High-frequency prediction of intraday crypto order flow relies most on features that capture the microstructure state of the market and its recent dynamics. Empirical and theoretical work in market microstructure shows that features derived from the limit order book and recent trades often outperform coarse indicators like raw price or isolated volume spikes. Rama Cont Imperial College London has characterized how queue sizes and book dynamics drive short-term price moves, and Jean-Philippe Bouchaud École Normale Supérieure and Capital Fund Management has emphasized the role of trade sign persistence and impact in predicting subsequent flow. Marcos López de Prado Cornell University advocates engineered features such as volume-weighted averages and volatility-adjusted returns for robust machine-learning inputs.
Core predictive features
The most informative features are those that encode order book imbalance, signed volume, and execution flow near the top-of-book, because they reflect the current supply-demand asymmetry and the pressure to lift or hit resting orders. Order book slope and multi-level depth capture potential latent liquidity beyond best quotes. Short-window realized volatility and recent returns provide momentum or mean-reversion signals that interact with order flow. Time features such as time-of-day and inter-arrival durations of trades indicate rhythm changes driven by regional participant activity or scheduled events. Features that aggregate past trade signs into an order flow imbalance metric are especially predictive of immediate continuation or reversal.
Relevance, causes, and consequences
These features are predictive because they proxy for the presence of informed trading, algorithmic liquidity provision, and transient liquidity gaps. When private information or execution algorithms create persistent signed volume, limit order book imbalance tends to precede price pressure, producing measurable predictability. In crypto markets, fragmentation across exchanges and 24/7 trading amplify the importance of exchange-specific book features and time-zone effects. Consequences of relying on these predictors include improved market-making performance and reduced adverse selection, but also potential amplification of volatility if many algorithms act on the same signals simultaneously, increasing systemic fragility. Regulatory and territorial factors shape which features matter: fee structures, custody and KYC regimes, and regional participant mixes influence cancelation behavior and the depth of order books. For robust models, practitioners should combine microstructural LOB features with engineered aggregates and validate across exchanges, citing established microstructure research from recognized authors and institutions to ensure defensible, explainable inputs.