Are Transformers Effective for Time Series? (DLinear / NLinear)

Proved that a simple single-layer linear model decomposing time series into trend and seasonal components outperforms complex multi-head attention Transformer architectures.

Serves as qualified reference for implementing task-time-series-forecasting in production systems.

Assumptions

• Underlying spatio-temporal continuity and domain distributional stability hold

Limitations

• Performance scaling and computational footprint depend on receptive field depth, sequence length, and resolution