Protocol liquidity, rule-based benchmarks, and USDT yield predictability in DeFi lending
Аннотация
This paper studies short-horizon predictability of the DefiLlama-reported USDT supply APY for the Aave V3 market on Ethereum. The empirical design uses independently sourced reserve-liquidity, crypto-market, and macro-financial variables and excludes processed utilization and borrowed-balance proxies that are mechanically constructed from APY. Expanding-window forecasts reveal substantial horizon dependence. Current-rate and horizon-matched moving-average rules dominate at one and three days. At seven days, State XGBoost records the lowest RMSE, but its squared-loss advantage over MA(7D) is not statistically significant. At fourteen days, the state model lowers RMSE from 0.608 to 0.434; pairwise loss tests favor the state model, while horizon-wise multiplicity adjustment leaves the absolute-loss result statistically decisive. TVL has the expected negative association with future APY but does not improve a strong moving-average benchmark as a linear correction. In ex-post high-yield episodes, all principal models have negative out-of-sample R-squared relative to the subgroup mean. The evidence points to strong near-term own-rate persistence and horizon-dependent value of independently observed financial states.