Multi-task end-to-end portfolio optimization with self-supervised graph attention network

End-to-end portfolio methods have attracted considerable attention owing to their capacity to simultaneously optimize the prediction and decision-making processes. However, existing methods mostly rely on single-task learning, which is insufficient to accommodate multi-dimensional investment objectives. In addition, the capture of correlations among stocks and sectors typically relies on static prior knowledge, failing to adequately incorporate their dynamic evolution. Therefore, this paper proposes a portfolio optimization model based on self-supervised graph attention network within a multi-task end-to-end framework. By integrating dynamic correlation capturing, multi-task prediction, and portfolio optimization, this model enables the dynamic and adaptive adjustment of return and risk weighting parameters, thereby automatically optimizing investment allocation in response to varying market conditions. Empirical results on the S&P 500 stocks and the robustness test on the CSI 300 stocks demonstrate that the proposed model achieves improved portfolio performance relative to existing representative models across key evaluation metrics, including cumulative return, Sharpe ratio, and maximum drawdown. These results are essential for enhancing decision-making efficiency and facilitating the development of adaptive quantitative investment strategies.