一种基于Transformer的迁移学习方法及其在金融时序预测中的应用
A Transfer Learning Method Based on Transformer and Its Application in Financial Time-Series Forecasting
王旸1
作者信息
摘要
金融市场预测通常被认为是数据挖掘中最具挑战性的任务之一.最近Transformer模型在提高金融时序预测(Financial Time-Series Forecasting,FTSF)的精度方面取得了成功,但由于作为隐含的复杂信息,且可用的标记数据较少,当前基准在该领域的泛化能力较差.为缓解干净数据不足导致的过拟合问题,提出一种Transformer结合对抗域适应的深度迁移学习框架TADA-FTSF,用于金融领域TSF任务,以提高深度预测模型的可靠性与准确性.
Abstract
Financial market forecasting is generally considered to be one of the most challenging tasks in data mining.Transformer model has recently been successful in improving the accuracy of Financial Time-Series Forecasting(FTSF).However,due to the complex information as implicit and the small amount of labeled data available,current benchmarks have poor generalization ability in this field.In order to alleviate the overfitting problem caused by the lack of clean data,a deep Transfer Learning framework named TADA-FTSF which combines with Transformer and adversarial domain adaptation is proposed for TSF tasks in the financial field,to improve the reliability and accuracy of the deep forecasting model.
关键词
金融时序预测/Transformer模型/域适应/迁移学习Key words
Financial Time-Series Forecasting/Transformer model/domain adaptation/Transfer Learning引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据