DiffLOB: Diffusion Models for Counterfactual Generation in Limit Order Books

Battle Count: 74

Highly relevant to quantitative trading. The paper directly addresses stress testing, scenario analysis, and counterfactual reasoning for limit order books - core needs in algorithmic trading. The ability to simulate 'what-if' market scenarios (liquidity shocks, volatility spikes, trend reversals) is critical for strategy robustness testing, risk management, and market making. The downstream task improvement (regime prediction) demonstrates practical value for trading signal generation. The LOB-level granularity makes it directly applicable to execution algorithms and market microstructure research.

Moderate-to-high complexity. Requires implementing DDPM with WaveNet backbone, dual-stage training (ControlNet-inspired), FiLM modulation, classifier-free guidance, and regime condition encoders. The architecture has multiple components (local/global encoders, control module, price level embeddings). Training requires careful hyperparameter tuning and the two-stage procedure. However, code is provided and the paper follows well-established diffusion model practices. Data preprocessing (LOBSTER format, price/volume normalization) adds additional complexity.

Code is publicly available on GitHub (https://github.com/ZhuoHan1998/DiffLOB). Data source (LOBSTER) is commercial and requires subscription. Training details are well-documented including hyperparameters (learning rate 1e-4, batch size 128, 100 noise levels, 16 residual blocks). Dual-stage training algorithm and sampling algorithm are provided in appendices. However, LOBSTER data access is a barrier to full reproducibility.