The Predictability of Risk: Linking User Action Frequency to Survival Model Performance

This paper explores the relationship between the predictability of user actions and their statistical frequency within the Aave DeFi lending protocol. Leveraging the FinSurvival dataset, we analyze the performance of deep learning survival models (DeepSurv and DeepHit) across various transaction pairs. By correlating model performance (C-index) with frequency metrics (Count and Lift), we find that a monolithic view of user behavior is insufficient. Our analysis reveals that for users who have never been liquidated, both high-frequency and statistically surprising actions are harder to predict. Conversely, for users who have experienced a liquidation, raw action frequency becomes a stronger indicator of unpredictability, especially prior to the event. Notably, we observe a behavioral shift post-liquidation, where the predictability of statistically surprising actions inverts, suggesting a move towards more deliberate behavior. These findings underscore the importance of user segmentation for building resilient AI-driven risk management systems in DeFi and provide a foundation for developing more adaptive and context-aware predictive models.