Research map / Time series / Long-term forecasting

TStruct: Learning Shared Temporal Structures for Long-Term Time-Series Forecasting

NeurIPS 2026 · Time series › Long-term forecasting · approach: Lightweight long-term forecasters

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TStruct, which learns a small set of shared temporal dependency matrices and mixes them with sample-variable-specific softmax weights (plus a low-rank channel-adaptive head) for long-term time-series forecasting.

Problem

Long-term time-series forecasting must model temporal dependencies that stay stable under noise and temporal distribution shift, while flexible attention tends to fit incidental interactions in low-SNR data.

Approach

Parameterize a set of K shared patch-level temporal dependency matrices and combine them per sample-variable using softmax weights from a static variable prior plus a dynamic router, yielding row-normalized aggregation weights instead of unconstrained attention; a PCAHead uses shared low-rank banks for variable-specific prediction heads.

Key results

• TStruct achieves the best average MSE on all seven benchmarks and best average MAE on six, ranking first in 13 of 14 dataset-metric comparisons (e.g., ETTh1 avg MSE 0.390 vs DUET 0.398).
• On Electricity TStruct reports average MSE 0.151 / MAE 0.244 and on Traffic 0.381 / 0.252 under the paper's protocol.
• Ablations show replacing the Temporal Structure Mixer with MHA or a shared linear mixer, or PCAHead with a linear head, degrades results; TSM is more robust to temporal shift and noise injection than MHA.

Surprising findings

• Replacing the temporal structure mixer with multi-head attention gives worse results and is more sensitive to temporally biased training data, with TSM trained on 40% of data outperforming the full-data MHA and Linear models on ETTh1.
• Electricity and Traffic show a smaller gap between TStruct and the shared linear mixer, suggesting simple shared mixing is already competitive there.

Limitations

• Gains on Traffic are mixed (second-best MAE), and the paper notes ablation gaps narrow on Electricity and Traffic.
• Baseline numbers are largely taken from original papers or released code with per-horizon look-back selection, so comparisons depend on those protocols.

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic

Metrics: MSE MAE

Baselines: PatchTST PhaseFormer DLinear CycleNet TQNet TimeEmb GTR DUET

Builds on: PatchTST RevIN DLinear CycleNet

Closest papers by meaning

• A Lightweight Sparse Interaction Network for Time Series Forecasting (AAAI 2025 · similarity 0.89)
• TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.88)
• SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (NeurIPS 2025 · similarity 0.87)
• TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.87)
• Sequence Complementor: Complementing Transformers for Time Series Forecasting with Learnable Sequences (AAAI 2025 · similarity 0.87)
• TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.87)