📈 economics

DeXposure-FM: A Time-series, Graph Foundation Model for Credit Exposures and Stability on Decentralized Financial Networks

DeXposure-FM is a novel time-series graph foundation model trained on a massive multi-chain dataset that outperforms existing methods in forecasting DeFi credit exposures and provides essential tools for macroprudential monitoring and systemic risk stress testing.

Original authors: Aijie Shu, Wenbin Wu, Gbenga Ibikunle, Fengxiang He

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aijie Shu, Wenbin Wu, Gbenga Ibikunle, Fengxiang He

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Mapping the Invisible Web of Money

Imagine the world of Decentralized Finance (DeFi) not as a bank, but as a giant, bustling marketplace built on a digital highway. In this marketplace, thousands of different "shops" (protocols) trade with each other.

The problem? In this digital world, shops don't usually sign formal contracts saying, "I owe you money." Instead, they hold each other's tokens (digital coins) as collateral. If Shop A holds Shop B's tokens, and Shop B's tokens crash in value, Shop A loses money too. This creates a hidden, invisible web of debt and risk. If one shop fails, the shock can ripple through the whole market, causing a chain reaction.

Until now, regulators and experts have struggled to see this web clearly because it changes every second and is incredibly complex.

DeXposure-FM is a new "super-spy" tool designed to map this invisible web, predict how it will change next week, and tell us which shops are most likely to cause a crash.

1. What is DeXposure-FM?

Think of DeXposure-FM as a crystal ball trained on a massive history book.

The Data: The researchers fed the model a massive dataset containing 43.7 million snapshots of the DeFi market from 2020 to 2025. This covers over 4,300 shops, 602 different digital highways (blockchains), and 24,000+ types of tokens.
The Brain: The model uses a "Foundation Model" (a type of AI that learns general patterns first, then specializes). Specifically, it uses a "GraphPFN" brain.
- Analogy: Imagine teaching a student to read by giving them millions of books first (pre-training). Then, you give them a specific textbook on DeFi and ask them to solve math problems based on it (fine-tuning). This model has "read" the entire history of DeFi interactions.
The Job: It looks at the market today and predicts what the web of connections will look like in 1, 4, 8, or 12 weeks.

2. How Does It Work? (The Three Tasks)

The model acts like a detective solving three different puzzles simultaneously:

Will a connection exist? (Link Prediction)
- Analogy: Will Shop A and Shop B still be trading with each other next month? The model predicts if a "line" will appear or disappear between them.
How strong is the connection? (Edge Weight)
- Analogy: If they are trading, how much money is at risk? Is it a small handshake deal or a massive loan? The model guesses the size of the exposure.
How big will the shops get? (Node TVL)
- Analogy: Will Shop A's vault (Total Value Locked) grow or shrink? The model predicts the size of the shops.

3. Why is This Better Than Just Guessing?

In the real world, things often stay the same for a while. If you just assumed "next week will look exactly like this week" (a method called Persistence), you'd be right most of the time because the DeFi market is very stable.

However, the paper found that DeXposure-FM shines when things change.

The "Bad News" Zone: When the market is chaotic (like during a crash or a sudden new rule), the "just copy last week" method fails miserably.
The Model's Superpower: DeXposure-FM is much better at predicting these chaotic moments. It correctly identified that connections would break or form in the "worst 20%" of difficult weeks where simple guesses failed. It acts like a weather forecast that is great at predicting sunny days but essential for predicting storms.

4. What Can We Do With This? (The Financial Tools)

Once the model predicts the future web, the researchers use it to build "dashboard tools" for regulators and risk managers:

Systemic Importance Score (SIS):
- Analogy: A "Most Wanted" list. The model ranks which shops are so connected that if they fall, they take everyone down with them. It helps authorities know who to watch closely.
Spillover Concentration:
- Analogy: A "Contagion Map." It shows if risk is spreading from one specific sector (like "Stablecoins") to another (like "Bridges"). It highlights where a fire might jump from one building to another.
Stress Testing (The "What If" Machine):
- Analogy: A simulation game. You can tell the model, "What if the biggest shop loses 50% of its value?" The model runs the simulation on its predicted future web to see how much total money would be lost.
- Key Finding: This simulation is most useful when the market is unstable. If the market is calm, the simulation doesn't add much value, but when things are shaky, it provides a crucial safety net.

5. What Are the Limits? (What the Paper Says)

The authors are very honest about what their tool cannot do:

It only sees the "On-Chain" world: It can see the digital tokens moving on the blockchain, but it cannot see money sitting in centralized banks, private deals between big investors, or off-chain reserves.
It's a weekly snapshot: It looks at the market once a week. It might miss very fast, flash crashes that happen in minutes (like a flash flood).
It's a guide, not a judge: The model gives a score or a ranking, but it doesn't replace human judgment. A shop might look "safe" on the graph but have a hidden software bug that the model can't see.

Summary

DeXposure-FM is a specialized AI that maps the hidden debts between digital finance shops. While it's good at predicting normal times, its real value is acting as an early warning system during chaotic times, helping regulators see where a financial shock might spread before it happens. It turns a complex, invisible web of risk into a clear, forward-looking map.

