms:

• $B$7 acts as a high-pass FIR filter, amplifying mid- and high-frequency energy, with a $B$8-fold zero at frequency $B$9 to remove polynomial trends.
• $BX_t = X_{t-1}$0 acts as a comb filter, introducing $BX_t = X_{t-1}$1-fold zeros at seasonal harmonics $BX_t = X_{t-1}$2, removing periodic content of period $BX_t = X_{t-1}$3.

Impulse responses are finite:

$BX_t = X_{t-1}$4

for $BX_t = X_{t-1}$5 for ordinary differencing, and analogous expressions for seasonal differencing (Wang et al., 2019).

Standard differencing robustly removes nominal trend and seasonal periodicities, transforming the series into wide-sense stationary form, but is susceptible to over-differencing (excess attenuation and noise amplification) and frequency leakage when true spectral peaks do not align exactly with seasonal harmonics (Wang et al., 2019).

3. Model Identification, Estimation, and Diagnostics

Box–Jenkins methodology governs the SARIMA modeling cycle (Sak et al., 2012, Chai, 2021, Tewari, 2020):

1. Identification: Visual inspection and autocorrelation (ACF/PACF) analysis suggest orders $BX_t = X_{t-1}$6 and $BX_t = X_{t-1}$7.
2. Order Selection: Informational criteria (AIC, BIC) and additional hypothesis testing (ADF tests for stationarity) are applied to select optimal model structure.
3. Parameter Estimation: Maximum likelihood methods estimate polynomials’ coefficients and innovation variance.
4. Diagnostic Checking: Residuals are assessed for white-noise properties and normality; Ljung–Box Q-tests and normality tests validate model adequacy.

Sparse parameterizations may be preferred—eliminating statistically non-significant terms for parsimony and stability in forecasting (e.g., SARIMA$BX_t = X_{t-1}$8 for Hong Kong air traffic) (Chai, 2021).

4. Extensions: Fractional Integration and Heavy-Tailed Innovations

Fractional seasonal ARIMA (SARFIMA, ARFISMA) extend classical integer-order differencing to real-valued indices $BX_t = X_{t-1}$9, using binomial or Gegenbauer expansions to define $s$0, $s$1 (Reisen et al., 2010, Ndongo et al., 2012). This admits both long-memory at frequency zero and seasonal frequencies.

Innovations may follow symmetric $s$2-stable ($s$3) laws with characteristic function $s$4, $s$5, modeling infinite-variance (“heavy-tailed”) regimes common in finance and telecommunications (Ndongo et al., 2012). Stationarity and invertibility require strict constraints, e.g., $s$6, $s$7, $s$8 (Reisen et al., 2010, Ndongo et al., 2012).

Semiparametric log-periodogram approaches yield consistent and $s$9-normal estimators for fractional parameters in models with one or more seasonal periods (Reisen et al., 2010).

5. Estimation Strategies: ECF, Whittle-MLE, and Simulation

Two principal estimation frameworks are dominant for seasonal fractional models:

• Empirical Characteristic Function (ECF): Minimizes integrated squared deviation of the empirical joint CF over overlapping blocks to the parametric model CF. Yields simultaneous, robust, consistent, and asymptotically normal estimators for all parameters ($$\Phi(B^s)\,\phi(B)\,(1 - B)^d\,(1 - B^s)^D\,X_t \;=\;\Theta(B^s)\,\theta(B)\,\varepsilon_t$$0, AR, MA, $$\Phi(B^s)\,\phi(B)\,(1 - B)^d\,(1 - B^s)^D\,X_t \;=\;\Theta(B^s)\,\theta(B)\,\varepsilon_t$$1) (Ndongo et al., 2012).
  • For ARFISMA–SαS, closed-form joint CF is available:

  $$\Phi(B^s)\,\phi(B)\,(1 - B)^d\,(1 - B^s)^D\,X_t \;=\;\Theta(B^s)\,\theta(B)\,\varepsilon_t$$2

• Two-Step Method (TSM): First, estimates $$\Phi(B^s)\,\phi(B)\,(1 - B)^d\,(1 - B^s)^D\,X_t \;=\;\Theta(B^s)\,\theta(B)\,\varepsilon_t$$3 via MCMC Whittle likelihood in the frequency domain; then $$\Phi(B^s)\,\phi(B)\,(1 - B)^d\,(1 - B^s)^D\,X_t \;=\;\Theta(B^s)\,\theta(B)\,\varepsilon_t$$4 via MLE on pseudo-innovations. TSM may lose efficiency when short-memory terms and stable parameters interact (Ndongo et al., 2012).

Monte Carlo studies demonstrate ECF’s superior finite-sample RMSE and MAE for simultaneous estimation in mixed fractional–short memory models (Ndongo et al., 2012).

Conditional simulation methodology enables generation of sample paths and continuation forecasts from fitted SARIMA models, as shown in R implementations; forecasts and their error structure closely agree with theoretical ARIMA predictions (Sak et al., 2012).

6. Computational and Methodological Advancements

Classical SARIMA is routinely implemented in statistical software via state-space Kalman filtering, conditional MLE, and recursive ARMA forecast equations (Sak et al., 2012, Tewari, 2020). In contrast, the univariate time-varying approach for periodic ARMA/SARIMA (Karanasos et al., 2014) eliminates recursions and high-dimensional VAR representations; analytic expressions for multi-step predictors and MSFE are constructed via continuant determinants of periodic coefficients (Karanasos et al., 2014). This maintains computational tractability regardless of seasonal period.

Recent work also critiques blunt-differencing approaches in SARIMA, recommending spectral analysis and filter synthesis methods (“ARMA–SIN”) where custom-designed FIR/IIR filters precisely remove nonstationary frequencies, yielding stationary series for ARMA modeling without midband distortion (Wang et al., 2019).

7. Applications, Limitations, and Interpretive Dimensions

SARIMA and its fractional/stable extensions are widely deployed in economics, finance, air-traffic forecasting, environmental monitoring, and telecommunication signal analysis (Chai, 2021, Tewari, 2020, Reisen et al., 2010). The ability to forecast under stable social conditions and to estimate counterfactual trajectories during crises or regime changes demonstrates SARIMA’s utility as both a predictive and impact-quantification tool (Chai, 2021).

Nevertheless, practical limitations, such as over-differencing, inability to completely isolate nonstationary frequency content, and sensitivity to short-memory misspecification, motivate more flexible transformation strategies and diagnostic procedures (Wang et al., 2019). Seasonal ARIMA theory continues to evolve through advances in long-memory estimation, heavy-tailed inference, and multiseasonal generalizations.