This paper introduces an enhanced transformer architecture that can effectively process numerical stock data and accurately forecast future stock returns. The proposed model outperforms over 100 traditional factor-based quantitative strategies in the Chinese stock market.
The authors advocate for using randomized quasi-Monte Carlo (RQMC) quadrature to enhance the scalability of Fourier methods in high dimensions, providing practical error estimates. They propose an efficient domain transformation procedure based on integrand regularity to improve RQMC convergence rates.