Crypto-Anomaly detection with Ripple

This was a collaboration between NYU’s Center for Data Science and Ripple. We were given a large dataset with granular data such as exchanges, cryptocurrency prices, volume and more. We were given the task to identify arbitrage opportunities and whether there were any market conditions that would be predictive of arbitrage opportunities.

Abstract: In cryptocurrency markets, the price of crypto assets can diverge across markets due to numerous reasons (e.g. exchange downtime and trade volumes). Therefore, outlier detection is extremely important for ensuring that erroneous market data does not distort price feeds. This project aims to detect anomalies (outliers) in the price of cryptocurrency transactions across exchanges and assets using various models including Z-score thresholds, Logistic Regression, Random Forest, Ensemble Voting, LSTM, and XGBoost. As a result, the XGBoost model achieves the highest AUC of 0.99, with taker fees being its most important feature. The final model has the potential of being deployed by Ripple as an inference layer on top of various financial models to ensure data quality.