TABL-ABM: A Hybrid Framework for Synthetic LOB Generation

Battle Count: 73

Directly addresses synthetic LOB data generation for trading applications. The framework enables: (1) training trading models on synthetic data including rare events, (2) stress testing execution strategies, (3) predicting market impact of trades, (4) generating realistic order flow for backtesting. The hybrid approach combining mechanistic ABM with data-driven deep learning is particularly relevant for generating data that generalizes beyond historical observations.

Requires implementing: (1) three separate TABL/BiNTABL neural network models with specific architectures, (2) extended Chiarella ABM with multiple trader types and calibration via grid search, (3) a simulated matching engine with price-time priority queuing, (4) novel Bayesian-based order deletion modeling with dynamic probability updates, (5) data preprocessing including OFI feature engineering and binning strategies. The combination of deep learning, ABM, and matching engine simulation adds significant integration complexity.

The paper provides detailed model architecture descriptions, hyperparameters (dropout=0.3, focal loss gamma=2, learning rate scheduler parameters, binning strategies), and ABM calibration equations. However, no code repository is provided, and some implementation details (exact grid search ranges, specific LOBSTER data day selection beyond AAPL 21/6/2012) are not fully specified. The LOBSTER data is publicly available but requires a subscription.