DLinear: Decomposition Linear Model for Time Series Forecasting
DLinear (Decomposition Linear Model for Forecasting) · Also known as: Decomposition Linear, DLinear Forecaster, Linear Decomposition Model, Ayrışım Doğrusal Modeli
DLinear is a lightweight time series forecasting model introduced by Zeng et al. at AAAI 2023. It challenges the prevailing assumption that Transformer-based architectures are necessary for accurate long-horizon forecasting. The model decomposes an input sequence into trend and seasonal components using a moving average filter, then applies separate single-layer linear transformations to each component before summing their outputs to produce the final forecast.
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
DLinear is well-suited for long-horizon univariate or channel-independent multivariate forecasting tasks where computational efficiency and interpretability matter. It works best when the time series exhibits a separable trend-seasonal structure. It is less appropriate when cross-variate interactions are essential, when the series lacks clear decomposable structure, or when probabilistic forecasts are required. Consider PatchTST or TSMixer when richer temporal representations or channel-mixing are needed.
Strengths & limitations
- Extremely lightweight architecture with far fewer parameters than Transformer-based models, enabling fast training and inference
- Decomposition step provides interpretable separation of trend and seasonal dynamics
- Competitive or superior long-horizon forecasting accuracy relative to many Transformer variants on standard benchmarks
- Channel-independent design avoids overfitting caused by spurious cross-variate correlations in high-dimensional data
- Cannot model non-linear temporal dependencies or complex interactions between variates
- Single-layer linear projection may underfit highly irregular or chaotic series
- Moving average decomposition assumes a relatively stationary seasonal period and may struggle with evolving seasonality
- Does not produce uncertainty estimates or probabilistic forecasts natively
Frequently asked
How does DLinear differ from a plain linear regression model?
Unlike a single linear regression applied to the raw series, DLinear first decomposes the input into trend and seasonal components using a moving average filter and then applies two separate linear layers to each component. This decomposition allows each linear layer to specialize on structurally different parts of the signal, improving forecast accuracy while retaining simplicity.
Is DLinear suitable for multivariate forecasting?
DLinear can be applied to multivariate data in a channel-independent manner, treating each variate as an independent univariate series. This avoids spurious cross-variate correlations that can harm generalization. However, it does not model inter-variable relationships, so tasks where cross-variate dynamics are critical may benefit from models like TSMixer.
What lookback window length should be used with DLinear?
The original paper explores multiple lookback window lengths, and performance can vary by dataset. A common practical choice is 336 or 512 time steps for long-horizon benchmarks, but the optimal length should be tuned via cross-validation on the target dataset. Longer windows generally help the trend decomposition but increase memory usage linearly.
Sources
- Zeng, A., Chen, M., Zhang, L., & Xu, Q. (2023). Are transformers effective for time series forecasting? AAAI. link ↗
How to cite this page
ScholarGate. (2026, June 2). DLinear (Decomposition Linear Model for Forecasting). ScholarGate. https://scholargate.app/en/deep-learning/dlinear
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.
Compare side by side →