FinSurvival: Benchmark for DeFi Survival Analysis

• FinSurvival is a benchmark suite for survival analysis in DeFi lending, defining 16 time-to-event prediction tasks with significant right censoring.
• It constructs survival records from Aave transaction data with engineered features that capture user behavior, market trends, and temporal dynamics.
• The benchmark highlights challenges like temporal drift and censoring, promoting robust methodologies applicable to broader Web3 and temporal analytics contexts.

Searching arXiv for the FinSurvival papers and related extensions to ground the article. FinSurvival is a public benchmark suite for survival analysis built from decentralized-finance lending data, and, in its later challenge form, a methodological case study in what its authors call “temporal Web3 intelligence.” It formulates user behavior in DeFi as right-censored time-to-event prediction: given an index event such as a borrow or deposit, the task is to predict the duration until a specified subsequent outcome event occurs. Across its published forms, FinSurvival uses Aave transaction histories to instantiate 16 survival tasks, emphasizes longitudinal temporal structure rather than static prediction, and positions DeFi event streams as a high-fidelity environment for benchmarking survival modeling under censoring, temporal drift, and evolving user behavior (Green et al., 7 Jul 2025).

1. Origins and conceptual framing

FinSurvival was introduced as “a suite of 16 survival modeling tasks from finance” derived from publicly available DeFi lending transactions, with the stated goal of supplying a large-scale, realistic, and freely available benchmark for AI survival modeling (Green et al., 7 Jul 2025). The original benchmark paper frames survival modeling as the problem of predicting time until an event occurs under censoring, and argues that public datasets suitable for evaluating modern survival methods are often small, private, expensive, or highly restricted. FinSurvival is presented as a response to that gap.

The later “FinSurvival Challenge 2025” broadened this framing. Rather than presenting the benchmark only as a finance dataset, it treated FinSurvival as a case study in “temporal Web3 intelligence,” emphasizing that many web-scale prediction problems are fundamentally temporal: users arrive, engage, disengage, default, churn, return, or evolve over long horizons, while the environment itself changes over time (Seneviratne et al., 26 Feb 2026). In that view, Web3 systems are distinctive because they expose immutable, time-stamped, public event streams, making censoring, temporal drift, and strategic behavior unusually visible.

This produces a dual identity. On one level, FinSurvival is a survival-analysis benchmark suite instantiated on Aave transaction histories. On another, it is a benchmark-design argument: DeFi is used as a high-fidelity sandbox for studying temporal prediction problems that also arise in conventional Web, FinTech, SaaS, e-commerce, and other user-lifecycle settings. This suggests that FinSurvival is best understood not only as a finance benchmark, but also as a template for longitudinal, censoring-aware temporal analytics.

2. Data sources and benchmark construction

The original FinSurvival benchmark is built from transaction data collected from The Graph, specifically the Aave V2 Ethereum subgraph, covering November 30, 2020 through September 30, 2024 (Green et al., 7 Jul 2025). The raw source includes five core transaction types relevant to lending behavior: deposits, borrows, repays, withdraws, and liquidations. The paper reports a raw transaction corpus of 1,977,491 transactions, 117,008 users, 38 raw features, and a CSV size of 668.8 MB, which expands into 7,698,497 survival records because each transaction can seed a time-to-event record under multiple index/outcome definitions.

The challenge edition re-instantiated the same benchmarking framework on a new protocol deployment: Aave v3 on Polygon. It used over 21.8 million transaction records, preserved the same benchmark codebase, and replaced the earlier 128-feature Aave v2/Ethereum dataset with a new Aave v3/Polygon dataset containing 90 engineered features (Seneviratne et al., 26 Feb 2026). The distinction matters because the challenge was not a simple reuse of the earlier benchmark, but a new longitudinal environment with different temporal and market conditions.

The benchmark construction rule is event-centric. A user action becomes an index event, the clock starts there, and the task is to predict how long it takes until a later outcome event for the same subject. In the original benchmark, subjects are defined primarily by user and coin (reserve), so a repayment must match the same reserve as the originating borrow (Green et al., 7 Jul 2025). The automated pipeline takes an event table, one or more subject columns, an observation period, an index event set, and an outcome event set; it filters to the observation period, groups by subject, performs a rolling join from each index event to the first later outcome event with the same subject, computes elapsed time, and uses the observation-period end as censoring time if no valid future outcome exists.

