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

Battle Count: 78

The paper is highly relevant to quantitative trading, specifically targeting high-frequency order flow modeling and mid-price prediction at 500ms horizons. It addresses core challenges in HFT: multi-scale liquidity dynamics, regime shifts, and adaptive resolution selection. The trading simulation results (Sharpe ratios of 2.78 for equities, 2.41 for crypto) and profitability up to 3bps transaction costs demonstrate practical viability. The model's ability to autonomously shift granularity during volatile periods offers direct advantages for algorithmic trading, market making, and execution optimization. The interpretable hidden states provide actionable regime insights for trading strategy adaptation.

The architecture is highly complex, combining multiple specialized components: dilated causal convolutions with exponential dilation rates, learnable wavelet transforms implemented as constrained 1D convolutional layers, LSTM for coarse-grained processing, multi-head attention mechanisms, gating networks conditioned on market microstructure variables, conditional normalizing flows with coupling layers and masked MLP transformers, volatility-adaptive transition models with temperature scaling, and GRU-based transition tracking. Training requires end-to-end optimization of all components simultaneously with a composite loss function. The normalizing flow Jacobian computations and forward algorithm for HMM likelihood add significant computational overhead. However, the paper reports manageable inference latency (0.83ms at 99th percentile) and only 12% overhead over vanilla NeuralHMM.

The paper provides detailed architectural specifications (dilation rates, kernel sizes, number of coupling layers, training hyperparameters like learning rate and batch size), dataset descriptions (NASDAQ, LSE, Binance), and evaluation protocols. However, no code repository is mentioned, and some implementation details (e.g., exact wavelet basis constraints, specific normalizing flow architecture choices) are described at a high level. The training protocol with 4-week rolling windows and specific GPU hardware (NVIDIA V100) is provided. Statistical significance testing via Diebold-Mariano tests is specified.