This research paper presents a novel approach to hedge fund portfolio construction by integrating PolyModel theory with iTransformer deep learning techniques. The study addresses the fundamental challenge of sparse financial time series data, which often hinders the application of machine learning methods in finance. PolyModel theory provides a robust framework for this context by enabling comprehensive risk factor selection that includes various financial indices, currencies, and commodity prices, thereby mirroring real-world market complexities. The implementation utilizes an extensive dataset of over 10,000 hedge funds spanning 29 years, creating quantitative measures such as Long-term Alpha, Long-term Ratio, and SVaR alongside traditional metrics like the Sharpe ratio.

The methodology combines PolyModel theory's data handling capabilities with iTransformer's advanced forecasting power for high-dimensional time series. The iTransformer model specifically addresses challenges in capturing market trends while efficiently controlling downside risk. This integration allows the development of multiple portfolio strategies that demonstrate superior performance compared to various benchmarks. The results show significant improvements in both Sharpe ratio and annualized returns, validating the effectiveness of this hybrid approach in achieving the dual objectives of high returns and low risks in hedge fund portfolio management.