Trend–Seasonal Decomposition

• Trend–seasonal decomposition is a method that separates time series data into a smooth trend, cyclic seasonal patterns, and a noise component for clearer analysis.
• It employs techniques like moving averages, LOESS smoothing, spectral methods, and neural network modules to extract hidden patterns and manage challenges such as incommensurate seasonalities and heteroscedasticity.
• This framework facilitates improved forecasting, anomaly detection, and uncertainty quantification by isolating noise and capturing both gradual trends and abrupt changes.

A trend–seasonal decomposition expresses a time series as the sum (or, in some cases, a function) of several interpretable components: a slowly-varying trend, one or more rapidly-varying seasonal (periodic or cyclic) components, and a residual or remainder representing noise and anomalies. This decomposition supports improved forecasting, anomaly detection, and interpretability in univariate and multivariate time series contexts. Advanced methods address challenges such as multiple or incommensurate seasonalities, spatiotemporal data, heteroscedasticity, and robustness to outliers or abrupt changes.

1. Formal Models and Mathematical Frameworks

Classically, the observed series is modeled as

$$y_t = \text{trend} + \text{seasonal} + \text{residual}$$

with extensions to multiple seasonalities and other components. Several specific model classes are prominent:

• Additive Decomposition: $y_t = T_t + \sum_{k=1}^K S_{k,t} + R_t$, with trend $T_t$, seasonality(ies) $S_{k,t}$, and remainder $R_t$ (Cho et al., 26 Jan 2026, Zhang et al., 2022, Li et al., 2020, Bandara et al., 2021, Dokumentov et al., 2020).
• Dispersion-aware Models: Models such as STD include a time-varying dispersion term, $y_t = T_t + D_t S_t$, making variance dynamics explicit (Dudek, 2022).
• Spatio-temporal Models: Extended as $$y(s,t) = \mu_t + s_t + c_t + z(s,t)^\top \beta + \xi(s,t) + \varepsilon(s,t)$$ to support spatial random effects and covariates (Laurini, 2017).

Contemporary neural architectures embed decomposition directly in feature transformation and prediction, notably in Transformer-based models (Cao et al., 2024, Qin et al., 2024, Zhang et al., 2022, Seo et al., 13 Jun 2025, Cao et al., 17 Feb 2025, Nematirad et al., 28 Mar 2025, Panta et al., 6 Feb 2026).

2. Trend and Seasonal Extraction Methodologies

A spectrum of methods is used for extracting trend and seasonal components:

• Moving Average / Median Filters: Both average-pooling and median-pooling over time windows provide simple and robust trend extraction (Li et al., 2020, Seo et al., 13 Jun 2025, Panta et al., 6 Feb 2026, Cao et al., 2024, Qin et al., 2024).
• LOESS and Local Regression: STL and its generalizations (e.g. MSTL for multiple seasonalities) apply Loess smoothing iteratively for both trend and each seasonal component (Bandara et al., 2021).
• Variance-normalized Decomposition: STD explicitly computes per-period means (trend) and unnormalized or normalized dispersions, with subsequent scaling of seasonal patterns (Dudek, 2022).
• Regularized and Robust Regression: Penalized regression formulations, as in STR, RobustSTL, and BASTION, solve for trends and seasonal surfaces under smoothness and/or sparsity constraints, enabling multiple seasonalities, irregular cycles, and robustness to abrupt changes or outliers (Dokumentov et al., 2020, Wen et al., 2018, Cho et al., 26 Jan 2026).
• Discrete and Continuous-domain Variational Methods: Convex programs with total-variation or generalized TV regularizers yield spline-based trend and periodic spline-based seasonal fits (Fageot, 15 May 2025).
• Piecewise-linear/Adaptive Regimes: LGTD replaces explicit seasonal frequency specification with local adaptive linear trends ("local trend regimes"), whose recurrence generates emergent seasonality (Sophaken et al., 8 Jan 2026).
• Spectral Approaches: Decomposition into frequency bands, or explicit construction of trigonometric seasonal terms, is used in both classical regression and deep learning pipelines (Gao et al., 2018, Zhang et al., 2022, Nematirad et al., 28 Mar 2025, Seo et al., 13 Jun 2025).

In modern neural networks, decomposition can be implemented as explicit layers (e.g. moving averages, Conv1D smoothers), modules with parameterized kernel sizes, or via parallel branches encoding trend and seasonality (Qin et al., 2024, Cao et al., 2024, Cao et al., 17 Feb 2025, Seo et al., 13 Jun 2025, Nematirad et al., 28 Mar 2025).

3. Handling Multiple Seasonalities and Heteroscedasticity

• Multiple Seasonal Patterns: Methods such as MSTL, STR, and BASTION generalize the STL paradigm to multiple coexisting seasonalities, supporting both integer and non-integer periods. STR achieves this via penalized regression over 2D surfaces or basis expansions, while MSTL applies Loess smoothing for each detected seasonal period (Bandara et al., 2021, Dokumentov et al., 2020, Cho et al., 26 Jan 2026).
• Dispersion Dynamics: STD decomposes at each cycle both trend and dispersion, normalizing seasonal components to unit (or known) variance, yielding explicit series for evolving volatility (Dudek, 2022).
• Regression-based Covariate and Topology Effects: STR incorporates exogenous covariates—static, smoothly time-varying, or seasonal—jointly with trend and seasonality, using ridge-style penalties for smoothness. Complex seasonality "topologies" (e.g., switching between weekday/weekend or holiday patterns) are modeled via cyclic graph adjacency and difference penalties (Dokumentov et al., 2020).
• Outlier and Break Robustness: Methods using $\ell_1$-based penalties (LAD), global-local shrinkage priors (horseshoe), or explicit non-local filtering provide adaptivity to abrupt regime shifts and point anomalies (Wen et al., 2018, Cho et al., 26 Jan 2026, Li et al., 2020).

