DeXposure: A Dataset and Benchmarks for Inter-protocol Credit Exposure in Decentralized Financial Networks

AI summary

Overview

Research area: Decentralized Finance (DeFi), financial network analysis, and machine learning on graphs.

Technical level: Advanced. The paper assumes familiarity with graph theory (directed weighted graphs, assortativity, degree centralization), temporal graph neural networks, vector autoregression, and TF-IDF text matching.

Scope: The paper releases DeXposure, a large-scale dataset of inter-protocol credit exposure in DeFi (43.7 million entries covering 4.3 thousand protocols, 602 blockchains, and 24.3 thousand tokens from 2020 to 2025), together with three machine learning benchmarks built on it.

What This Paper Is About

DeFi protocols lend, trade, and hold each other's tokens, so when one protocol fails, others holding its tokens are exposed. Existing research studied this only for small sets of protocols, via case studies, or with transaction-level data, leaving no system-wide quantitative map of who owes what to whom across the ecosystem. The goal of this paper is to build that map: a large-scale, longitudinal dataset that infers "value-linked credit exposure" between protocols from changes in Total Value Locked (TVL), and to package it as benchmarks for machine learning research.

Key Contributions

1. The DeXposure dataset — described as the first large-scale dataset for inter-protocol credit exposure in decentralized financial networks, comprising 43.7 million entries across 4.3 thousand protocols, 602 blockchains, and 24.3 thousand unique tokens, spanning 2020 to 2025, with daily JSON (and CSV) snapshots of nodes and directed exposure links.

2. A new measure: value-linked credit exposure — financial dependency relationships between protocols inferred from changes in TVL. Protocol p has credit exposure to protocol q when the tokens issued by q are currently locked in p; the strength of exposure is measured by token value flows between them over a time interval.

3. A token-to-protocol mapping procedure — a four-stage fall-back pipeline (DefiLlama metadata lookup, manual expert mapping for high-TVL tokens, TF-IDF text similarity with cosine similarity and a threshold, and a catch-all "Primary Market Tokens" category) that attributes tokens held in one protocol to the protocol that issues them.

4. Three machine learning benchmarks — (1) graph clustering for global network measurement, (2) vector autoregression for sector-level exposure dynamics during the Terra and FTX shocks, and (3) temporal graph neural networks for dynamic link prediction on temporal graphs.

Main Findings

• Rapid growth of network volume: The authors report a rapid growth of network volume over the period covered by the dataset.
• Concentration in key protocols: The analysis reveals a trend of concentration toward key protocols.
• Declining network density: Network density — the ratio of actual connections to possible connections — declines over time.
• Distinct shock propagation across sectors: Major shocks (Terra and FTX) propagate differently across sectors, including lending platforms, trading exchanges, and asset management protocols.
• Benchmark metric values not reported in the available text: The truncated paper content includes the formal definitions of metrics such as degree centralization, degree coefficient of variation, degree distribution entropy, top 10% degree concentration, and assortativity, but the numerical results for the benchmarks (including link prediction scores) are not reported in the content available here.
• Related-work context cited: Prior work is summarized as covering small protocol sets — for example, compositions of 23 major DeFi protocols on Ethereum, systemic risk in 30 DeFi protocols, 12 major DeFi applications over a 34-month period, 5.8 million transactions of three DeFi tokens on Ethereum, a 2-year analysis of the Curve ecosystem, and 100 DeFi protocols analyzed with a "Total Value Redeemable" measure — which the authors position as motivation for a system-wide dataset.

Methodology in Plain English

• Data collection. Daily TVL data is retrieved from DefiLlama for protocols, chains, and tokens. The paper describes a heatmap of data availability and frequency across time for DeFi protocols, where darker shades indicate more frequent updates (for example hourly) and lighter shades less frequent updates (for example daily).
• Formal data model. Protocols, chains, and tokens are defined as sets, with a mapping from protocol-chain pairs to subsets of tokens. Each token at a time t carries an amount and a USD value. These aggregate into global, protocol-wise, and chain-wise token states.
• Defining a credit link. If protocol q issues a token and protocol p currently holds that token, then p is exposed to q. A worked example is given: a user wraps ETH into WETH, then uses WETH as collateral in MakerDAO to generate DAI, creating credit exposure from MakerDAO to the WETH protocol.
• Turning holdings into a network. Each daily snapshot is a weighted directed graph. Nodes are protocols weighted by their total asset value in USD. Edges are built from token value flows between protocols over a time interval, with the direction reversed when the flow is negative so weights stay positive. The edge weight represents the change in credit exposure from protocol p to the issuing protocol q.
• Matching tokens to issuers. The issuing protocol for a token is determined by a four-step fall-back procedure: direct metadata lookup, manual mapping for the highest-TVL tokens, TF-IDF vectorization of token and protocol descriptions plus cosine similarity above a threshold, and finally a fall-back protocol named after the token itself.
• Benchmarking. On the resulting temporal graph sequence, the authors run three experiments: network measurement and clustering, vector autoregression around the Terra and FTX shocks, and temporal graph neural networks for dynamic link prediction.

