LightTS: Light Sampling-oriented MLP for Multivariate Time-Series Forecasting
LightTS (Light Sampling-oriented MLP) · Also known as: Light Sampling-oriented MLP, LightMLP, Hafif Örnekleme Tabanlı MLP, Lightweight Time-Series MLP
LightTS is a lightweight, MLP-based architecture for multivariate time-series forecasting introduced by Tianping Zhang and colleagues in 2022. Motivated by the observation that simpler models can match or surpass heavy Transformer-based architectures, LightTS applies an interval-sampling strategy to decompose long input sequences into multiple sub-sequences and processes each with compact Chunk-MLP and Continuous-MLP modules. The design prioritizes computational efficiency while preserving both local and global temporal patterns.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
LightTS is well-suited for multivariate long-horizon forecasting tasks where computational resources are limited or inference speed is critical, such as real-time monitoring or edge deployment. It assumes stationarity or slow distribution shift and works best when inter-variate correlations are moderate. It is less suitable when the series exhibit strong non-stationarity, abrupt change points, or require explicit probabilistic uncertainty quantification. Alternatives include DLinear for even simpler baselines, TSMixer for token-mixing variants, or PatchTST when patch-based attention is preferred.
Strengths & limitations
- Highly efficient: linear complexity in sequence length enables fast training and inference on long sequences
- Simple architecture with few hyperparameters, making it easy to tune and reproduce
- Two-stage MLP hierarchy captures both local chunk patterns and global sequence continuity
- Competitive forecasting accuracy relative to its computational footprint compared to heavier Transformer models
- Lacks explicit attention mechanisms, which may limit its ability to capture complex long-range dependencies
- Interval sampling can discard high-frequency temporal information if S is too large
- Does not natively model probabilistic uncertainty or provide prediction intervals
- Performance may degrade under strong distribution shift or highly non-stationary regimes without additional adaptation
Frequently asked
How does LightTS differ from a standard MLP applied directly to the time series?
A plain MLP flattens the entire input and has quadratic parameter growth with sequence length. LightTS uses interval sampling to shorten sub-sequences before applying two structured MLPs—Chunk-MLP for local patterns and Continuous-MLP for global patterns—keeping parameter counts and computation manageable even for long input windows.
Is LightTS suitable for univariate forecasting?
LightTS was designed and evaluated primarily in multivariate settings where N variates are modeled jointly. For purely univariate tasks, simpler linear or statistical models may be equally competitive and faster to train, though LightTS can technically be applied by setting N to one.
What is the role of the sampling interval S and how should it be chosen?
The sampling interval S controls the tradeoff between computational efficiency and temporal resolution. A small S preserves more temporal detail but increases sub-sequence length; a large S reduces computation but risks losing high-frequency structure. In practice S is treated as a hyperparameter tuned via validation set performance.
Sources
- Zhang, T., Zhang, Y., Cao, W., Bian, J., Yi, X., Zheng, S., & Li, J. (2022). Less is more: Fast multivariate time series forecasting with light sampling-oriented MLP structures. arXiv preprint. link ↗
How to cite this page
ScholarGate. (2026, June 2). LightTS (Light Sampling-oriented MLP). ScholarGate. https://scholargate.app/en/deep-learning/lightts
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- DLinearDeep learning↔ compare
- Multilayer PerceptronDeep learning↔ compare
- TSMixerDeep learning↔ compare