This paper challenges the effectiveness of Transformer-based solutions for long-term time series forecasting (LTSF). While Transformers have shown success in various tasks, they may not be suitable for LTSF due to their permutation-invariant self-attention mechanism, which can lead to temporal information loss. The authors introduce LTSF-Linear, a simple one-layer linear model, which outperforms existing Transformer-based models on nine real-world datasets. LTSF-Linear achieves significant improvements in forecasting accuracy, often by a large margin. The study also highlights that existing Transformer-based models fail to extract temporal relations effectively, as their performance does not improve with larger look-back windows. The results suggest that Transformers may overfit to temporal noise rather than capturing true temporal patterns. The paper advocates for a re-evaluation of Transformer-based solutions for other time series analysis tasks. The findings indicate that while Transformers are not inherently ineffective, their performance in LTSF may be overstated, and simpler models like LTSF-Linear can serve as effective baselines. The study also explores the impact of various design elements in Transformer-based models, including positional encoding, sub-series embedding, and decoding strategies. Overall, the paper calls for a more critical examination of the effectiveness of Transformer-based approaches in time series forecasting.This paper challenges the effectiveness of Transformer-based solutions for long-term time series forecasting (LTSF). While Transformers have shown success in various tasks, they may not be suitable for LTSF due to their permutation-invariant self-attention mechanism, which can lead to temporal information loss. The authors introduce LTSF-Linear, a simple one-layer linear model, which outperforms existing Transformer-based models on nine real-world datasets. LTSF-Linear achieves significant improvements in forecasting accuracy, often by a large margin. The study also highlights that existing Transformer-based models fail to extract temporal relations effectively, as their performance does not improve with larger look-back windows. The results suggest that Transformers may overfit to temporal noise rather than capturing true temporal patterns. The paper advocates for a re-evaluation of Transformer-based solutions for other time series analysis tasks. The findings indicate that while Transformers are not inherently ineffective, their performance in LTSF may be overstated, and simpler models like LTSF-Linear can serve as effective baselines. The study also explores the impact of various design elements in Transformer-based models, including positional encoding, sub-series embedding, and decoding strategies. Overall, the paper calls for a more critical examination of the effectiveness of Transformer-based approaches in time series forecasting.