TL;DR
AsyTO breaks the classic trade-off between parameter efficiency and model flexibility in multivariate time series forecasting by compressing the forecasting operator rather than the data series. By decoupling history-reading and future-writing into distinct, shared temporal factors with variable-specific gains, it achieves state-of-the-art accuracy among lightweight models while scaling linearly with the number of variables.
What Problem Does This Paper Address?
When forecasting multivariate time series, researchers face a structural dilemma. On one hand, sharing a single temporal predictor across all variables is highly parameter-efficient but forces completely different variables (e.g., temperature vs. humidity) to follow an identical history-to-future mapping. On the other hand, learning an independent predictor for each variable preserves unique behaviors but causes the parameter count to explode quadratically with the product of the variable count, context length, and forecast horizon ($O(C \times L \times H)$).
Existing lightweight models try to mitigate this by compressing the time-series data using spectral methods or downsampling. However, they typically use the same mathematical object (like a symmetric kernel or basis set) to handle both the past and the future. The authors argue that this structural symmetry is fundamentally flawed: reading historical evidence is an extraction problem that focuses on specific informative lags, whereas writing a future forecast requires coordinating an entire trajectory.
Methods
AsyTO (Asymmetric Temporal Operator) resolves this dilemma through a multi-stage architecture that isolates periodic trends and factorizes the remaining residual mapping:
- Low-Rank Periodic Prototype ($Q = UV^\top$): The model first extracts recurring seasonal patterns shared across variables, leaving a periodicity-exclusive residual.
- Asymmetric Temporal Operator (AsyTO): Instead of materializing a massive dense operator tensor, AsyTO uses CANDECOMP/PARAFAC (CP) factorization to split the operator into three distinct components:
- History-side factors ($\Psi$): Shared across variables to read historical patterns.
- Future-side factors ($\Phi$): Shared across variables to write the future trajectory. Crucially, $\Psi \neq \Phi$, allowing for history-future asymmetry.
- Variable-mode response ($G$): Variable-specific gains that dictate how strongly each variable responds to a given temporal mode.
- Cycle-Separable Temporal Factorization (CeST): To prevent the shared temporal factors from dominating the parameter budget, CeST factorizes them into a Kronecker product of a cycle-index factor (how many cycles ago) and a within-cycle phase factor (where in the cycle).
- Residual Echo: A simple gating mechanism that passes a phase-conditioned copy of the observation from exactly one cycle prior as a direct local cue.
Datasets
The model was evaluated across 11 standard multivariate time-series benchmarks:
- ETT Family: ETTh1, ETTh2, ETTm1, ETTm2 (Electricity Transformer Temperature)
- Electricity
- Weather
- Traffic
- PEMS Datasets: PEMS03, PEMS04, PEMS07, PEMS08 (Public California transportation data)
No new datasets were introduced in this paper.
Key Insights and Ideas
- The Necessity of Asymmetry: Forcing history-reading and future-writing operators to be symmetric ($\Psi = \Phi$) consistently harms performance. The paper demonstrates that history modes tend to be highly localized in time, whereas future modes spread smoothly across the forecasting horizon.
- Compressing Operators over Series: Instead of trying to compress the input data streams, compressing the underlying forecasting operator preserves the target-only property (predicting a variable solely from its own history) while keeping compute requirements minimal.
- Plug-and-Play Transferability: Remarkably, the low-rank periodic prototype and CeST modules can be attached to frozen weights of other independent lightweight backbones (like PhaseFormer or MixLinear). Doing so yields significant error reductions (up to 68.4% on PEMS datasets) without updating any backbone weights, proving that these modules provide a universally transferable structural correction.
Key Results
AsyTO achieves the best lightweight error in 30 out of 44 dataset-horizon settings, firmly establishing itself on the accuracy-compute Pareto frontier.
- Compute and Parameter Efficiency: On the Electricity dataset, AsyTO uses 47.5×, 62.2×, and 590.4× fewer MACs than TQNet, iTransformer, and PatchTST, respectively. Its parameter footprint is 155× smaller than iTransformer.
- Linear Scaling: Because the temporal factors are shared, adding a new variable to the model configuration scales linearly, costing only a handful of parameters ($Rp$) to update the prototype coordinate map.
Limitations and Future Work
The primary limitation of AsyTO stems from its channel-independent design. Each variable is predicted strictly from its own historical context. While this keeps the model lightweight and prevents overfitting, it fails on datasets where cross-variable information sharing is vital.
For instance, AsyTO consistently underperformed on the Traffic dataset (which contains 862 highly correlated sensors). The authors acknowledge this boundary and suggest that incorporating selective cross-variable interaction is the most natural next step for future iterations of the architecture.
Summary updated