pith. sign in

arxiv: 2604.14583 · v1 · submitted 2026-04-16 · 💻 cs.LG

From Risk to Rescue: An Agentic Survival Analysis Framework for Liquidation Prevention

Fernando Spadea , Oshani Seneviratne This is my paper

Pith reviewed 2026-05-10 11:54 UTC · model grok-4.3

classification 💻 cs.LG
keywords DeFiliquidation preventionsurvival analysisautonomous agentcounterfactual optimizationAave v3risk managementtime-to-event modeling
0
0 comments X

The pith

An autonomous agent using survival analysis prevents DeFi liquidations where static health-factor rules fail by simulating minimal interventions that carry zero risk of worsening any position.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops an agent that perceives liquidation risk through time-to-event modeling, then actively executes protocol-faithful fixes rather than issuing passive alerts. It derives a normalized return period from a Cox model and pairs it with a volatility filter to isolate genuine threats from dust positions. A counterfactual loop then searches for the smallest capital adjustment that averts liquidation. When tested on thousands of high-risk Aave v3 profiles inside a faithful simulator, the agent rescues positions that static thresholds would have liquidated without ever increasing risk for any user.

Core claim

The agent combines a return period metric from a numerically stable XGBoost Cox proportional hazards model, a volatility-adjusted trend score, and a counterfactual optimization loop to select and execute the minimum-capital intervention that prevents liquidation, achieving this in imminent-risk cases where static rules fail while maintaining a zero worsening rate and filtering out negligible dust events.

What carries the argument

The counterfactual optimization loop that simulates potential user actions to identify the minimum capital required to mitigate risk, driven by the return period metric and volatility-adjusted trend score.

If this is right

  • The agent prevents liquidations in scenarios where static thresholds fail.
  • It maintains a zero worsening rate across all executed interventions.
  • It distinguishes actionable financial risks from negligible dust events and optimizes capital use accordingly.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the zero-worsening guarantee holds beyond simulation, the agent could serve as a safety layer for other autonomous financial tools that currently lack such bounds.
  • Protocol designers might consider embedding similar intervention loops to reduce systemic liquidation cascades during volatility spikes.

Load-bearing premise

The high-fidelity Aave v3 simulator accurately captures real user behavior, market dynamics, and liquidation mechanics without artifacts that favor the agent.

What would settle it

Live application of the agent to actual high-risk Aave v3 positions followed by comparison of liquidation outcomes and risk changes against a static-rule baseline on the same cohort.

Figures

Figures reproduced from arXiv: 2604.14583 by Fernando Spadea, Oshani Seneviratne.

Figure 1
Figure 1. Figure 1: Overview of our agentic liquidation-prevention action [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Simulator state fidelity and transaction feasibility. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Liquidation prediction behavior under dynamic risk modeling. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
read the original abstract

Decentralized Finance (DeFi) lending protocols like Aave v3 rely on over-collateralization to secure loans, yet users frequently face liquidation due to volatile market conditions. Existing risk management tools utilize static health-factor thresholds, which are reactive and fail to distinguish between administrative "dust" cleanup and genuine insolvency. In this work, we propose an autonomous agent that leverages time-to-event (survival) analysis and moves beyond prediction to execution. Unlike passive risk signals, this agent perceives risk, simulates counterfactual futures, and executes protocol-faithful interventions to proactively prevent liquidations. We introduce a return period metric derived from a numerically stable XGBoost Cox proportional hazards model to normalize risk across transaction types, coupled with a volatility-adjusted trend score to filter transient market noise. To select optimal interventions, we implement a counterfactual optimization loop that simulates potential user actions to find the minimum capital required to mitigate risk. We validate our approach using a high-fidelity, protocol-faithful Aave v3 simulator on a cohort of 4,882 high-risk user profiles. The results demonstrate the agent's ability to prevent liquidations in imminent-risk scenarios where static rules fail, effectively "saving the unsavable" while maintaining a zero worsening rate, providing a critical safety guarantee often missing in autonomous financial agents. Furthermore, the system successfully differentiates between actionable financial risks and negligible dust events, optimizing capital efficiency where static rules fail.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes an autonomous agentic framework for DeFi liquidation prevention in protocols like Aave v3. It combines survival analysis via a numerically stable XGBoost Cox proportional hazards model to derive a return period metric, a volatility-adjusted trend score to filter noise, and a counterfactual optimization loop that simulates user actions to identify minimal capital interventions. The approach is validated on a cohort of 4,882 high-risk user profiles using a high-fidelity Aave v3 simulator, with claims that the agent prevents liquidations in imminent-risk cases where static health-factor thresholds fail, achieves a zero worsening rate, and distinguishes actionable risks from dust events.

