基于增量式因果Transformer的高频金融时序预测OA

High-frequency financial data exhibit pronounced nonlinear characteristics and high volatility,significantly complicating accurate market forecasting.Traditional predictive models,reliant upon static correlations among features,fail to capture dynamic causal structures,thereby lacking robust generalization across varying market conditions.To address these challenges,this study pro-poses a novel forecasting model,the meta-causal transformer(MCT),which synergistically integrates meta-learning with dynamic causal inference.Specifically,an incremental causal discovery algorithm,leveraging a sliding-window approach and adaptive Grang-er causality testing,is introduced to dynamically reconstruct causal relationships among market variables.An adaptive decay mecha-nism further enhances this capability,allowing real-time tracking of rapidly evolving causal patterns.The causal-constrained Trans-former architecture utilizes attention masks explicitly to eliminate spurious correlations driven by market noise,thus strengthening the model's interpretability and predictive robustness.Additionally,a meta-learning framework is employed to ensure rapid adapta-tion of model parameters to diverse causal scenarios during market regime shifts.Empirical evaluations on Level 2 high-frequency order book datasets from the Chinese A-share and U.S.stock markets demonstrate that the MCT model achieves substantial improve-ments,yielding a 7%-15%increase in directional prediction accuracy and a reduction in prediction latency by 1-3 time steps com-pared to state-of-the-art benchmarks.This research provides an interpretable and dynamically adaptive causal inference methodolo-gy,offering robust decision-making support for real-time high-frequency trading systems.