Enhancing the pricing efficiency of financial assets with an optimized bayesian network based on efficient fusion

Abstract

To address the limitations of traditional pricing models regarding accuracy and adaptability in high-frequency trading, this study presents a Transformer-based Efficiently-Fused Optimized Bayesian Network (Trans-EFOBN) for financial asset pricing. The framework integrates a masked transformer with temporal logic constraints to extract sequential features and combines a Dynamic Bayesian Network (DBN) to establish hierarchical structural dependencies between macro factors and micro market variables. This design does not aim to establish strict econometric causality but instead leverages an end-to-end learning mechanism to simultaneously optimize feature representation and network parameters. Empirical analyses utilizing minute-level high-frequency data of the CSI 300 constituent stocks from 2019 to 2024 in the Wind database demonstrate substantial performance gains: the mean absolute error (MAE) decreases to 0.037 (approximately 25% lower than the baseline static Bayesian model), while R² attains 0.86. In simulated trading scenarios incorporating transaction costs and slippage, the proposed model yields an annualized return of 14.2% and a Sharpe ratio of 0.95. The results indicate that integrating structural dependency logic with dynamic probabilistic inference significantly enhances asset pricing efficiency and interpretability, providing robust technical support for high-frequency quantitative trading.