Significance. If the simulator fidelity and empirical results hold, the work could advance proactive risk management in DeFi by moving beyond reactive static rules to simulation-driven interventions with explicit safety guarantees. The combination of survival modeling with agentic execution and the emphasis on capital efficiency are potentially valuable contributions to autonomous financial agents.

major comments (3)
  1. [Abstract and Results] Abstract and Results section: The central claims of preventing liquidations in imminent-risk scenarios, 'saving the unsavable,' and maintaining a zero worsening rate on 4,882 profiles rest on unshown empirical evidence; no performance metrics, confidence intervals, ablation studies, or simulator validation details against on-chain traces are supplied, making it impossible to evaluate the strength of the safety guarantee.
  2. [Methodology] Methodology (counterfactual optimization loop and return period definition): The return period metric is produced by the fitted XGBoost Cox model and the counterfactual loop optimizes interventions against outputs from the same model, so reported risk reductions are partly a function of the fitted parameters rather than independent predictions; this circularity undermines the claim of an independent safety guarantee.
  3. [Simulator validation] Simulator validation subsection: The high-fidelity Aave v3 simulator is asserted to be protocol-faithful, but no evidence is provided that it reproduces real user decision-making after interventions, market-impact/slippage effects, or exact liquidation mechanics under volatility; without such calibration, the zero worsening rate and differentiation of dust vs. risk become simulator artifacts rather than transferable results.
minor comments (2)
  1. [Methodology] Clarify the exact functional form and numerical stability procedure for the XGBoost Cox model and the volatility-adjusted trend score, including any free parameters.
  2. [Methodology] Add equations defining the return period metric and the counterfactual objective function to improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We appreciate the referee's thorough evaluation of our manuscript on the agentic survival analysis framework for DeFi liquidation prevention. We have prepared point-by-point responses to the major comments and will incorporate revisions to enhance the empirical support and methodological transparency.

read point-by-point responses

  1. Referee: Abstract and Results section: The central claims of preventing liquidations in imminent-risk scenarios, 'saving the unsavable,' and maintaining a zero worsening rate on 4,882 profiles rest on unshown empirical evidence; no performance metrics, confidence intervals, ablation studies, or simulator validation details against on-chain traces are supplied, making it impossible to evaluate the strength of the safety guarantee.

    Authors: We agree that the abstract and results presentation would benefit from more explicit quantitative support. Although the manuscript describes the outcomes on the 4,882 profiles, including the zero worsening rate, we will revise the abstract to incorporate specific performance metrics such as the proportion of prevented liquidations and capital efficiency gains. We will also add confidence intervals using bootstrap resampling of the cohort, ablation studies that isolate the contribution of the XGBoost Cox model and the counterfactual loop, and a detailed account of simulator validation against historical on-chain liquidation traces from Aave v3. These changes will provide the necessary evidence to substantiate the safety guarantees. revision: yes

  2. Referee: Methodology (counterfactual optimization loop and return period definition): The return period metric is produced by the fitted XGBoost Cox model and the counterfactual loop optimizes interventions against outputs from the same model, so reported risk reductions are partly a function of the fitted parameters rather than independent predictions; this circularity undermines the claim of an independent safety guarantee.

    Authors: We value this observation on potential circularity. The XGBoost Cox model is trained once on historical data to estimate the baseline hazard and covariate effects. For each high-risk profile, the return period is calculated using this trained model. The counterfactual optimization then applies hypothetical interventions to the profile features and evaluates the updated return period using the unchanged model parameters. This is analogous to using a predictive model for what-if analysis in other domains and does not involve refitting during the loop. To mitigate concerns, we will include in the revised manuscript an explicit statement on the separation between training and simulation phases, along with validation on a held-out test set of profiles to show that predicted risk reductions align with observed outcomes where interventions occurred naturally. revision: partial

  3. Referee: Simulator validation subsection: The high-fidelity Aave v3 simulator is asserted to be protocol-faithful, but no evidence is provided that it reproduces real user decision-making after interventions, market-impact/slippage effects, or exact liquidation mechanics under volatility; without such calibration, the zero worsening rate and differentiation of dust vs. risk become simulator artifacts rather than transferable results.