The original paper formalizes survival records as

$$\{(T_i, \delta_i, \mathbf{X}_i) : i = 1, 2, \ldots, n\},$$

where $T_i$ is the observed time, $\delta_i$ is the event indicator, and $\mathbf{X}_i$ is the feature vector (Green et al., 7 Jul 2025). Here, $T_i$ is either time from index event to outcome event or time from index event to end of follow-up, while $\delta_i=1$ indicates an observed outcome and $\delta_i=0$ indicates censoring.

Benchmark instance	Data source	Scale
Original FinSurvival	Aave V2 Ethereum via The Graph	7,698,497 survival records, 128 features
FinSurvival Challenge 2025	Aave v3 Polygon	over 21.8 million transaction records, 90 engineered features

This progression from Aave v2/Ethereum to Aave v3/Polygon reflects a broader methodological point made in the later paper: benchmark code and task logic can remain stable while the underlying temporal environment changes. A plausible implication is that FinSurvival is intended as a reusable pipeline for generating future survival benchmarks as DeFi protocols and chains evolve.

3. Task suite, temporal design, and censoring

FinSurvival’s task family is generated from ordered transitions among four basic transaction types—Borrow, Deposit, Repay, and Withdraw—with Liquidated used as an outcome event. The original paper states that the 16 survival tasks are the 12 basic pairings among borrow/deposit/repay/withdraw plus 4 liquidation-outcome tasks (Green et al., 7 Jul 2025). The challenge paper presents the same family as all ordered transitions among four index event types, excluding self-transitions, with Liquidated included as an outcome for each index event, yielding the same 16 index–outcome task pairs (Seneviratne et al., 26 Feb 2026).

The challenge paper lists the task structure explicitly:

• Borrow $\rightarrow$ {Deposit, Repay, Withdraw, Liquidated}
• Deposit $\rightarrow$ {Borrow, Repay, Withdraw, Liquidated}
• Repay $\rightarrow$ {Borrow, Deposit, Withdraw, Liquidated}
• Withdraw $T_i$0 {Borrow, Deposit, Repay, Liquidated}

Conceptually, each task is a time-to-event problem: given that a user has just experienced an index event, predict the elapsed time until a specified outcome event occurs. The challenge paper also notes that this implicitly defines a competing-risks setting, because several mutually exclusive future outcomes are possible after a given index event, and if the target outcome is not observed during the observation window, the instance is right-censored (Seneviratne et al., 26 Feb 2026).

Temporal design is central in both versions, but the challenge makes it especially explicit. The original benchmark uses a temporal train/test split with cutoff date July 1, 2022, yielding roughly 60% of transactions before the cutoff and 40% after, with a 30-day buffer at the end of both train and test periods so that included index events have enough follow-up time to realize an outcome (Green et al., 7 Jul 2025). The challenge uses a strict chronological split: training spans March 12, 2022 to July 2, 2024, and testing spans July 3, 2024 to August 21, 2025 (Seneviratne et al., 26 Feb 2026). The challenge paper repeatedly emphasizes that this is meant to expose non-stationarity rather than hide it.

Censoring is treated as a first-class property rather than a nuisance to be filtered away. The original benchmark reports no left-censored records because the authors collected all Aave V2 Ethereum transactions since launch, but it reports a mean censoring rate of 81.26% across the 16 tasks (Green et al., 7 Jul 2025). The challenge paper similarly defines the setting as right-censored time-to-event prediction and argues that incomplete user trajectories should be retained because incomplete user lifecycles are normal in web-scale systems (Seneviratne et al., 26 Feb 2026).

The original benchmark also provides illustrative task statistics showing the heterogeneity of difficulty. For example, borrow $T_i$1 repay has 267,010 records, mean delay 78.80 days, and 16.62% censored; deposit $T_i$2 withdraw has 629,082 records, mean delay 198.42 days, and 57.15% censored; withdraw $T_i$3 liquidated has 399,704 records, mean delay 360.26 days, and 99.34% censored; and deposit $T_i$4 liquidated has 507,437 records and 98.73% censored (Green et al., 7 Jul 2025). These figures make clear that liquidation tasks are extremely sparse relative to more routine behavioral transitions.

