Research map / Time series / Long-term forecasting
Unlocking the Power of LSTM for Long Term Time Series Forecasting
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 P-sLSTM, an sLSTM-based long-term time series forecaster that adds patching and channel independence to counter sLSTM's short-memory behavior, reporting competitive or SOTA results on five multivariate datasets.
Problem
LSTM and sLSTM struggle to capture long-term dependencies in multivariate long-term time series forecasting; the paper argues sLSTM's exponential gating does not guarantee long memory.
Approach
Builds on sLSTM by applying patching to each univariate channel (channel independence with a shared backbone), so the recurrent model processes shorter patch sequences, and provides a Markov-chain / geometric-ergodicity analysis of sLSTM memory.
Key results
- P-sLSTM ranks first in 23 and second in 10 of 40 settings across five datasets (Weather, Electricity, Solar, ETTm1, PEMS03).
- On Weather with horizon 96, P-sLSTM achieves MSE 0.149 / MAE 0.208 versus sLSTM 0.172 / 0.261 and LSTM 0.236 / 0.304.
- Channel-independent P-sLSTM shows higher training error but lower validation and test error than the channel-mixing variant on Weather, supporting overfitting reduction; P-sLSTM trains faster than iTransformer (52.54 s vs 79.12 s on Weather).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add patching and channel independence to sLSTM | sLSTM · Weather, horizon 96 | MSE: 0.172 -> 0.149 | Helped |
| add patching and channel independence to sLSTM | sLSTM · Electricity, horizon 96 | MSE: 0.281 -> 0.130 | Helped |
| add patching and channel independence to sLSTM | sLSTM · PEMS03, horizon 12 | MSE: 0.103 -> 0.077 | Hurt |
| remove memory mixing (ablation) | P-sLSTM · Weather, horizon 96, look-back 336 | MSE: 0.149 -> 0.151 | Hurt |
| remove memory mixing (ablation) | P-sLSTM · Weather, horizon 96, look-back 336 | MAE: 0.208 -> 0.209 | Hurt |
| remove memory mixing (ablation) | P-sLSTM · Weather, horizon 192, look-back 336 | MSE: 0.197 -> 0.201 | Hurt |
| remove memory mixing (ablation) | P-sLSTM · Weather, horizon 336, look-back 336 | MAE: 0.297 -> 0.306 | Hurt |
| remove memory mixing (ablation) | P-sLSTM · Weather, horizon 720, look-back 336 | MSE: 0.320 -> 0.320 | No clear effect |
| replace channel independence with channel mixing (ablation) | P-sLSTM · Weather, horizons 96, 192, 336, 720 | test MSE: 0.167 -> 0.227; 0.180 -> 0.234; 0.190 -> 0.265; 0.196 -> 0.260 | Hurt |
| increase patch size (ablation) | P-sLSTM · Weather | MSE | Mixed |
| increase look-back window using patching (ablation) | P-sLSTM · Weather, horizon 720 | MSE | Helped |
Surprising findings
- Channel independence lowers test error despite higher training error, which the authors attribute to reduced overfitting.
- P-sLSTM does not perform well on PEMS03, which the authors attribute to noise without a denoising mechanism.
Limitations
- Performance is not uniformly best; P-sLSTM underperforms on PEMS03 and is weaker than iTransformer in some settings.
- Optimal patch size varies by dataset and must be found by trial and error.
- Theoretical analysis relies on assumptions (e.g., no external input, bounded forget gate) that may not hold in practice; the paper notes gradient clipping implicitly bounds the forget gate.
Experimental setup
Datasets: Weather Electricity Solar ETTm1 PEMS03
Benchmarks: Weather Electricity Solar ETTm1 PEMS03
Metrics: MSE MAE
Baselines: sLSTM LSTM iTransformer FEDformer Crossformer RLinear DLinear TimesNet Mamba
Builds on: xLSTM sLSTM PatchTST LSTM
Closest papers by meaning
- xLSTM: Extended Long Short-Term Memory (NeurIPS 2024 · similarity 0.86)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.86)
- PGN: The RNN's New Successor is Effective for Long-Range Time Series Forecasting (NeurIPS 2024 · similarity 0.86)
- xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories (NeurIPS 2025 · similarity 0.86)
- Sequence Complementor: Complementing Transformers for Time Series Forecasting with Learnable Sequences (AAAI 2025 · similarity 0.85)
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.85)
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.