    Authors: We concur that calibration evidence is vital. The simulator implements the exact Aave v3 liquidation logic, including health factor thresholds, bonus calculations, and interest rate models as per the protocol's smart contracts. For market impact and slippage, we use empirical liquidity curves from major DEXes at the time of each profile. While real user decision-making post-intervention cannot be perfectly replicated without behavioral data, the simulator assumes protocol-compliant responses. In the revision, we will augment the simulator validation subsection with comparisons of simulated liquidation frequencies to actual on-chain events, sensitivity analyses for volatility and slippage parameters, and case studies showing how the agent differentiates dust from risk in line with historical patterns. This will confirm the results are grounded rather than artifacts. revision: yes

Circularity Check

1 steps flagged

Counterfactual optimization loop minimizes risk metric produced by the same fitted XGBoost Cox model

specific steps
  1. fitted input called prediction [Abstract]
    "We introduce a return period metric derived from a numerically stable XGBoost Cox proportional hazards model to normalize risk across transaction types, coupled with a volatility-adjusted trend score to filter transient market noise. To select optimal interventions, we implement a counterfactual optimization loop that simulates potential user actions to find the minimum capital required to mitigate risk."

    The optimization loop selects capital adjustments that reduce the return-period value produced by the fitted Cox model; the subsequent claim that the agent 'prevents liquidations' and achieves 'zero worsening rate' is therefore evaluated on the same model's predictions rather than on an external liquidation oracle.

full rationale

The paper derives a return-period risk metric from a fitted XGBoost Cox model, then uses that exact metric inside a counterfactual optimization loop to select interventions. Validation results (risk reduction, zero worsening rate, saving unsavable cases) are therefore measured against the model's own outputs rather than an independent ground-truth liquidation process. This matches the fitted-input-called-prediction pattern; the central safety guarantee is partly forced by construction once the model parameters are fixed. No other circular steps (self-citation chains, ansatz smuggling, or renaming) appear in the provided derivation.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 2 invented entities

The framework depends on a fitted machine-learning model whose parameters are not independent of the data and on a custom simulator whose fidelity is asserted but not independently verified.

free parameters (2)
  • XGBoost Cox model hyperparameters and coefficients
    Fitted to historical liquidation data to produce the return-period metric.
  • Volatility-adjusted trend score parameters
    Chosen or tuned to filter noise; exact values not stated.
axioms (1)
  • domain assumption The Aave v3 simulator faithfully reproduces protocol liquidation logic and user transaction outcomes.
    Invoked for all validation results.
invented entities (2)
  • Return period metric no independent evidence
    purpose: Normalize risk across different transaction types
    Derived directly from the fitted Cox model outputs.
  • Volatility-adjusted trend score no independent evidence
    purpose: Filter transient market noise from genuine risk signals
    New composite score introduced in the paper.

pith-pipeline@v0.9.0 · 5555 in / 1550 out tokens · 46521 ms · 2026-05-10T11:54:24.764823+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

  1. [1]

    Aave v3 technical paper,

    E. Frangella and L. Herskind, “Aave v3 technical paper,” Aave, Tech. Rep., jan 2022, accessed: 2026-01-11. [Online]. Available: https://github.com/aave/aave-v3-core/blob/master/techpaper/Aave V3 Technical Paper.pdf

    work page 2022

  2. [2]

    Health Factor & Liquidations

    Aave. Health Factor & Liquidations. Aave. Accessed: 2025-12-24. [Online]. Available: https://aave.com/help/borrowing/liquidations

    work page 2025

  3. [3]

    Fin- survival: A suite of large scale survival modeling tasks from finance,

    A. Green, Z. Nie, H. Qin, O. Seneviratne, and K. P. Bennett, “Fin- survival: A suite of large scale survival modeling tasks from finance,” Journal of Data-centric Machine Learning Research, 2026

    work page 2026

  4. [4]

    Regression models and life-tables,

    D. R. Cox, “Regression models and life-tables,”Journal of the Royal Statistical Society: Series B (Methodological), vol. 34, no. 2, pp. 187– 202, 1972

    work page 1972

  5. [5]

