Paper: arXiv 2603.20456

Authors: Tianzuo Hu

Abstract

We propose a Neural Hidden Markov Model (HMM) with Adaptive Granularity Attention (AGA) for high-frequency order flow modeling. The model addresses the challenge of capturing multi-scale temporal dynamics in financial markets, where fine-grained microstructure signals and coarse-grained liquidity trends coexist. The proposed framework integrates parallel multi-resolution encoders, including a dilated convolutional network for tick-level patterns and a wavelet-LSTM module for low-frequency dynamics. A gating mechanism conditioned on local volatility and transaction intensity adaptively fuses multi-scale representations, while a multi-head attention layer further enhances temporal dependency modeling. Within this architecture, a Neural HMM with conditional normalizing flow emissions is employed to jointly model latent market regimes and complex observation distributions. Empirical results on high-frequency limit order book data demonstrate that the proposed model outperforms fixed-resolution baselines in predicting short-term price movements and liquidity shocks. The adaptive granularity mechanism enables the model to dynamically adjust its focus across time scales, providing improved performance particularly during volatile market conditions.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents an advanced hybrid methodology combining Neural HMMs, normalizing flows, and multi-scale attention to tackle high-frequency market dynamics. It demonstrates strong empirical rigor using limit order book data, though replication may require detailed implementation efforts given the complex neural architecture.

Research Flowchart

  flowchart TD
    A[Research Goal: Model Multi-Scale Dynamics in High-Frequency Order Flow] --> B[Data Input: Limit Order Book Tick Data]
    B --> C[Parallel Multi-Resolution Encoders<br>Dilated CNN & Wavelet-LSTM]
    C --> D[Adaptive Granularity Attention<br>Gated Fusion based on Volatility & Intensity]
    D --> E[Neural HMM with Conditional<br>Normalizing Flow Emissions]
    E --> F[Outcomes: Dynamic Scale-Focus & Superior Prediction<br>of Price Movements & Liquidity Shocks]

Cited References (38)

  1. Adaptive Hierarchical Hidden Markov Models for Structural Market Change (2025) — link
  2. Attention-Based Multi-Asset Order Flow Networks for Enhanced Mid-Price Prediction (2025) — link
  3. Incorporating Liquidity and Order Flow Imbalances for Intraday Stock Volatility Prediction (2025) — link
  4. Time Series Prediction Based on Multi-Scale Feature Extraction (2024) — link
  5. Ensemble of temporal Transformers for financial time series (2024) — link
  6. Multi-scale adaptive attention-based time-variant neural networks for multi-step time series forecasting (2023) — link
  7. Multi-Scale Convolutional Network with Channel Attention Mechanism for rolling bearing fault diagnosis (2022) — link
  8. Dilated convolution with learnable spacings (2021) — link
and 30 more
  1. A review on the attention mechanism of deep learning (2021) — link
  2. An Effective Optimization Method for Machine Learning Based on ADAM (2020) — link
  3. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation (2020) — link
  4. AT-LSTM: An Attention-based LSTM Model for Financial Time Series Prediction (2019) — link
  5. Testing Stylized Facts of Bitcoin Limit Order Books (2019) — link
  6. Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods (2017) — link
  7. Modelling the stock market using a multi-scale approach (2017)
  8. Comparing Predictive Accuracy, Twenty Years Later: A Personal Perspective on the Use and Abuse of Diebold–Mariano Tests (2015) — link
  9. Optimal Execution with Dynamic Order Flow Imbalance (2014) — link
  10. Market Microstructure in Practice (2013) — link
  11. Realized Volatility: A Review (2008) — link
  12. Optimal execution strategies in limit order books with general shape functions (2007) — link
  13. A Dynamic Approach To Feature Weighting (2002) — link
  14. Segmenting time series with a hybrid neural networks - hidden Markov model (2002)
  15. The London Stock Exchange: A History (2001) — link
  16. Wavelet methods in (financial) time-series processing (2000) — link
  17. Market Microstructure Theory (1995) — link
  18. Studies on cross-prediction (2025)
  19. High and low frequency attention network for long-term traffic flow prediction (2025)
  20. TCN-MultiRes : A temporal convolutional network with fixed multi-resolution branches (2022)
  21. LSTM-ATTN : An attention-enhanced LSTM network (2019)
  22. Market Manipulation Detection Based on Classification Methods (2017) — link
  23. Recurrent Neural Networks Approach to the Financial Forecast of Google Assets (2017)
  24. Discrete Wavelet Transform: A Signal Processing (2016)
  25. Sharpe Ratio (SR) : Risk-adjusted returns from simulated trading (2012)
  26. The Sharpe ratio efficient frontier (2012)
  27. Hidden Markov Models In Finance (2007) — link
  28. NeuralHMM : A vanilla Neural HMM without adaptive granularity (2002)
  29. We introduce a gating mechanism that conditions resolution weights on measurable market microstructure variables, providing explainable adaptation compared to black-box alternatives
  30. MS-TCT: Multi-scale temporal convolutional transformer for action detection

0 of 38 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2518 linked papers.