Research map / Time series / Long-term forecasting
xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories
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 xLSTM-Mixer, combining a shared NLinear initial forecast, sLSTM-based refinement over variate tokens with learnable initial embedding, and multi-view (forward/reversed) mixing, evaluated on long-term multivariate forecasting benchmarks with low memory.
Problem
Long-term multivariate time series forecasting needs accurate models that capture temporal and cross-variate dependencies while remaining memory- and compute-efficient, as Transformers scale quadratically.
Approach
Combine a channel-independent linear forecast shared across variates with refinement by a stack of sLSTM blocks that stride over variate tokens (each token holding a full time series), plus a learnable initial token and multi-view forecasting from original and reversed variate order, fused by a linear view-mixing layer.
Key results
- xLSTM-Mixer achieves best MSE in 11 of 28 and best MAE in 16 of 28 long-term forecasting settings against recent baselines.
- On GIFT-Eval, xLSTM-Mixer ranks 2nd overall by CRPS (0.510) and is the top purely supervised model.
- Ablation shows removing the sLSTM (using LSTM) raises MSE by 7.0% and variate recurrence removal raises MSE by 4.7%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use xLSTM-Mixer rather than competing methods | long-term forecasting benchmarks · Weather, Electricity, Traffic, ETTh1, ETTh2, ETTm1, ETTm2 | MSE and MAE: best results in 11 out of 28 cases for MSE and 16 out of 28 cases for MAE | Mixed |
| Use xLSTM-Mixer for probabilistic forecasting | GIFT-Eval · GIFT-Eval | CRPS: 0.498 -> 0.510 | Mixed |
| Use sLSTM instead of mLSTM, LSTM, or GRU blocks (ablation) | xLSTM-Mixer · Weather and ETTm1 forecasting | MSE and MAE: omitting xLSTM (only having LSTM, #3) raises the MAE by 6.2%/the MSE by 7.0% | Helped |
| Recur over variates rather than time (ablation) | xLSTM-Mixer · Weather and ETTm1 forecasting | MSE and MAE: variate recurrence (#5) by 4.3%/4.7% | Helped |
| Add a learned initial embedding token (ablation) | xLSTM-Mixer · Weather and ETTm1 forecasting | MSE and MAE: learning initial token embedding (#7) by 0.4%/0.7% | Helped |
| Add multi-view mixing (ablation) | xLSTM-Mixer · Weather and ETTm1 forecasting | MSE and MAE: view mixing (#8) by 0.6%/0.7% | Helped |
| Add NLinear time mixing (ablation) | xLSTM-Mixer · Weather and ETTm1 forecasting | MSE and MAE: time mixing (#11) by 2.7%/3.1% | Helped |
| Use DLinear rather than NLinear for time mixing (ablation) | xLSTM-Mixer · Weather and ETTm1 forecasting | MSE and MAE | Mixed |
| Add more than two randomly permuted views (ablation) | xLSTM-Mixer · multivariate forecasting | forecasting accuracy | No clear effect |
Surprising findings
- Reported xLSTMTime results could not be replicated with the official implementation or paper details, per the authors.
- Ensembling over more than two variate orderings did not improve accuracy.
Limitations
- Assumes all variates are sampled on a uniform time grid; irregular or missing timestamps need preprocessing.
- Striding over variates ties runtime and memory to the number of channels, a bottleneck for very high-dimensional series.
- Multi-view mixing makes attribution of temporal versus cross-channel contributions difficult.
Experimental setup
Datasets: Weather Electricity Traffic ETTh1 ETTh2 ETTm1 ETTm2 GIFT-Eval
Benchmarks: Weather Electricity Traffic ETTh1 ETTh2 ETTm1 ETTm2 GIFT-Eval
Metrics: MSE MAE MASE CRPS
Baselines: xLSTMTime LSTM TimeMixer++ TimeMixer TSMixer CycleNet DLinear TiDE S-Mamba Chimera PatchTST iTransformer
Builds on: xLSTM xLSTMTime NLinear RevIN MLP-Mixer
Closest papers by meaning
- TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · similarity 0.89)
- WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.89)
- TiRex-2: Generalizing TiRex to Multivariate Data and Streaming (NeurIPS 2026 · similarity 0.89)
- U-Mixer: An Unet-Mixer Architecture with Stationarity Correction for Time Series Forecasting (AAAI 2024 · similarity 0.89)
- Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series (NeurIPS 2024 · similarity 0.89)
- WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.89)
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.