LOB-ID: Evaluating Synthetic Market Data by Inception Distances

Battle Count: 70

Highly relevant to quantitative trading. The paper directly addresses the critical challenge of evaluating synthetic market data used in backtesting, execution analysis, and risk measurement. LOB-ID provides a principled way to assess whether generated order-book trajectories preserve joint temporal and cross-level structure, which is essential for realistic simulation-based strategy development. The adversarial attack analyses reveal weaknesses in existing evaluation methods that practitioners rely on. The model comparison (ZI, Hawkes, GAN, S5, Diffusion) covers the spectrum of current LOB generators used in quantitative finance.

Moderate to high complexity. Requires: (1) training DeepLOB on Level-2 LOB data with proper preprocessing (z-scoring, event-indexed windows of T=100, 40 features); (2) extracting 96-dimensional Inception embeddings; (3) implementing FID with Gaussian moment estimation; (4) implementing MIND with 256 random 1D projections and exact 2-Wasserstein computation via sorting; (5) constructing Level-2 books from message-level data; (6) training or calibrating five generative models. The adversarial attack experiments add further complexity with constrained optimization. However, the core pipeline is well-defined and modular.

The paper provides detailed methodology including DeepLOB architecture, training procedure, data splits, and metric formulations. However, no GitHub repository is mentioned. The HKEX Level-2 data (September-December 2025) is proprietary and likely not freely available. The five generative models (ZI, CH, LOBGAN, LOBS5, DiffLOB) are described but implementations are not explicitly released. The temporal split and book-reconstruction pipeline are specified, aiding partial reproducibility.