Skip to content

Explainable Patterns in Cryptocurrency Microstructure

Authors: Bartosz Bieganowski, Robert Ślepaczuk Institution: University of Warsaw, Faculty of Economic Sciences — Department of Quantitative Finance and Machine Learning Year: 2026 (February 2026) arXiv: 2602.00776 Categories: q-fin.TR, q-fin.CP, q-fin.ST

Plain-language abstract

Do crypto markets share a universal microstructure? This paper trains CatBoost gradient-boosted trees with a direction-aware GMADL objective on Binance Futures perpetual contracts for five assets spanning an order of magnitude in market capitalisation — BTC, LTC, ETC, ENJ, ROSE (market-cap ranks 1 / 20 / 40 / 60 / 100 at the start of 2022). SHAP analysis shows that both feature importance rankings and partial-effect shapes are strikingly similar across all five assets despite very different liquidity profiles. A major flash crash provides a natural experiment: the taker strategy collapses while the maker strategy survives — empirical validation of classical adverse-selection theory. Authors argue for a portable universal feature library for crypto short-horizon returns.

Key contributions

  1. Cross-asset SHAP universality — feature importance rankings are highly Spearman-correlated across five assets; dependence-curve shapes coincide after relative-price / relative-flow normalisation.
  2. Direction-aware GMADL objective — Generalized Mean Absolute Directional Loss, rewarding sign-correct forecasts scaled by return magnitude. Models are trained with squared-error loss but selected by GMADL on inner-fold performance.
  3. Tick-size modulation: cross-asset, the high-quantile SHAP magnitude of OBI rises with effective tick size. Validated with a clever natural experiment — W/USDT spot (tick \(10^{-4}\)) vs W/USDT perp (tick \(10^{-5}\)) — spot OBI correlates at \(c = 0.94\) with the futures mid's position within the spot spread. OBI in large-tick markets is a visible proxy for the latent continuous price, linking directly to Stoikov's microprice.
  4. Paired taker / maker backtest — top-of-book taker uses conservative unfavourable-side inventory marking; a fixed-depth maker complement allows the flash-crash adverse-selection contrast.
  5. Flash-crash stress test — taker and maker performance diverge sharply during a major crash in a way that matches Glosten–Milgrom predictions about uninformed liquidity providers being picked off.

Method summary

Data

  • Binance Futures perpetual contracts: BTC, LTC, ETC, ENJ, ROSE.
  • 1-second frequency, 1 January 2022 → 12 October 2025.
  • Top-of-book quotes synchronised with trades.
  • Target: 3-second log return of mid price, \(r_{t \to t+3s}\).

Feature families

  • Top-of-book metrics — mid, spread, best-bid/ask sizes.
  • Order / trade imbalances — net signed flows.
  • VWAP-to-mid deviations — for buy- and sell-side trades separately.

Features kept in original scale (tree models are scale-invariant). Relative measures (spread/mid, VWAP/mid) preferred for cross-asset comparability.

Model & training

  • CatBoost (ordered boosting, TreeExplainer-compatible).
  • Optuna TPE Bayesian hyperparameter optimisation (depth, iterations, lr, ℓ₂ leaf reg, subsampling temperature, discretisation granularity).
  • Nested time-series CV: inner CV inside each training window for tuning; outer walk-forward folds with a purge window between train and test to prevent leakage from slow-moving features.
  • Models trained with squared-error loss; the GMADL-scored checkpoint is retained for explanation + backtest (squared-error variant kept as robustness check).

Backtests

  • Taker: signals derived by thresholding predictions, \(\hat R_t > \theta\). Buy at best ask, sell at best bid. Inventory marked to the unfavourable side of the book — systematically pessimistic accounting. Position changes only on signal flips. Latency not modelled → upper bound in fastest regime.
  • Maker: fixed-depth limit-order strategy for the flash-crash contrast.

Main results

Cross-asset stability

  • SHAP feature-importance rankings are highly Spearman-correlated across all five assets.
  • Same three families dominate mean-absolute SHAP everywhere: OFI, bid-ask spread, VWAP-to-mid deviations.
  • Dependence-curve shapes agree:
  • OFI — largely monotone with concavity at the extremes (diminishing marginal effect as pressure accumulates).
  • Spread — wider spread → attenuated predictive effect (adverse selection rises, confidence falls).
  • VWAP-to-mid — asymmetric; coherent with transient pressure followed by microstructure reversion as depth replenishes.

Tick-size effect

Across assets, high-quantile imbalance SHAP rises with effective relative tick size. The W/USDT spot-vs-perp natural experiment (finer tick in perp) gives a correlation of 0.94 between spot OBI and the perp-implied continuous price location within the spot spread — i.e. in large-tick markets OBI is essentially a readable microprice. Connects directly to the small/medium/large-tick taxonomy in deep-lob-forecasting.

Taker backtest (gross, fixed notional)

Asset ARC ASD IR* MDD Buy-and-hold ARC
BTC 0.13 0.53 0.25 0.29 0.83
LTC 0.07 0.99 0.07 0.64 0.53
ETC 5.78 0.64 8.97 0.24 −0.10
ENJ 4.06 0.62 6.58 0.26 −0.67
ROSE 7.00 1.33 5.28 0.43 −0.71

Signal quality is strongest on smaller-cap assets where the universal feature set is less priced in — BTC gets similar IR to buy-and-hold, while ROSE/ETC/ENJ dominate their own buy-and-hold on a risk-adjusted basis.

Flash-crash case study

During a major flash crash event, the taker strategy's equity collapses; the maker strategy remains resilient. The divergence is directionally correct under adverse-selection theory (Glosten–Milgrom): in a one-way market, takers are hit at increasingly unfavourable prices while makers who stand aside are protected — but the paper's fixed-depth maker gets picked off at the onset, matching the picking-off predictions too.

Limitations

  • Binance Futures only — one venue, perpetual contracts; spot, traditional exchanges, and fragmented crypto venues untested.
  • 1-second frequency — tick-level and minute-level behaviour may differ; execution latency not modelled (results = upper bound under zero latency).
  • Five assets — strong universality signal but a small sample; "long tail" below rank 100 not covered.
  • No walk-forward online deployment — all backtests are offline replays of historical data, no live trading validation.
  • Flash-crash analysis is a case study, not a systematic stress test across multiple events or regimes.

Connections