Explainable Patterns in Cryptocurrency Microstructure

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.

💡 Research Summary

**
The paper investigates whether a compact, universal set of high‑frequency microstructure features can predict short‑horizon returns across cryptocurrency assets that differ dramatically in market capitalization and liquidity. Using Binance Futures perpetual contract data at a 1‑second granularity from 1 January 2022 to 12 October 2025, the authors construct a unified feature library for five assets – Bitcoin (BTC), Litecoin (LTC), Ethereum Classic (ETC), Enjin Coin (ENJ) and ROSE – that spans three intuitive groups: (1) top‑of‑book metrics (mid‑price, spread, level‑1 volumes), (2) order‑flow and trade‑imbalance measures (net signed volume over short windows), and (3) VWAP‑to‑mid price deviations for both buy and sell trades. All features are kept in their raw scale; relative ratios (e.g., spread/mid) are used where scale‑invariance is required.

Modeling is performed with CatBoost, a gradient‑boosted decision‑tree algorithm well‑suited for tabular data. Two loss functions are compared: the conventional mean‑squared‑error (MSE) and a finance‑specific Generalized Mean‑Absolute Directional Loss (GMADL). GMADL penalises sign‑incorrect predictions more heavily and weights errors by the absolute magnitude of realized returns, reflecting the practical importance of correctly anticipating large price moves. A rolling‑window time‑series cross‑validation scheme with a purge gap between training and test periods is employed to avoid look‑ahead bias. Hyper‑parameters are tuned only inside the training window; the outer validation set is never used for tuning.

After training, the authors apply TreeSHAP to obtain both global feature importance rankings and local partial‑dependence‑like SHAP dependence plots. The results are strikingly consistent across all five assets. The three most important features are (i) order‑flow imbalance, (ii) spread, and (iii) VWAP‑mid deviation. The SHAP dependence curves reveal: (a) a monotonic but concave relationship for order‑flow imbalance—large imbalances increase predicted returns but with diminishing marginal effect, echoing Kyle’s linear price‑impact model; (b) a negative, roughly linear effect of spread, consistent with Glosten‑Milgrom’s adverse‑selection argument that wider spreads signal higher execution risk; (c) asymmetric effects for VWAP‑mid deviation, where buy‑side deviations generate positive SHAP values (upward pressure) and sell‑side deviations generate negative values (downward pressure), aligning with the “square‑root impact” law observed in empirical market impact studies.

To test economic relevance, two simple trading strategies are back‑tested. The “taker” strategy executes market orders on the side indicated by the model’s sign (buy when predicted positive, sell when predicted negative) at the best available price, but only when the absolute SHAP contribution exceeds a modest threshold (|SHAP| > 0.05). The “maker” strategy places limit orders on the opposite side at a fixed depth (e.g., five price levels) to provide liquidity. In normal market periods the taker strategy achieves an annualized Sharpe ratio of about 1.2, while the maker strategy yields roughly 0.9, demonstrating that the SHAP‑driven signals are tradable. During a major flash‑crash on 15 November 2023 (≈30 % price drop within seconds), the taker strategy suffers a severe loss (‑45 %), whereas the maker strategy limits its loss and even records a modest gain (+5 %). This divergence empirically validates classic adverse‑selection theory: aggressive liquidity‑taking during extreme order‑flow spikes incurs high execution risk, while passive liquidity provision can be relatively protected.

Portability is further confirmed by applying the same feature set, hyper‑parameters, and model architecture to each asset without any asset‑specific tuning. Out‑of‑sample R² values range from 0.02 to 0.04, and GMADL scores from 0.78 to 0.84, indicating that predictive performance is stable despite a ten‑fold market‑cap spread. The authors argue that relative price and flow measures capture the essential information needed for short‑horizon forecasting, making the approach scalable to a broader crypto universe.

In summary, the paper makes three substantive contributions: (1) it demonstrates that a small, interpretable set of microstructure features yields consistent importance rankings and functional relationships across heterogeneous crypto assets; (2) it links SHAP‑derived dependence shapes directly to established microstructure theory (order‑flow impact, spread‑based adverse selection, and price‑impact curvature); and (3) it provides a robustness analysis by contrasting taker and maker strategies during a flash‑crash, thereby offering empirical support for theoretical risk mechanisms. The authors suggest future work on extending the framework to spot markets, integrating raw order‑book images via deep learning, and building real‑time risk‑aware execution systems that combine the interpretability of SHAP with the predictive power of more complex models.