When Does Order Flow Matter? State-Dependent L2 Liquidity-State Transitions in Crypto Futures

Battle Count: 89

Highly relevant for quantitative trading in crypto futures markets. The paper establishes that pre-event L2 liquidity state is the first-order predictor of post-event liquidity regimes, providing a state-first design principle for market microstructure models. The finding that order flow adds incremental value only as an overlay (ETH-dominant, stress-amplified) directly informs execution strategy design, liquidity-aware quoting, and event-window trading. The staged evaluation protocol and baseline serve as a benchmark that RL, execution-policy, or LLM-based layers must exceed. However, the paper is a prediction study, not a trading strategy paper, and reports no profit or execution results.

Moderate complexity. The core models (logit, shallow gradient boosting) are straightforward to implement. However, the full evaluation protocol requires careful implementation of: rolling monthly out-of-sample folds, event-clustered bootstrap resampling (500 resamples), blocked permutation tests (64 draws with month/symbol/state blocks), matched non-event window construction (shifted by whole weeks), train-fold-only tercile threshold estimation, and multi-scale L2 feature engineering from top-20 order book snapshots. The data pipeline aligning L2 order book, trade flow, and macro calendar at one-minute cadence over 3.5 years adds engineering complexity.

The paper provides detailed evaluation protocol specifications including rolling monthly folds, event-clustered bootstrap (500 resamples), blocked permutation tests (64 draws), and explicit model hyperparameters (depth-3 histogram GBC, 60 iterations, lr=0.05, L2 reg=1.0). Code, evaluation contracts, manifests, artifact hashes, and non-proprietary derived artifacts are stated to be released upon acceptance. Licensed raw inputs (Binance order book, trade flow, macro calendar) cannot be redistributed. The dataset composition (47,513 windows, 9,773 events, 40 monthly folds) is fully specified.