Relevance
9/10
Directly addresses a core operational challenge in quantitative trading: anticipating transient liquidity withdrawal at sub-second horizons. The LWI target is interpretable and maps to actionable decisions for execution algorithms, market makers, and risk managers. The horizon-dependent model selection findings (XGBoost at 1s/5s, HAR at 2s) provide practical guidance for system design. The theoretical aggregation bound offers a principled way to assess whether forecastability is genuine or mechanical. The regime-dependence finding (tree dominance is not universal) is critical for production model selection. The paper explicitly targets liquidity withdrawal rather than mid-price direction, aligning with operational priorities in execution and surveillance.
Implementation Complexity
7/10
The pipeline involves: (1) reconstructing the limit order book from raw MBO event streams at nanosecond resolution, (2) resampling to a 250ms grid with warm start, (3) computing LWI with denominator stabilization, (4) engineering a panel of microstructure features (OFI, QI, spreads, depths, rolling statistics, activity rates), (5) multi-method feature selection (MI, XGBoost importance, LASSO) with cross-ticker consensus, (6) implementing AR(5), HAR, and XGBoost with walk-forward CV and embargo, (7) computing the theoretical aggregation bound via method-of-moments calibration, and (8) Diebold-Mariano testing with Newey-West variances. The MBO data processing and feature engineering are the most complex components; the models themselves are standard. Latency constraints for real-time deployment add engineering complexity.
Reproducibility
3/5
The public-data replication (LOBSTER AAPL, June 21, 2012) is fully reproducible with committed code, derived panels, and SHA-256 hashes in a replication repository. However, the primary commercial Databento MBO data (July 30, 2025, 7 symbols) cannot be redistributed, making the headline Section 5 results author-reported and not independently verifiable. The theoretical Proposition 1 and its proof are fully self-contained. Feature construction and model evaluation protocols are described in sufficient detail for reimplementation.