4. Computational Complexity and Online Methods

• Batch Algorithms: Classical STL and its generalizations have complexity $O(n w)$ per fit (n=series length, w=window), with worst-case cost scaling as $O(\text{num\_seasons}\times n\times w)$ (Bandara et al., 2021).
• Online and Real-Time Decomposition: OneShotSTL achieves $O(1)$ amortized time per new sample via iteratively reweighted least-squares banded-system updates with online LDL$^\top$ factorization (He et al., 2023).
• Multi-Scale/Resolution Decomposition: Down-sampling accelerates long-period extraction, as in the multi-scale ADMM algorithm (Yang et al., 2021).
• Deep Learning Integration: Conv1D-based decomposition modules, average-pooling, and efficient attention mechanisms allow end-to-end training within large neural architectures with linear or near-linear scaling (Cao et al., 2024, Nematirad et al., 28 Mar 2025, Panta et al., 6 Feb 2026).
• Explicit Linearity: Dual-MLP and trend-specific normalization modules (e.g., RevIN) in hybrid architectures maintain linear per-sample complexity and improve scaling (Panta et al., 6 Feb 2026).

5. Applications in Forecasting, Anomaly Detection, and Uncertainty Quantification

Decomposed components enable more interpretable and accurate downstream modeling:

• Forecasting Pipelines: Trend, seasonality, and dispersion can be separately forecasted via ARIMA/ETS models, ML regressors, or neural architectures. Additive recombination generates the final forecast (Dudek, 2022, Panta et al., 6 Feb 2026, Cho et al., 26 Jan 2026, Qin et al., 2024).
• Anomaly Detection: Robust decomposition leaves anomalies in the residuals, which can be detected via median/MAD-sensitive ESD tests (Li et al., 2020), end-to-end learning as in TADNet with mask separation (Zhang et al., 2023), or as residual spikes in the remainder component. Residual calibration and outlier scores arise naturally.
• Uncertainty Quantification: Model-based frameworks (e.g., STR and BASTION) provide closed-form confidence intervals or posterior credible sets for all components via covariance propagation or Bayesian sampling (Cho et al., 26 Jan 2026, Dokumentov et al., 2020).
• Interpretability and Attribution: Component-specific modules and gating (e.g., Ister’s dot-attention) allow attribution of forecasted values to trend or seasonal sources, and explicit output of per-channel contributions (Cao et al., 2024, Qin et al., 2024).

6. Robustness, Model Selection, and Practical Considerations

• Outlier Immunity: Use of median filters, $\ell_1$ trend filtering, or horseshoe shrinkage achieves robustness to spikes and regime shifts (Li et al., 2020, Wen et al., 2018, Cho et al., 26 Jan 2026).
• Model Selection: Information criteria (BIC, cross-validation) determine optimal polynomial orders, number of trigonometric terms, or penalty weights (Gao et al., 2018, Dokumentov et al., 2020).
• Hyperparameterization: Kernels, moving averages, window sizes, and penalty strengths are chosen based on data periodicity, cross-validation, or end-to-end optimization as part of a neural pipeline (Qin et al., 2024, Seo et al., 13 Jun 2025).
• Spatiotemporal and Multivariate Extension: Methods accommodate high-dimensional and spatial data via factor models (Gao et al., 2018), spatiotemporal projections and GMRFs (Laurini, 2017), and dual-path neural processing (Cao et al., 17 Feb 2025).

7. Recent Advances and Empirical Performance

Recent research demonstrates:

• Linear-Time, Differentiable Decomposition: Methods such as KEDformer, OneShotSTL, and patch-based models introduce fully differentiable, decomposition-aware neural components with constant-time per sample performance (Qin et al., 2024, He et al., 2023).
• Variance-awareness and Pattern Generalization: STD achieves explicit heteroscedasticity modeling, while BASTION provides robust uncertainty bands even under abrupt breaks or stochastic volatility (Dudek, 2022, Cho et al., 26 Jan 2026).
• Component-Specific Masked Modeling and Contrastive Learning: ST-MTM and SPDNet leverage decomposition-aware, component-wise masking and loss to focus model capacity on semantic substructures (Seo et al., 13 Jun 2025, Nematirad et al., 28 Mar 2025).
• Empirical Validation: Across a range of energy, web, atmospheric, and financial datasets, decomposition-enhanced models consistently outperform or complement baselines lacking explicit trend/seasonal separation, with ~10% MSE reduction recurring in benchmarks (Panta et al., 6 Feb 2026).
• Flexible, Topology-Aware Multiple Seasonality: STR enables modeling of arbitrary seasonal topologies, time-varying covariate effects, and outputs immediate component confidence intervals (Dokumentov et al., 2020).

Research continues to advance trend–seasonal decomposition as a fundamental primitive for interpretable, robust, and efficient time series modeling, with growing integration in forecasting, anomaly detection, and self-supervised representation learning.