Research map / Time series / Long-term forecasting
TimePro: Efficient Multivariate Long-term Time Series Forecasting with Variable- and Time-Aware Hyper-state
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 TimePro, a Mamba-based multivariate long-term forecaster whose HyperMamba module reconstructs variable- and time-aware hyper-states by adaptively sampling important time points from the variable-scan state.
Problem
Long-term multivariate time series forecasting suffers from the multi-delay issue: different variables affect the target over different time lags, but existing models (Mamba and Transformer variants) process all variables or time points uniformly, limiting variable-relationship and temporal represen
Approach
Scan along the variable dimension to get a plain state, then learn offsets to adaptively sample key time points within each variable (deformable-style sampling with linear interpolation) and fuse them to tune the state into a time- and variable-aware hyper-state, all with linear complexity in the number of variables.
Key results
- TimePro achieves average MSE 0.169 on ECL and 0.352 on Exchange with MAE 0.262 and 0.399, claiming 12 firsts and 2 seconds across 16 metrics in the main table.
- On ETTh1 average MSE 0.438 vs 0.449 for SOFTS and 0.454 for iTransformer (L=96).
- Uses only 67% of S-Mamba's parameters and 78% of its GFLOPs, with linear complexity O(NL) versus iTransformer's O(N^2+NL).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaptive time tuning instead of a linear projection to integrate variable states (ablation) | TimePro · Exchange | MSE: 0.360 -> 0.352 | Helped |
| use adaptive time tuning instead of a linear projection to integrate variable states (ablation) | TimePro · ETTh1 | MSE: 0.451 -> 0.438 | Helped |
| remove the depthwise convolution before scanning (ablation) | HyperMamba · Exchange | MSE: 0.358 -> 0.358 | No clear effect |
| remove the linear projection after scanning (ablation) | HyperMamba · Exchange | MSE: 0.358 -> 0.356 | Helped |
| replace selective scan with Hyper-Scan (ablation) | Mamba + Hyper-Scan · Exchange | MSE: 0.374 -> 0.358 | Helped |
| add adaptive variable-and-time scanning instead of variable-only scanning (ablation) | Hyper-Scan · Exchange | MSE: 0.367 -> 0.352 | Helped |
| add adaptive variable-and-time scanning instead of time-only scanning (ablation) | Hyper-Scan · Exchange | MSE: 0.367 -> 0.352 | Helped |
| add adaptive variable-and-time scanning instead of scanning variable then time (ablation) | Hyper-Scan · ETTh1 | MSE: 0.450 -> 0.438 | Helped |
| increase feature dimension (ablation) | TimePro · most datasets | forecasting performance | Mixed |
| increase the number of encoder layers (ablation) | TimePro · Weather | MAE | No clear effect |
| increase patch length from 8 to 32 (ablation) | TimePro · Exchange | MAE | No clear effect |
Surprising findings
- Its MSE on Weather (0.251) ties S-Mamba and the paper reports only modest gains over SOFTS on Solar-Energy, where SOFTS is better, indicating the gains are not uniform across datasets.
Limitations
- Gains over strongest baselines are small and mixed (e.g., Solar-Energy and some ETTh2 horizons favor SOFTS or S-Mamba).
- Results are reported on standard benchmarks with fixed lookback L=96 in the main table; the number of sampled points M=9 is fixed as a constant.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather ECL Exchange Solar-Energy
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather ECL Exchange Solar-Energy
Metrics: MSE MAE
Baselines: S-Mamba SOFTS iTransformer PatchTST Crossformer TiDE TimesNet DLinear SCINet Stationary
Builds on: Mamba PatchTST iTransformer RevIN S-Mamba
Closest papers by meaning
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.87)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.87)
- Timer-XL: Long-Context Transformers for Unified Time Series Forecasting (ICLR 2025 · similarity 0.87)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.87)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.87)
- TStruct: Learning Shared Temporal Structures for Long-Term Time-Series Forecasting (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.