search

BennyTMT / LLMsForTimeSeries

54 stars 7 forks source link

I am confused #4

Closed nightforwar closed 4 weeks ago

nightforwar commented 1 month ago

Thank you for the article. I have reproduced some of the results, and I plan to present this paper at the group meeting next week. However, I have some questions. After reading it, I feel a bit confused. It seems that after removing the LLM, the model is just a simple linear layer, yet the prediction performance is even better. It even outperforms some more complex architectures, such as Transformers or MLP-based models like TimesNet. This is quite difficult to understand. Why hasn’t anyone used a simple linear layer to achieve these results before? Thanks to the authors!

BennyTMT commented 1 month ago

There is a highly influential paper that might be helpful to you. "Are Transformers Effective for Time Series Forecasting?"

From my personal experience and experiments, Patching is also a very useful technique and a key contributor to TSF performance. For more details Please refer to the final section in our paper.

Thanks for the interest in the repo and paper, Good luck with your presentation : )