FlowLOB Generates Realistic, Controllable Limit Order Books Efficiently

Limit Order Book (LOB) simulators are vital tools for financial practitioners, but existing models often fall short in combining realism, computational efficiency, scenario controllability, and generalization to new instruments. This research presents FlowLOB, a novel conditional flow-matching generator designed to produce realistic LOB trajectories. FlowLOB addresses the limitations of previous agent-based and deep generative simulators by excelling in these critical areas. The model was trained on diverse Hong Kong Exchange (HKEX) symbols at various sampling frequencies, using a tick-relative representation, and demonstrated successful zero-shot transfer to unseen instruments. A controlled comparison with diffusion models, using identical data, architecture, and budget, revealed FlowLOB's superior sampling efficiency. Flow matching achieved its best quality with only 10 ODE-solver steps, whereas diffusion models required many more function evaluations to reach comparable fidelity. At this efficient operating point, FlowLOB significantly improved realism across most distributional metrics compared to baselines at finer sampling frequencies. Furthermore, it demonstrated strong counterfactual controllability, meaning that changes in scenario conditions accurately shifted generated statistics towards corresponding real-world tail regimes. This combination of realism, efficiency, and control positions FlowLOB as a powerful tool for market simulation and analysis.