An anomaly-detection model trained on one-minute Bitcoin and Ethereum ETP tape from Xetra and Nasdaq Stockholm reached AUC-ROC of 0.82 when forecasting structural breaks one bar ahead, according to a paper published on arXiv. The framework codifies three microstructure failure modes and routes each through machine-learning classifiers. For execution algorithms priced against European crypto ETPs, the result defines a measurable lead time before dislocation propagates into fillable alpha.
The Signal Stack
Three indicators structure the detection layer. Cross-venue divergence measures price dislocation between the Xetra and Nasdaq Stockholm listings of identical underlying ETPs—a function of latency arbitrage between correlated venues. Unrecovered price drops flag one-directional moves that fail to retrace within a fixed bar window, functioning as a proxy for institutional flow exhaustion. Momentum reversals isolate the statistical signature of trend failure at one-minute resolution. The paper evaluates each signal independently and in combination, enabling regime-specific anomaly profiles across BTC and ETH products.
Backtest Boundaries
The dataset comprises one-minute bars across Bitcoin and Ethereum ETPs across two venues, with an undisclosed lookback window. Classifier architecture was not specified in the available abstract. Reported performance peaks at AUC-ROC of 0.82, prediction horizon one bar forward. The source does not disclose standard deviation of false-positive rates, slippage-adjusted PnL, or post-deployment Sharpe ratio. The 0.82 figure is therefore a classifier output, not a strategy output.
What Desks Should Test
- Replicate the three indicators against current tick data; verify divergence thresholds hold across both BTC and ETH ETP pairs.
- Benchmark classifier output against logistic regression and random-forest baselines. An AUC-ROC of 0.82 has limited alpha value without comparative framing.
- Measure execution slippage inside flagged windows. The signal is actionable only if fills occur within the one-bar prediction horizon.
- Define recalibration frequency. Anomaly profiles compress during low-volatility regimes; model edge decays without parameter updates.
A 0.82 AUC-ROC on one-minute ETP tape is a starting condition, not a deployed system. The dataset is public, the framework is documented, the implementation gap is quantifiable. Until slippage, drawdown, and Sharpe-adjusted return are published, the signal remains research-grade.