Fraudes em ICOS: uma abordagem utilizando machine learning

Abstract

This article examines whether machine learning can support the detection of fraud in Initial Coin Offerings (ICOs). Using a dataset of 162 ICOs and 11 financial, informational, and ESG variables, we apply the K-means clustering algorithm to compute Euclidean distances to cluster centroids and classify projects above the 95th percentile as anomalies. Fraud labels, derived from on-chain activity on the Ethereum network (Etherscan API), are then used as ground truth to evaluate the model’s performance. The algorithm flags nine ICOs as anomalous, all of which are fraudulent, resulting in high precision and no false positives; however, sensitivity is very low, as only 9 out of 130 fraudulent ICOs are detected. Variable importance analysis shows that anomaly detection is primarily driven by token price and hard and soft caps, while informational signals and ESG scores play a residual role. These findings highlight the limitations of distance-based unsupervised methods and point to the need for richer data and supervised modeling approaches.