Neural Hidden Markov Model with Adaptive Granularity Attention for High-Frequency Order Flow Modeling

TL;DR

This paper introduces a Neural Hidden Markov Model with Adaptive Granularity Attention that captures multi-scale temporal dynamics in high-frequency financial data, improving prediction of price movements and liquidity shocks.

Contribution

It presents a novel framework combining multi-resolution encoders, adaptive fusion, and a neural HMM with normalizing flows for better market regime modeling.

Findings

Outperforms fixed-resolution models in predicting short-term price movements.

Effectively captures multi-scale market dynamics during volatile periods.

Adaptive mechanism improves model focus and prediction accuracy.

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.