时间序列预测是金融领域的重要研究课题,对提高决策效率和风险管理具有重要意义。本文提出并构建了多种基于Transformer架构的时间序列预测模型,包括Vanilla Transformer、Informer、FEDformer、Autoformer、PatchTST和TimesNet,并以浦...时间序列预测是金融领域的重要研究课题,对提高决策效率和风险管理具有重要意义。本文提出并构建了多种基于Transformer架构的时间序列预测模型,包括Vanilla Transformer、Informer、FEDformer、Autoformer、PatchTST和TimesNet,并以浦发银行股票收盘价作为预测对象,探讨了单变量和多变量输入下这些模型的预测性能。实验结果表明,基于Transformer架构的模型在预测精度和稳定性方面显著优于传统模型(ARIMA和VAR)。其中Vanilla Transformer在单变量预测中表现最佳(MSE = 0.0455, MAE = 0.1480),而TimesNet在多变量预测中性能最优(MSE = 0.0463, MAE = 0.1513)。研究表明,基于Transformer的模型在处理复杂时间序列问题时具有显著优势,为金融时间序列预测提供了新的方法和参考。未来可进一步探索Transformer与其他深度学习技术的结合,以提升模型的泛化能力和实用性。Time series forecasting is an important research topic in the financial field, with significant implications for improving decision-making efficiency and risk management. This paper proposes and constructs several time series forecasting models based on the Transformer architecture, including Vanilla Transformer, Informer, FEDformer, Autoformer, PatchTST, and TimesNet. Using the closing price of Shanghai Pudong Development Bank stock as the forecasting target, the paper examines the predictive performance of these models under univariate and multivariate inputs. Experimental results show that Transformer-based models significantly outperform traditional models (ARIMA and VAR) in terms of prediction accuracy and stability. Among them, the Vanilla Transformer performs the best in univariate forecasting (MSE = 0.0455, MAE = 0.1480), while TimesNet has the best performance in multivariate forecasting (MSE = 0.0463, MAE = 0.1513). The study demonstrates that Transformer-based models have a significant advantage in handling complex time series problems, providing new methods and references for financial time series forecasting. Future research can further explore the combination of Transformer with other deep learning technologies to enhance the model’s generalization ability and practical utility.展开更多
文摘时间序列预测是金融领域的重要研究课题,对提高决策效率和风险管理具有重要意义。本文提出并构建了多种基于Transformer架构的时间序列预测模型,包括Vanilla Transformer、Informer、FEDformer、Autoformer、PatchTST和TimesNet,并以浦发银行股票收盘价作为预测对象,探讨了单变量和多变量输入下这些模型的预测性能。实验结果表明,基于Transformer架构的模型在预测精度和稳定性方面显著优于传统模型(ARIMA和VAR)。其中Vanilla Transformer在单变量预测中表现最佳(MSE = 0.0455, MAE = 0.1480),而TimesNet在多变量预测中性能最优(MSE = 0.0463, MAE = 0.1513)。研究表明,基于Transformer的模型在处理复杂时间序列问题时具有显著优势,为金融时间序列预测提供了新的方法和参考。未来可进一步探索Transformer与其他深度学习技术的结合,以提升模型的泛化能力和实用性。Time series forecasting is an important research topic in the financial field, with significant implications for improving decision-making efficiency and risk management. This paper proposes and constructs several time series forecasting models based on the Transformer architecture, including Vanilla Transformer, Informer, FEDformer, Autoformer, PatchTST, and TimesNet. Using the closing price of Shanghai Pudong Development Bank stock as the forecasting target, the paper examines the predictive performance of these models under univariate and multivariate inputs. Experimental results show that Transformer-based models significantly outperform traditional models (ARIMA and VAR) in terms of prediction accuracy and stability. Among them, the Vanilla Transformer performs the best in univariate forecasting (MSE = 0.0455, MAE = 0.1480), while TimesNet has the best performance in multivariate forecasting (MSE = 0.0463, MAE = 0.1513). The study demonstrates that Transformer-based models have a significant advantage in handling complex time series problems, providing new methods and references for financial time series forecasting. Future research can further explore the combination of Transformer with other deep learning technologies to enhance the model’s generalization ability and practical utility.
