The Prediction on Short-Term Liquidity of Decentralized Finance Driven by Machine Learning

Abstract

Total Value Locked (TVL) explicitly reflects the total asset users deposit in Decentralized Finance (DeFi) protocols, similarly to the Asset Under Management (AUM) in traditional finance. This exposes liquidity providers to the risk of short-term liquidity depletion and highlight the urgent need for quantifiable and predictive risk management tools. As its short-term fluctuations can be effectively characterized by on-chain static features (e.g., fee tier, volatility), dynamic features (e.g., token balance changes, slippage), and technical indicators (e.g., MA), and since posterior calibration methods based on high-accuracy point forecasting models can provide more reliable estimations of downside risk boundaries, this study focuses on the USD Coin - Ethereum pool (0.3 % fee tier) of Uniswap V3. The model is constructed using 17 features, with eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine employed for TVL growth-rate regression forecasting. The results are compared with those from Long Short-Term Memory, Gated Recurrent Unit, and Naïve baselines. Furthermore, the research introduces the Liquidity-at-Risk (LaR95) metric to estimate downside risk through both residual-based and quantile regression approaches, and conduct interpretability analysis using SHAP values. XGBoost obviously outperforms Deep learning models on directional accuracy. XGBoost demonstrates a significantly superior performance to deep learning models in predicting the direction of TVL changes. The residual-based LaR95(liquidity-at-risk at the 95% confidence level) derived from its point forecasts exhibits a coverage rate closely aligned with the theoretical level, validating the effectiveness, robustness, and interpretability of the “high-accuracy prediction and residual calibration” framework in DeFi risk management.