Are Transformers Effective for Time Series Forecasting?

Recently, there has been a surge of Transformer-based solutions for the long-term time series forecasting (LTSF) task. Despite growing performance over past few years, we question validity this line research in work. Specifically, Transformers is arguably most successful solution to extract semantic correlations among elements long sequence. However, modeling, are temporal relations an ordered set continuous points. While employing positional encoding and using tokens embed sub-series facilitate preserving some ordering information, nature permutation-invariant self-attention mechanism inevitably results information loss. To validate our claim, introduce embarrassingly simple one-layer linear models named LTSF-Linear comparison. Experimental on nine real-life datasets show that surprisingly outperforms existing sophisticated LTSF all cases, often by large margin. Moreover, conduct comprehensive empirical studies explore impacts various design their relation extraction capability. We hope surprising finding opens up new directions also advocate revisiting other analysis tasks (e.g., anomaly detection) future.