Scholarlink

The accurate prediction of financial time series is critical in efficient portfolio management as well as development of trading strategies. This paper proposes CryptoGpt, a complete framework, which combines reversible instance normalization (RevIn), patch embedding, and a pretrained GPT-2 backbone for cryptocurrency daily prices prediction. Reversible instance normalization addresses the instability of a nonstationary data, and patch embedding captures local patterns and reduces sequence length. By fine-tuning only a small prediction head, CryptoGpt takes advantage of the rich contextual representations made available by a large language model without having to re-train it. Our proposed model competes with state-of-the-art transformer baselines, achieving lower MAPE, MAE, RMSE and higher R \(^{2}\) on several major cryptocurrencies, while maintaining competitive performance on the remaining ones. These results highlight the consideration of cross-domain pretraining in contrast to a feasible and correct method of univariate financial forecasting.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CryptoGpt: An LLM-Driven Transfer Learning Approach to Cryptocurrencies Time Series Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Amine Batsi,
  • Mohamed Biniz,
  • Imane Khattabi,
  • Ibtissam Chouklati,
  • Samir Boukil

摘要

The accurate prediction of financial time series is critical in efficient portfolio management as well as development of trading strategies. This paper proposes CryptoGpt, a complete framework, which combines reversible instance normalization (RevIn), patch embedding, and a pretrained GPT-2 backbone for cryptocurrency daily prices prediction. Reversible instance normalization addresses the instability of a nonstationary data, and patch embedding captures local patterns and reduces sequence length. By fine-tuning only a small prediction head, CryptoGpt takes advantage of the rich contextual representations made available by a large language model without having to re-train it. Our proposed model competes with state-of-the-art transformer baselines, achieving lower MAPE, MAE, RMSE and higher R \(^{2}\) on several major cryptocurrencies, while maintaining competitive performance on the remaining ones. These results highlight the consideration of cross-domain pretraining in contrast to a feasible and correct method of univariate financial forecasting.