FlowLOB: Efficient and Controllable Limit Order Book Generation with Flow Matching

Battle Count: 69

Highly relevant to quantitative trading. FlowLOB directly addresses core needs of practitioners: generating realistic LOB trajectories for backtesting, stress-testing execution algorithms, and training RL agents. The 10-step Euler sampling makes it computationally feasible for large-scale rollouts across multiple symbols and regimes. The tick-relative representation enables zero-shot transfer to unseen instruments, critical for multi-asset strategies. The counterfactual controllability (liquidity, imbalance, volatility, trend) allows scenario-based stress testing. The paper explicitly mentions applications in dynamic execution, portfolio risk analysis, and RL-based trading. The L2 book state output is directly suitable for snapshot-based trading policies.

Moderate to high complexity. Requires implementing: (1) flow matching training with linear interpolation paths, (2) adaLN-Zero transformer architecture (~100M parameters), (3) tick-relative LOB representation with log volume transforms, (4) multi-frequency data resampling pipeline, (5) scenario conditioning with classifier-free guidance (0.5 dropout), (6) ODE sampling with Euler/Heun/RK4 solvers, (7) matched diffusion baseline for comparison. The architecture is standard (DiT-style), but the data preprocessing for HKEX L2 data, multi-symbol training, and the evaluation pipeline (W1, KS, KL, JS distances) add complexity. Training requires significant GPU resources for 500 epochs on ~100M parameter model.

The paper provides detailed training settings (AdamW, batch size 1024, 500 epochs, early stopping patience 20, conditioning dropout 0.5), architecture details (~100M adaLN-Zero transformer), data description (HKEX L2 data, 8 symbols, 3 frequencies), and evaluation protocol. However, no code or model weights are explicitly mentioned as available. The data is from HKEX which may require subscription. The experimental setup is well-described but full reproduction requires access to the specific HKEX data and computational resources for ~100M parameter model training.