TiDE: Time-series Dense Encoder
TiDE (Time-series Dense Encoder) · Also known as: Time-series Dense Encoder, TiDE model, Dense Encoder for Long-term Forecasting, Yoğun Kodlayıcı Zaman Serisi Modeli
TiDE (Time-series Dense Encoder) is an MLP-based encoder-decoder architecture for long-term multivariate time-series forecasting, introduced by Abhimanyu Das and colleagues at Google Research in 2023. The model encodes past time-series observations together with static and dynamic covariates through stacked dense (MLP) layers, then decodes a latent representation into future forecasts. TiDE demonstrates that simple linear and dense architectures can match or outperform Transformer-based models on standard long-term forecasting benchmarks while being significantly faster.
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
TiDE is well-suited for long-horizon multivariate time-series forecasting tasks where covariate information is available and computational efficiency matters. It performs best when the lookback window is long and future covariates (such as calendar features) are known in advance. TiDE assumes stationarity within segments and benefits from standard normalisation. It is less appropriate for very short sequences or tasks requiring explicit attention to sparse event-driven patterns. Alternatives include DLinear for ultra-lightweight baselines, TSMixer for channel-mixing MLP approaches, and Transformer variants (e.g., PatchTST) when long-range dependencies are highly irregular.
Strengths & limitations
- Competitive long-term forecasting accuracy with substantially lower computational cost than Transformer-based models
- Explicit integration of static and dynamic covariates through dedicated projection layers
- Residual connection from raw input enables direct modelling of linear trends and seasonal components
- Simple architecture is easy to implement, debug, and fine-tune without specialised hardware
- Flattening the lookback window loses local temporal structure that convolutional or attention layers might capture
- Performance can degrade when future covariate information is unavailable or unreliable
- Dense MLP layers scale quadratically with lookback length, limiting very long context windows
- May underperform Transformer models on datasets with complex, irregular long-range dependencies
Frequently asked
How does TiDE differ from DLinear if both avoid attention mechanisms?
DLinear uses a single linear layer per channel with a seasonal-trend decomposition, making it extremely lightweight but limited in expressiveness. TiDE adds stacked MLP encoder and decoder blocks plus explicit covariate projection, giving it greater capacity to model nonlinear interactions and leverage auxiliary features, at the cost of more parameters and slightly higher training time.
Does TiDE handle multivariate inputs by mixing channels?
By default TiDE follows a channel-independent strategy: each variate is encoded and decoded separately with shared weights, similar to PatchTST. This avoids overfitting cross-channel noise on smaller datasets but may miss inter-variable relationships. Channel-mixing variants can be constructed by concatenating all variates before encoding.
What normalisation strategy is recommended with TiDE?
The authors and subsequent practitioners recommend using instance normalisation such as Reversible Instance Normalisation (RevIN), which normalises each input sequence at inference time and reverses the transformation on the output. This mitigates non-stationarity and distribution shift between training and test windows, which is a common failure mode for dense MLP forecasters.
Sources
- Das, A., Kong, W., Leach, A., Mathur, S., Sen, R., & Yu, R. (2023). Long-term forecasting with TiDE: Time-series dense encoder. Transactions on Machine Learning Research. link ↗
How to cite this page
ScholarGate. (2026, June 2). TiDE (Time-series Dense Encoder). ScholarGate. https://scholargate.app/en/deep-learning/tide
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