Seasonal persistence dominates monthly tropospheric ozone predictability in Campo Grande, Brazil: SARIMA benchmarking and an SSA-based hybrid approach

Tropospheric ozone (O₃) is a secondary atmospheric pollutant whose temporal variability is controlled by the interaction of photochemical processes, meteorological conditions, and precursor emissions, making reliable forecasting essential for air-quality management. This study investigates the predictability of monthly mean tropospheric ozone concentrations in Campo Grande, Brazil, during the period 2000–2018, with emphasis on the relative performance of statistical, machine-learning, and hybrid forecasting approaches. Singular Spectrum Analysis (SSA) was first employed to decompose the ozone time series into structured seasonal and residual components. Forecasting performance was then evaluated using a Seasonal Autoregressive Integrated Moving Average (SARIMA) model, Support Vector Regression (SVR), Gaussian Support Vector Regression (GSVR), a conventional SSA–GSVR model, and a novel SSA–SARIMA–SVR hybrid framework under a rolling-origin validation strategy. The SSA decomposition revealed that the monthly ozone series is dominated by a persistent annual cycle, indicating that seasonal variability constitutes the principal source of forecast predictability. Among the evaluated models, SARIMA consistently achieved the lowest forecasting errors across all prediction horizons, demonstrating that the dominant temporal dynamics are adequately represented by a parsimonious seasonal stochastic model. The proposed SSA–SARIMA–SVR framework produced competitive forecasting performance while providing a physically interpretable separation between deterministic seasonal variability and irregular nonlinear fluctuations, substantially improving the stability of conventional SSA-based hybrid modeling. These findings demonstrate that forecasting performance depends primarily on the intrinsic statistical structure of the ozone series rather than on model complexity alone. The proposed framework provides an interpretable methodology for atmospheric time-series analysis and offers a flexible basis for future multivariate forecasting systems integrating meteorological and emission-related predictors.