Technical Summary: Adaptive Oscillatory-State Alignment for Time Series Forecasting (AOSNET)

1. Problem Statement

Long-term time series forecasting relies heavily on capturing recurring temporal structures. While existing methods often employ inductive biases to model periodicity, they typically assume that recurrence can be represented as a fixed repeated template. These approaches fall into three categories:

1. Explicit Period Estimation: Aligning observations based on a pre-defined or estimated dominant period (e.g., CycleNet, TQNet).
2. Global Spectral Components: Modeling periodicity through global frequency representations (e.g., FEDformer, FILM).
3. Learnable Periodic Embeddings: Using fixed templates indexed by cycle positions.

The authors argue that these fixed-template assumptions are fundamentally mismatched to real-world temporal dynamics, which are rarely rigidly periodic. Real-world signals often exhibit non-stationary oscillatory behavior characterized by three coupled factors:

• Amplitude Modulation: The intensity (envelope) of the signal varies over time ($dA/dt \neq 0$).
• Phase Drift: Peaks and troughs arrive earlier or later than predicted by a nominal period ($d\delta/dt \neq 0$).
• Local Frequency Variation: The instantaneous oscillation rate changes, causing consecutive cycles to have different durations ($d\omega/dt \neq 0$).

Under these conditions, methods relying on rigid templates or fixed periods fail to adapt to evolving local states, leading to performance degradation as non-stationarity intensifies.

2. Methodology: AOSNET

The paper proposes AOSNET, a Hilbert-guided forecasting framework that reformulates periodic modeling from "fixed template matching" to adaptive oscillatory-state alignment.

Core Concept: Oscillatory-State Space

Instead of treating periodicity as a repeated cycle, AOSNET operates in an oscillatory-state space defined by three local descriptors derived from the analytic signal via the Hilbert transform:

1. Log-Amplitude Envelope ($a$): $a = \log(|z| + \epsilon)$, where $z$ is the analytic signal.
2. Phase Representation ($c, s$): Continuous sine and cosine components of the phase ($\cos \phi, \sin \phi$) to avoid discontinuities from phase unwrapping.
3. Instantaneous Frequency ($\omega$): Estimated via phase increments between consecutive samples ($\omega_\ell = \angle(z_\ell z^*_{\ell-1})$).

These descriptors are computed using a differentiable, FFT-based implementation, requiring no manual period selection or autocorrelation search.

Architecture Components

1. Global Oscillatory Prior ($P$): A learnable, channel-wise tensor shared across all samples. Unlike fixed templates, $P$ is not constrained to a specific period length. It serves as a flexible oscillatory reference.
2. Descriptor-Conditioned Adaptive Gate:
   • The model computes analytic descriptors for both the observed sequence ($X$) and the global prior ($P$).
   • These descriptors are concatenated and processed by a lightweight 1D convolutional gate.
   • The gate produces a fusion coefficient $G \in [0, 1]$ for each time step and channel.
   • State Alignment: The input is adaptively interpolated: $\tilde{X} = (1 - G) \odot X + G \odot P$.
     • If the local state matches the prior, $G \approx 0$ (preserving the observation).
     • If there is a mismatch (e.g., phase drift), $G > 0$ (softly correcting the input toward the learned reference).
3. Dual-Path Forecasting Head:
   • Attention Path: Uses Multi-Head Self-Attention (MHA) on channel tokens to refine cross-variate dependencies after alignment.
   • Base Path: A direct temporal projection for stable per-channel extrapolation.
   • Fusion: The final prediction is a learnable weighted sum of the two paths.

3. Key Contributions

1. Reformulation of Periodicity: The paper shifts the paradigm from fixed-template matching to adaptive oscillatory-state alignment, explicitly identifying amplitude modulation, phase drift, and local frequency variation as key sources of mismatch for existing methods.
2. Hilbert-Guided Framework: AOSNET introduces a fully differentiable mechanism to extract and align local oscillatory states (envelope, phase, frequency) without requiring predefined periods.
3. Learnable Global Prior: The global prior acts as a flexible reference interpreted through local dynamics rather than a rigid repeated template.
4. Empirical Validation: The method achieves state-of-the-art (SOTA) or highly competitive accuracy on eight standard benchmarks (including ETT, Electricity, Traffic, Weather) and two cloud workload traces, with significantly faster inference speeds compared to complex Transformer-based baselines.

4. Experimental Results

• Benchmark Performance: AOSNET achieved the best or second-best MSE on 7 out of 8 datasets and the best MAE on 7 out of 8. It consistently outperformed fixed-template baselines (TQNet, CycleNet), particularly on datasets with complex dynamics like Electricity and ETTh2.
• Non-Stationary Robustness: Controlled synthetic experiments (isolating amplitude, phase, and frequency variations) demonstrated that AOSNET's advantage grows as non-stationarity intensifies. While fixed-template methods degraded significantly under phase drift and frequency variation, AOSNET maintained stable performance.
• Efficiency: AOSNET is highly efficient, requiring only 1.45M parameters and achieving an inference speed of 9.15 ms/batch (1.4× faster than TQNet, 7.9× faster than PatchTST).
• Ablation Studies: Removing the Adaptive Oscillatory State (AOS) module or the Global Prior resulted in significant performance drops (up to +15.5% MSE degradation), confirming the necessity of the alignment mechanism.
• Internal Analysis: Visualization of the learned prior confirmed it is oscillatory but not rigidly periodic. The adaptive gate was shown to respond primarily to instantaneous frequency and phase descriptors, validating the state-alignment hypothesis.

5. Significance and Claims

The paper claims that adaptive oscillatory-state alignment is a superior inductive bias for long-term forecasting in non-stationary environments. The core insight is that "same cycle phase does not imply same oscillatory state." By moving away from rigid period indexing and instead aligning local dynamical states (amplitude, phase, frequency) against a flexible reference, AOSNET effectively handles the amplitude modulation, phase drift, and frequency variations inherent in real-world data.

The authors position AOSNET as a lightweight, fully differentiable solution that bridges the gap between signal processing (Hilbert transform) and deep learning, offering a robust alternative to methods that assume globally stable periodic patterns.

6. Limitations

The paper acknowledges several limitations:

• Scope: The current study focuses on point forecasting with a compact backbone; the benefits in larger Transformer architectures or probabilistic forecasting settings remain unexplored.
• Resolution: The Hilbert descriptors are computed on fixed look-back windows, which may be less suitable for multi-resolution structures or irregularly sampled data.
• Prior Conditioning: The global prior is shared across all samples. Future work could explore making the prior conditional on covariates, regimes, or cluster-level states.
• Evaluation: Experiments focus on standard benchmarks and controlled non-stationary oscillations; broader evaluation on datasets with abrupt regime changes, missing values, or exogenous events is needed.