    Efficiency mode (e-mode) — Aave help & support,

    Aave, “Efficiency mode (e-mode) — Aave help & support,” https://aa ve.com/help/borrowing/e-mode, 2026, accessed: 2026-01-11

    work page 2026

  6. [6]

    Isolation mode — Aave help & support,

    ——, “Isolation mode — Aave help & support,” https://aave.com/help/ supplying/isolation-mode, 2026, accessed: 2026-01-11

    work page 2026

  7. [7]

    Enabling cross-language data integration and scalable analytics in decentralized finance,

    C. Flynn, K. P. Bennett, J. S. Erickson, A. Green, and O. Seneviratne, “Enabling cross-language data integration and scalable analytics in decentralized finance,” in2023 IEEE International Conference on Big Data (BigData), 2023, pp. 4290–4299

    work page 2023

  8. [8]

    Characterizing common quarterly behaviors in defi lending protocols,

    A. Green, M. Giannattasio, K. Wang, J. S. Erickson, O. Seneviratne, and K. P. Bennett, “Characterizing common quarterly behaviors in defi lending protocols,” inMathematical Research for Blockchain Economy, P. Pardalos, I. Kotsireas, W. J. Knottenbelt, and S. Leonardos, Eds. Cham: Springer Nature Switzerland, 2023, pp. 62–77

    work page 2023

  9. [9]

    Defi survival analysis: Insights into risks and user behaviors,

    A. Green, C. Cammilleri, J. S. Erickson, O. Seneviratne, and K. P. Bennett, “Defi survival analysis: Insights into risks and user behaviors,” inMathematical Research for Blockchain Economy, P. Pardalos, I. Kot- sireas, Y . Guo, and W. Knottenbelt, Eds. Cham: Springer International Publishing, 2023, pp. 127–141

    work page 2023

  10. [10]

    DeFi Survival Analysis: Insights Into the Emerging Decentralized Financial Ecosystem,

    A. M. Green, M. P. Giannattasio, J. S. Erickson, O. Seneviratne, and K. P. Bennett, “DeFi Survival Analysis: Insights Into the Emerging Decentralized Financial Ecosystem,”ACM Distributed Ledger Technologies (DLT) Research and Practice, 2023. [Online]. Available: https://doi.org/10.1145/3638064

    work page doi:10.1145/3638064 2023

  11. [11]

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

    F. Spadea, A. Talukder, K. P. Bennett, and O. Seneviratne, “The Predictability of Risk: Linking User Action Frequency to Survival Model Performance,”TechRxiv, 2025, presented at the AI for Resilient & Responsible DeFi Dynamics (AR2D2) Workshop at the 6th ACM International Conference on AI in Finance (ICAIF ’25). [Online]. Available: https://www.techrxiv.o...

    work page arXiv 2025

  12. [12]

    Liquidity math in uniswap v3,

    A. Elsts, “Liquidity math in uniswap v3,”Available at SSRN 4575232, 2021

    work page 2021

  13. [13]

    The suitability of using uniswap v2 model to analyze v3 data,

    N. Chemaya and D. Liu, “The suitability of using uniswap v2 model to analyze v3 data,”Finance Research Letters, vol. 59, p. 104717, 2024

    work page 2024

  14. [14]

    Lehar and C

    A. Lehar and C. A. Parlour,Systemic fragility in decentralized markets. Bank for International Settlements, Monetary and Economic Depart- ment, 2022

    work page 2022

  15. [15]

    Cefi vs. defi–comparing centralized to decentralized finance,

    K. Qin, L. Zhou, Y . Afonin, L. Lazzaretti, and A. Gervais, “Cefi vs. defi–comparing centralized to decentralized finance,”arXiv preprint arXiv:2106.08157, 2021

    work page arXiv 2021

  16. [16]

    Aave V4’s new liquidation engine,

    Aave Labs, “Aave V4’s new liquidation engine,” https://aave.com/blog/ aave-v4-liquidations, 2024, accessed: 2026-01-11

    work page 2024

  17. [17]

    An empirical study of defi liquidations: Incentives, risks, and instabilities,

    K. Qin, L. Zhou, P. Gamito, P. Jovanovic, and A. Gervais, “An empirical study of defi liquidations: Incentives, risks, and instabilities,” inProceedings of the 21st ACM internet measurement conference, 2021, pp. 336–350

    work page 2021

  18. [18]

