Explainable Patterns in Cryptocurrency Microstructure

TL;DR

This work investigates whether short-horizon crypto returns admit a universal microstructure representation across assets with wide capitalization differences. It combines a compact feature library of top-of-book metrics, order-flow imbalances, and VWAP deviations with a direction-aware GMADL objective in CatBoost, evaluated via forward-looking, purged rolling cross-validation. Across five cryptocurrencies, the authors show cross-asset invariance in SHAP importance and dependence shapes, linking these patterns to microstructure theory and validating tradability through taker and maker backtests. A major robustness test during the October 2025 flash crash demonstrates that taker strategies are more robust to extreme events, while maker strategies suffer from severe adverse selection, highlighting systemic risks of homogeneous algorithmic trading and motivating universal feature libraries for crypto markets.

Abstract

We document stable cross-asset patterns in cryptocurrency limit-order-book microstructure: the same engineered order book and trade features exhibit remarkably similar predictive importance and SHAP dependence shapes across assets spanning an order of magnitude in market capitalization (BTC, LTC, ETC, ENJ, ROSE). The data covers Binance Futures perpetual contract order books and trades on 1-second frequency starting from January 1st, 2022 up to October 12th, 2025. Using a unified CatBoost modeling pipeline with a direction-aware GMADL objective and time-series cross validation, we show that feature rankings and partial effects are stable across assets despite heterogeneous liquidity and volatility. We connect these SHAP structures to microstructure theory (order flow imbalance, spread, and adverse selection) and validate tradability via a conservative top-of-book taker backtest as well as fixed depth maker backtest. Our primary novelty is a robustness analysis of a major flash crash, where the divergent performance of our taker and maker strategies empirically validates classic microstructure theories of adverse selection and highlights the systemic risks of algorithmic trading. Our results suggest a portable microstructure representation of short-horizon returns and motivate universal feature libraries for crypto markets.

Figures (22)

• Figure 1: The raw feature correlation heatmap shows some common patterns across assets, such as positive correlation in spreads and volatility, positive correlation in cumulative depths on the same side across various levels, or negative correlation in order book imbalance and VWAP-to-mid deviations. Correlation values remain quite similar across assets.
• Figure 2: Schematic of walk-forward cross-validation with purging. In each fold, the model is trained on a rolling historical window and evaluated on a sequential future period, with a temporal gap (purge window) in between. This gap helps prevent information leakage by ensuring that slowly evolving features or delayed market responses do not contaminate the validation set, thus providing a more realistic estimate of true, forward-looking predictive performance.
• Figure 3: Global SHAP summaries for BTC, LTC, and ETC shown side by side.
• Figure 4: SHAP dependence plots (top feature: orderbook imbalance) for BTC, ETC, and ROSE shown side by side.
• Figure 5: Association between relative tick size and high-quantile imbalance SHAP values across assets. Larger relative ticks correspond to stronger imbalance contributions.