Stock conditional drawdown at risk portfolio optimization based on gated bidirectional temporal convolution and discrete cosine graph neural networks on hypervariable graphs

With the development of machine learning technology, the application of stock prediction in financial portfolio optimization has become increasingly important. This study proposes an intelligent portfolio optimization method that combines gated bidirectional temporal convolution-discrete cosine graph neural network (TDGNN) with the mean-conditional drawdown at risk (Mean-CDaR) model, aiming to improve the risk-return performance of the portfolio. The method consists of two main stages: first, the data is converted into a hypervariable graph through the TDGNN model, the gated bidirectional temporal convolution layer is used to capture the temporal dynamic characteristics, and the discrete cosine graph neural network is combined to effectively model the complex spatiotemporal relationship in the stock market; second, the Mean-CDaR model is used for portfolio optimization, and the maximum drawdown is used as a measurement indicator to achieve precise risk control. Experimental results show that on the CSI 300, S&P500, and Nikkei 225 data sets, TDGNN and Mean-CDaR models perform significantly better than traditional methods, with $R^2$ of 0.9991, 0.9991, and 0.9983, respectively. Under the assumption of no transaction costs, the cumulative returns are 0.42, 0.62, and 0.93, respectively; considering 0.05 % transaction costs, the cumulative returns are 0.1, 0.25, and 0.49, respectively. The study shows that this method not only effectively captures the spatiotemporal dependency of stock data but also effectively controls risks while improving returns, providing investors with a robust and efficient decision support system.