    (2025, dec) Aave Introduces V4 Liquidation Engine to Enhance Efficiency

    Binance News. (2025, dec) Aave Introduces V4 Liquidation Engine to Enhance Efficiency. Binance Square. [Online]. Available: https: //www.binance.com/en/square/post/12-12-2025-aave-introduces-v4-liq uidation-engine-to-enhance-efficiency-33602881765090

    work page 2025

  19. [19]

    (2021, jun) Aave protocol liquidation retrospective: May

    Watson Fu. (2021, jun) Aave protocol liquidation retrospective: May

    work page 2021

  20. [20]

    [Online]

    Gauntlet. [Online]. Available: https://www.gauntlet.xyz/resources /aave-protocol-liquidation-retrospective-may-2021

    work page 2021

  21. [21]

    All-in-one crypto lending dashboard with automation,

    DeFi Saver, “All-in-one crypto lending dashboard with automation,” ht tps://defisaver.com/features/lending-and-borrowing, 2024, accessed: 2024-05-22

    work page 2024

  22. [22]

    Instadapp: The unified platform for everything DeFi,

    Instadapp, “Instadapp: The unified platform for everything DeFi,” https: //instadapp.io, 2026, accessed: 2026-01-11

    work page 2026

  23. [23]

    0xjacobzhao and L. Sokolin. (2025, aug) AgentFi 101: The Definitive Guide to Decentralized Financial Agents. Fintech Blueprint. [Online]. Available: https://lex.substack.com/p/agentfi-101-the-definitive-guide

    work page 2025

  24. [24]

    Elsa: AI crypto copilot,

    Elsa AI, “Elsa: AI crypto copilot,” https://heyelsa.ai, 2026, accessed: 2026-01-11

    work page 2026

  25. [25]

    Bankr: The AI agent for your crypto portfolio,

    Bankr, “Bankr: The AI agent for your crypto portfolio,” https://bankr.bot, 2026, accessed: 2026-01-11

    work page 2026

  26. [26]

    Bringing AI agents onchain: Introducing the ConsoleKit,

    Brahma, “Bringing AI agents onchain: Introducing the ConsoleKit,” ht tps://brahma.fi/blog/bringing-ai-agents-onchain-introducing-the-conso lekit, Jan. 2025, accessed: 2026-01-11

    work page 2025

  27. [27]

    DeFi liquidation predictors — research suite,

    The Satoshi Terminal, “DeFi liquidation predictors — research suite,” https://docs.satoshiterminal.io/research-suite/analytics/defi-analytics /insights-and-predictive-analytics/defi-liquidation-predictors, 2025, accessed: 2026-01-11

    work page 2025

  28. [28]

    Aave liquidation MCP server,

    clumsynonono, “Aave liquidation MCP server,” https://glama.ai/mcp/s ervers/@clumsynonono/aave-liquidation-mcp, 2026, accessed: 2026- 01-11

    work page 2026

  29. [29]

    Autofinsurv: An automatic xgboost-cox optimization framework for financial survival analysis,

    R.-L. Jian, “Autofinsurv: An automatic xgboost-cox optimization framework for financial survival analysis,” Singapore, November 2025. [Online]. Available: https://finsurvival.github.io/papers/1 Hierarchical Feature Engineering.pdf

    work page 2025

  30. [30]

    Finsurvival challenge

    K. P. Bennett, C. Curran, A. M. Green, A. Pavao, O. Seneviratne, and F. Spadea, “Finsurvival challenge.” [Online]. Available: https: //finsurvival.github.io

    work page

  31. [31]

    Benchmarking Temporal Web3 Intelligence: Lessons from the FinSurvival 2025 Challenge,

    O. Seneviratne, F. Spadea, A. Pavao, A. M. Green, and K. P. Ben- nett, “Benchmarking Temporal Web3 Intelligence: Lessons from the FinSurvival 2025 Challenge,”ACM, 2026, 16th Temporal Web Analytics Workshop (TempWeb 2026)

    work page 2025

  32. [32]

    Instabilities using cox ph for forecasting or stress testing loan portfolios,

    J. L. Breeden, A. Bellotti, E. Leonova, and A. Yablonski, “Instabilities using cox ph for forecasting or stress testing loan portfolios,” inCredit Scoring and Credit Control Conference XIV, 2015

    work page 2015