ScholarGate
Get started
Machine learningDeep learningTime-series forecastingAlgorithm

TSMixer: All-MLP Architecture for Time Series Forecasting

Also known as: All-MLP Time Series Mixer, Time Series Mixer, TSMixer Forecasting Model, Zaman Serisi Karıştırıcı

OriginatorSi-An Chen et al. (Google)Year2023Sources1Related methods8

TSMixer is a multivariate time-series forecasting model introduced by Si-An Chen and colleagues at Google in 2023. It challenges the prevailing dominance of Transformer-based architectures by demonstrating that a simple stack of interleaved MLP layers — alternating between mixing along the time axis and mixing across feature channels — achieves strong forecasting accuracy while remaining computationally efficient and easy to interpret architecturally.

Key highlights

  • Achieves competitive accuracy on multivariate benchmarks using only MLP layers, without attention or recurrence
  • Computationally efficient: linear complexity in time and features, fast training and inference
  • Explicitly models both temporal dynamics and cross-variate correlations through alternating mixing
  • Simple, interpretable architecture that is easy to implement and extend

Intuition

This section is available to Pro members. Upgrade to Pro

How it works

This section is available to Pro members. Upgrade to Pro

When to use it

Use TSMixer for multivariate long-horizon forecasting tasks where inter-variable dependencies carry predictive value. It suits scenarios with regular sampling, stable feature sets, and sufficient historical data. The model is appropriate when computational efficiency and scalability matter. It is less suited for univariate-only problems (where simpler linear models may suffice), irregularly sampled data, or tasks where temporal attention patterns are highly irregular. Alternatives include DLinear for purely linear baselines and PatchTST or Autoformer for attention-based approaches.

Strengths & limitations

Strengths
  • Achieves competitive accuracy on multivariate benchmarks using only MLP layers, without attention or recurrence
  • Computationally efficient: linear complexity in time and features, fast training and inference
  • Explicitly models both temporal dynamics and cross-variate correlations through alternating mixing
  • Simple, interpretable architecture that is easy to implement and extend
Limitations
  • May underperform on univariate or weakly correlated multivariate tasks where cross-feature mixing provides little benefit
  • The fixed look-back window makes it less flexible for variable-length contexts compared to attention-based models
  • Lacks explicit inductive biases for periodicity or trend decomposition that domain-specific models provide
  • Performance is sensitive to the choice of look-back window length and number of Mixer Blocks

Common pitfalls

This section is available to Pro members. Upgrade to Pro

Applications

This section is available to Pro members. Upgrade to Pro

Frequently asked

How does TSMixer differ from DLinear?

DLinear decomposes a series into trend and remainder components and applies separate linear layers to each, operating purely in the time dimension. TSMixer adds a second mixing dimension over features (channels), allowing it to capture cross-variate correlations through alternating temporal and feature MLP blocks. TSMixer is generally more expressive on multivariate tasks.

Does TSMixer use any attention mechanism?

No. TSMixer is an all-MLP architecture. It replaces self-attention with simple shared linear layers applied alternately along the time and feature axes. This design choice reduces computational complexity while maintaining competitive accuracy on standard multivariate forecasting benchmarks, as demonstrated in the original paper.

What look-back window length should I use with TSMixer?

The original paper evaluates multiple look-back window lengths and finds that longer windows generally improve performance on benchmarks with longer patterns. Common choices range from 96 to 512 time steps. The optimal length depends on the dominant temporal patterns in your data and should be validated empirically on a held-out validation split.

Sources

  1. 1.
    Chen, S.-A., Li, C.-L., Yoder, N., Arik, S. O., & Pfister, T. (2023). TSMixer: An all-MLP architecture for time series forecasting. Transactions on Machine Learning Research.
    https://arxiv.org/abs/2303.06053

You have read it. What now?

Ask about this methodSimilar methods

Cite this page

ScholarGate. (2026, June 2). TSMixer. ScholarGate. https://scholargate.app/deep-learning/tsmixer

TSMixer: All-MLP Architecture for Time Series Forecasting | ScholarGate