加载中...
本文提出将自适应粒度注意力(AGA)与神经隐马尔可夫模型(Neural HMM)结合,用于高频限价单簿的多尺度特征提取与时序建模;模型通过并行的扩张卷积(tick级)与可学习小波+LSTM(分钟级)提取多分辨率特征,利用基于局部波动率与成交频率的门控动态融合,并以条件正规流替代高斯发射以刻画非高斯观测分布,实证显示在500ms中期价格预测与流动性冲击检测上显著优于固定分辨率基准
| Model | Accuracy (%) | MCC | Volatile Regime F1 | Sharpe Ratio |
|---|---|---|---|---|
| HMM-GARCH | 58.1 | 0.412 | 0.621 | 1.87 |
| LSTM-ATTN | 62.4 | 0.503 | 0.658 | 2.15 |
| TCN-MultiRes | 63.6 | 0.527 | 0.672 | 2.34 |
| NeuralHMM | 60.8 | 0.463 | 0.643 | 2.01 |
| Wavelet-HMM | 59.3 | 0.431 | 0.629 | 1.92 |
| AGA-Neural HMM | 68.3 | 0.581 | 0.713 | 2.78 |
| 变体 | Accuracy (%) | △ vs. Full |
|---|---|---|
| Full AGA-Neural HMM | 68.3 | — |
| -AGA mechanism | 61.1 | -7.2 |
| -Dilated Conv | 64.2 | -4.1 |
| -Wavelet-LSTM | 65.5 | -2.8 |
| Gaussian emissions | 66.0 | -2.3 |
| Fixed transitions | 67.1 | -1.2 |
元数据与概览(引言与报告概览)
逐节深度解读
图表与数表深度解读(关键环节)
估值/性能衡量与实际可部署性(对高频应用的讨论)
风险因素与作者识别的局限(Section 7)
批判性视角与细微差别
结论性综合
2603.20456.pdf