4. Feature design and evaluation framework

Feature construction is one of FinSurvival’s defining design choices. The original benchmark provides 128 features per record, including raw transaction fields and engineered summaries of user history, market history, and time (Green et al., 7 Jul 2025). It specifies 19 temporal features, 45 user-history features, and 40 market-history features, all computed using only information available up to and including the index event. The challenge version reduces the provided set to 90 engineered features, grouped into three families: user-centric, market-level, and transactional features (Seneviratne et al., 26 Feb 2026). User-centric features capture historical activity, total deposit and borrow sums, and interaction frequency; market-level features capture protocol-wide liquidity, average deposit amounts, and asset prices; transactional features include current transaction size in USD and cyclical temporal encodings such as sine/cosine representations of time-of-day.

Both versions emphasize anti-leakage design. The challenge paper states a strict anti-leakage principle: no forward-looking or post-outcome features were included; all covariates reflected information available before the event being predicted (Seneviratne et al., 26 Feb 2026). This makes the benchmark operationally realistic while reducing the burden on participants to reconstruct full protocol state from raw blockchain logs.

The formal evaluation metric for the challenge is the Concordance Index (C-index), defined as

$T_i$5

where $T_i$6 are observed event times, $T_i$7 are predicted risk scores, and $T_i$8 is the indicator function (Seneviratne et al., 26 Feb 2026). The leaderboard score is the average of the 16 task-level C-index values. The challenge paper justifies this choice on three grounds: DeFi event times are heavy-tailed, the benchmark is intentionally non-stationary, and the C-index is censoring-aware because it only compares pairs whose ordering is knowable.

The original benchmark likewise uses C-index for survival prediction and reports that $T_i$9 is random ranking and $\delta_i$0 is perfect concordance (Green et al., 7 Jul 2025). It also introduces a companion classification benchmark by thresholding each survival task using the restricted mean survival time (RMST). For each task, RMST is estimated from the Kaplan–Meier curve, the threshold is chosen by increasing $\delta_i$1 in one-day increments until the change in RMST is less than 5%, and labels are assigned as class 1 if the event time is $\delta_i$2, class 0 if time-to-event is $\delta_i$3, with cases censored before the threshold dropped as ambiguous (Green et al., 7 Jul 2025).

This bifurcation between survival prediction and thresholded classification is one of the benchmark’s more informative design choices. The original paper finds that strong performance on continuous survival prediction does not transfer automatically to RMST-thresholded classification: XGBoost is best for survival regression, while logistic regression and elastic net are strongest in classification (Green et al., 7 Jul 2025). This suggests that “predict when” and “predict whether soon enough” are materially different tasks even when derived from the same event histories.

5. Baselines, challenge results, and methodological lessons

The original benchmark evaluates six survival methods: Cox proportional hazards, Accelerated Failure Time (AFT, Weibull), gradient boosting machine (GBM), XGBoost with AFT objective / censored likelihood handling, DeepSurv, and DeepHit (Green et al., 7 Jul 2025). XGBoost is the best overall survival model, with mean C-index 0.783 and mean Borda rank 1.312; AFT is second with mean C-index 0.747 and mean Borda rank 1.750; Cox is markedly weaker at mean C-index 0.522; DeepSurv and DeepHit perform poorly overall at 0.461 and 0.441, respectively. The original paper’s qualitative conclusion is that exact survival prediction remains difficult, heavily censored tasks are challenging, and deep survival models underperform strong classical baselines on this data.

The challenge results sharpen that conclusion. Among the top 16 contestants, the average C-index was 0.750, while the top three teams reached 0.914, 0.849, and 0.847 (Seneviratne et al., 26 Feb 2026).

Rank	Team	Summary
1st	Balancehero	XGBoost AFT, hierarchical features, 10,000+ engineered features, average C-index 0.914
2nd	AutoFinSurv	XGBoost Cox, original 90 features, average C-index 0.849
3rd	FinBoost	mixed XGBoost ensemble with augmented 90+ features, average C-index 0.847

The winning solution, Balancehero, is the clearest empirical statement of the benchmark’s central thesis. It rebuilt transaction histories at multiple levels of abstraction and generated over 10,000 features across 12 interrelated tables, spanning user-level behavioral trajectories, market-level context, and transaction- and liquidation-specific dynamics (Seneviratne et al., 26 Feb 2026). The team used a multi-level indexing scheme such as user_index, market_index, and trx_index to preserve temporal and relational integrity while enabling large-scale joins and aggregations without leakage. The paper also highlights missing-data recovery: unobserved or partially logged transactions and liquidation events were inferred using temporal ordering and count-based validation, which was especially important for liquidation-related outcomes.