Why This Matters

Impact on research. The paper contributes both a data resource and a machine learning benchmark suite. It provides longitudinal, ecosystem-wide data on interconnected protocol balance sheets, which previous datasets — largely segmented by verticals such as fraud detection, specific DEXs, or event types — did not offer. It also introduces a measurable definition of credit exposure that can be reused and challenged.

Real-world applications.

• Financial stability monitoring: tracking exposure concentrations and network topology over time as inputs to systemic risk assessment and early warning systems.
• Regulatory and policy analysis: giving central bankers, regulators, and policymakers a quantitative, system-wide view of how DeFi protocols depend on one another.
• DeFi market modeling: supporting modeling of how leverage, composability, and interconnectedness evolve.
• Shock propagation analysis: examining how stress at one protocol or sector (as with Terra and FTX) travels through lending platforms, exchanges, and asset management protocols.

Industry relevance. The dataset and code are released publicly, and results are presented through an interactive DeFi visualisation tool hosted by the Cambridge Centre for Alternative Finance, making the output usable by practitioners as well as researchers.

Future Directions

• Extending the mapping: the token-to-protocol mapping relies on metadata, manual curation for high-TVL tokens, and TF-IDF similarity thresholds; improving coverage and accuracy for tokens that fall into the catch-all "Primary Market Tokens" category is an open problem.
• Addressing TVL accounting limits: the paper discusses double counting and rehypothecation in TVL, and argues that excluding double counting by a token's stage in circulation may hide leverage and interconnectedness; how best to reconcile credit and liquidity risk in these measures remains unresolved.
• Broadening the benchmark tasks: the paper frames graph clustering, vector autoregression, and temporal link prediction as starting points and states that it envisions the dataset supporting further work in machine learning and financial risk monitoring, DeFi market modeling, and related areas.
• Deeper temporal and cross-chain analysis: the dataset covers 602 blockchains and spans 2020 to 2025, leaving open questions about cross-chain exposure pathways and about modeling temporal relationships beyond snapshot-by-snapshot construction.

Target Audience

Machine learning researchers working on temporal graphs, graph clustering, and link prediction with financial applications; financial stability researchers, central bankers, and regulators studying DeFi systemic risk; and DeFi analysts and data scientists who need a large-scale, longitudinal dataset of protocol-level credit exposure.

Authors’ abstract

We curate the DeXposure dataset, the first large-scale dataset for inter-protocol credit exposure in decentralized financial networks, covering global markets of 43.7 million entries across 4.3 thousand protocols, 602 blockchains, and 24.3 thousand tokens, from 2020 to 2025. A new measure, value-linked credit exposure between protocols, is defined as the inferred financial dependency relationships derived from changes in Total Value Locked (TVL). We develop a token-to-protocol model using DefiLlama metadata to infer inter-protocol credit exposure from the token's stock dynamics, as reported by the protocols. Based on the curated dataset, we develop three benchmarks for machine learning research with financial applications: (1) graph clustering for global network measurement, tracking the structural evolution of credit exposure networks, (2) vector autoregression for sector-level credit exposure dynamics during major shocks (Terra and FTX), and (3) temporal graph neural networks for dynamic link prediction on temporal graphs. From the analysis, we observe (1) a rapid growth of network volume, (2) a trend of concentration to key protocols, (3) a decline of network density (the ratio of actual connections to possible connections), and (4) distinct shock propagation across sectors, such as lending platforms, trading exchanges, and asset management protocols. The DeXposure dataset and code have been released publicly. We envision they will help with research and practice in machine learning as well as financial risk monitoring, policy analysis, DeFi market modeling, amongst others. The dataset also contributes to machine learning research by offering benchmarks for graph clustering, vector autoregression, and temporal graph analysis.