Anticipating Crypto Success: An XAI Framework for Early-Stage Token Viability Using Deployment Features

Abstract

The exponential growth of blockchain-based tokens has heightened the need for reliable methods to assess their longterm viability at deployment, a stage where historical market data is absent and risks such as scams and project failures are prevalent. This study introduces an explainable machine learning framework to predict token viability using static features available at launch, including smart contract properties (e.g., mintability, centralization), deployment details (e.g., network), and metadata (e.g., presence of an icon). We collected 100,000 ERC-20 tokens from Ethereum, Binance Smart Chain, and Polygon and analyzed their characteristics available at deployment and derived features. We labeled them as live or failed based on post-deployment scores derived from liquidity, transfer frequency, and holder distribution. Among the models evaluated, XGBoost with class-weight adjustment excelled, creating an enriched token set that contained, on average, 11 times more live tokens than the original dataset, surpassing other classification models in identifying viable tokens. SHAP analysis highlighted key predictors: tokens with icons, complex yet high-quality code, and deployment on Ethereum were more likely to succeed, while Polygon deployments correlated with higher risk. Though effective as an early filter, the framework's modest standalone precision underscores its role as part of a broader strategy integrating post-launch data. This approach advances early-stage token evaluation, enhancing investor decision-making and risk assessment in decentralized finance.