Scholarlink

<p>In this paper, we propose a Transformer-based model for multi-asset price prediction with long-term forecasting capabilities, integrating a pre-trained Anomaly Transformer (AT) for anomaly detection and a Decomposition-based Transformer for time-series forecasting. The proposed model addresses the challenges of noisy and anomalous input data, which are common in financial datasets. By embedding anomaly detection within the prediction architecture, our model enhances performance without distorting the data characteristics, as can occur in traditional pre-processing pipelines. We evaluate the model on a dataset containing 26 assets, including Foreign Exchange Majors, Commodities, US Stocks, Indices, and Cryptocurrencies. Experimental results demonstrate that AT + iTransformer outperforms baseline models, including Long Short-Term Memory (LSTM) and some latest transformer-based methods, across various performance metrics. In the time lag 6 and time lead 1 configuration, AT + iTransformer achieves the best results, with an <i>R</i><sup><b>2</b></sup> of 0.912, RMSE of 6.802, and MAE of 4.389. These findings highlight the effectiveness of the proposed approach for multi-asset price forecasting.</p>

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

Asset Price Prediction Using a Transformer-Based Model with Pre-Trained Anomaly Detection Feature Extraction

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • Bao Bui Quoc,
  • Quang Truong Dang,
  • Tai Nguyen Duc,
  • Anh Son Ta

摘要

In this paper, we propose a Transformer-based model for multi-asset price prediction with long-term forecasting capabilities, integrating a pre-trained Anomaly Transformer (AT) for anomaly detection and a Decomposition-based Transformer for time-series forecasting. The proposed model addresses the challenges of noisy and anomalous input data, which are common in financial datasets. By embedding anomaly detection within the prediction architecture, our model enhances performance without distorting the data characteristics, as can occur in traditional pre-processing pipelines. We evaluate the model on a dataset containing 26 assets, including Foreign Exchange Majors, Commodities, US Stocks, Indices, and Cryptocurrencies. Experimental results demonstrate that AT + iTransformer outperforms baseline models, including Long Short-Term Memory (LSTM) and some latest transformer-based methods, across various performance metrics. In the time lag 6 and time lead 1 configuration, AT + iTransformer achieves the best results, with an R2 of 0.912, RMSE of 6.802, and MAE of 4.389. These findings highlight the effectiveness of the proposed approach for multi-asset price forecasting.