Its model architecture was a two-stage XGBoost AFT framework: stage one trained across shared outcome types to learn transferable survival patterns; stage two fine-tuned for specific index–outcome tasks (Seneviratne et al., 26 Feb 2026). The second-place team, AutoFinSurv, instead used the original 90 features with XGBoost and a Cox proportional hazards objective, optimizing hyperparameters across all 16 tasks with Optuna. The third-place team, FinBoost, compared classical and deep survival models including Weibull AFT and DeepSurv, then settled on a tree-based ensemble blending “Conservative,” “Balanced,” and “Aggressive” risk postures.

The authors’ most important empirical lesson is that feature engineering dominated end-to-end modeling (Seneviratne et al., 26 Feb 2026). They give three reasons. First, DeFi behavior is stateful but sparse, and important latent variables are not directly observed. Second, strict temporal splits induce substantial distribution shift, making normalized, market-aware engineered features more robust than models that overfit old regimes. Third, tree-based survival objectives such as Cox and AFT in XGBoost align naturally with a ranking metric like the C-index. The paper summarizes this more conceptually by arguing that the hidden variable in benchmark performance is the method’s implicit theory of time: assumptions about memory, recency, persistence, and risk accumulation are often encoded more effectively through hierarchical temporal representations than inferred directly from sparse event streams.

6. Cross-domain significance, limitations, and extensions

FinSurvival is explicitly positioned as transferable beyond DeFi. The challenge discussion maps DeFi survival tasks to other web domains: Deposit $\delta_i$4 Withdraw corresponds to Subscription $\delta_i$5 Churn in SaaS; Borrow $\delta_i$6 Liquidate corresponds to Credit Line $\delta_i$7 Default in FinTech; Repay $\delta_i$8 Borrow corresponds to Feature Engagement $\delta_i$9 Upsell in e-commerce; and Index Event $\mathbf{X}_i$0 Censoring corresponds to User Session $\mathbf{X}_i$1 Inactivity/Logout (Seneviratne et al., 26 Feb 2026). This mapping is central to the claim that FinSurvival is not merely a protocol-specific finance benchmark, but a general template for user evolution modeling under delayed, censored, and non-stationary outcomes.

The benchmark-design principles extracted from the challenge follow directly from that claim. The paper recommends that next-generation temporal web benchmarks be “longitudinal first,” use explicit censoring rather than filtering incomplete cases, rely on fixed centralized evaluation pipelines, encourage behavioral state modeling, support cross-domain transferability, and make data, code, and evaluation logic openly accessible (Seneviratne et al., 26 Feb 2026). In reproducibility terms, the challenge was hosted on Codabench, used fixed train/test splits, required top-ranked teams to submit open-source code and technical reports, and verified final results by re-running those submissions.

The limitations are equally explicit. The original benchmark covers one protocol family, Aave V2 on Ethereum, while the challenge covers Aave v3 on Polygon; in both cases, model behavior may partly reflect protocol-specific mechanics (Green et al., 7 Jul 2025). The challenge paper adds that off-chain influences such as sentiment, governance, or macro news are absent even though on-chain events are fully visible, and that concordance evaluates only relative ranking, not calibration or absolute-time accuracy (Seneviratne et al., 26 Feb 2026). The original paper similarly notes that competing risks are present but not modeled directly, despite the pipeline supporting them in principle (Green et al., 7 Jul 2025).

Future directions in the challenge paper include multi-protocol and cross-chain data, richer contextual signals, complementary evaluation metrics beyond concordance, hybrid methods that combine automated optimization with structured feature construction, and multitask or cross-chain survival modeling (Seneviratne et al., 26 Feb 2026). A subsequent paper, “From Risk to Rescue: An Agentic Survival Analysis Framework for Liquidation Prevention,” extends the FinSurvival style of modeling from prediction to intervention by embedding survival estimates inside an autonomous decision loop for liquidation prevention on Aave v3 (Spadea et al., 16 Apr 2026). This suggests a broader trajectory for FinSurvival-related work: from benchmark construction, to temporal representation learning, to operational risk intervention systems.

Taken together, the published FinSurvival work establishes a benchmark family in which time is not a secondary attribute of labels but the organizing analytic dimension. Its central contribution is therefore methodological as much as empirical: it demonstrates that realistic temporal benchmarks should preserve censoring, expose non-stationarity, and make temporal state representation measurable. In that sense, FinSurvival is both a concrete suite of DeFi survival tasks and an argument about how longitudinal benchmarks should be designed for survival analysis in Web3 and beyond (Seneviratne et al., 26 Feb 2026).