Technical Summary: DeXposure-FM

Problem Statement
Decentralized Finance (DeFi) operates as a dynamic, interconnected ecosystem where credit exposure is implicit and token-mediated rather than defined by bilateral contracts. Protocols hold tokens issued by other protocols as collateral, liquidity, or reserves, creating a complex network of dependencies. When a widely held token experiences a shock (e.g., price collapse, exploit, or governance failure), losses can cascade across protocols and chains. While policy authorities have highlighted these vulnerabilities, quantitative tools for monitoring DeFi systemic risk remain limited. Existing approaches face three primary constraints:

Data Complexity: Exposures are not reported in standardized balance sheets but must be reconstructed from on-chain states, resulting in high-dimensional, non-stationary, and heteroskedastic data.
Model Limitations: Classical time-series tools (e.g., VAR) lack the capacity to represent coupled node-level states, edge-level exposures, and meso-structural evolution. Conversely, standard Graph Neural Networks (GNNs) are typically trained on narrow tasks with limited datasets, lacking the reusability of foundation models.
Measurement Gaps: There is a need for a unified framework that can forecast inter-protocol credit exposures and translate those forecasts into macroprudential indicators (e.g., systemic importance, contagion losses) for regulatory monitoring.

Methodology
The authors introduce DeXposure-FM, a domain-specific time-series graph foundation model (GFM) designed to forecast inter-protocol credit exposures. The approach is built upon the DeXposure dataset, which aggregates 43.7 million weekly inter-protocol exposure observations from 2020–2025, covering 4,300+ protocols, 602 blockchains, and 24,300+ tokens.

Data Formalization: The DeFi ecosystem is modeled as a weighted, directed multigraph where nodes represent protocols and edges represent token-mediated USD exposures. Edge weights are derived from the flow of token values between protocols, calculated by aggregating changes in Total Value Locked (TVL) and token composition.
Architecture:
- Encoder: The model utilizes a GraphPFN (Graph Prior-data Fitted Network) encoder, a transformer-based architecture that fuses graph structure with tabular node features (e.g., log-scaled TVL, token composition, sector encoding). The encoder is initialized with open-source GraphPFN weights and fine-tuned end-to-end.
- Prediction Heads: The encoder feeds into lightweight, multi-horizon heads ( $h \in \{1, 4, 8, 12\}$ $h \in {1, 4, 8, 12}$ weeks) that perform three joint tasks:
  1. Edge Existence: Binary classification of whether an exposure link exists.
  2. Edge Weight: Regression of exposure magnitude changes using residual learning in log-space ( $\log(1+w)$ ).
  3. Node TVL Change: Prediction of protocol-level TVL dynamics.
Training Pipeline: The model is optimized using Adam with a multi-task loss function combining weighted binary cross-entropy (for edge existence) and Smooth L1 (Huber) loss (for edge weights and TVL changes). The training employs time-based expanding window splits to prevent look-ahead bias and handles class imbalance via negative sampling (5:1 ratio).

Key Contributions

Domain-Specific Foundation Model: DeXposure-FM is the first GFM tailored specifically for DeFi credit exposures, adapting the GraphPFN paradigm to dynamic, graph-tabular financial networks.
Forecast-Then-Measure Pipeline: The model does not merely predict graph states; it serves as a forward-looking measurement layer. Predicted graphs ( $\hat{G}_{t+h}$ $\hat{G}_{t + h}$ ) are fed into deterministic functions to generate macroprudential indicators, including:
- Protocol-level Systemic Importance Scores (SIS).
- Sector-to-sector spillover matrices and concentration indices (HHI).
- DebtRank-style scenario stress losses under counterfactual shocks.
Empirical Verification on Strict Hold-outs: The model is evaluated on a strict 2025 hold-out set, demonstrating performance beyond standard persistence baselines in critical regimes.

Empirical Results

Edge Prediction: On the strict 2025 hold-out, DeXposure-FM achieves an AUROC of 0.993–0.995 and AUPRC of 0.967–0.973 for edge existence, significantly outperforming the persistence baseline (AUPRC 0.58–0.64) and a frozen GraphPFN probe.
Magnitude Prediction: The model reduces Root Mean Square Error (RMSE) for edge weights compared to neural baselines (GraphPFN-frozen and ROLAND). While persistence remains competitive for highly stable magnitudes, DeXposure-FM shows lower RMSE across all horizons.
Stress Testing: In predictive contagion stress testing, the model's value is most pronounced in the worst-20% of cases where the persistence baseline fails (high-error tail regimes). In these regimes, DeXposure-FM achieves win rates of 83–100% in reducing Mean Absolute Error (MAE) for system losses, particularly under bridge-sector shocks where network topology evolves rapidly.
Financial Tools: The forecasted graphs successfully support the calculation of forward-looking SIS rankings and spillover concentrations, providing actionable alerts for structural changes and regime shifts (e.g., Terra/Luna and FTX collapses).

Significance and Claims
The paper positions DeXposure-FM as a critical advancement in the toolkit for understanding systemic risk in DeFi. Its primary significance lies in bridging the gap between raw on-chain data and policy-relevant risk metrics.

Complementarity: The authors explicitly state that the model is designed to complement, not replace, qualitative judgment. It provides quantitative summaries (rankings, concentration changes) that must be triangulated with code audits and governance analysis.
Regime Sensitivity: The model is most valuable in structural-change and tail-error regimes where simple carry-forward assumptions (persistence) are least reliable. It serves as a "backstop" for scenario analysis when exposures reallocate non-stationarily.
Reproducibility: The work emphasizes open science, with model checkpoints and code publicly available to foster community development and further research into decentralized financial stability.

The authors acknowledge current limitations, including the reliance on TVL as a proxy (which lacks risk-weighting), the exclusion of off-chain exposures (CEX, OTC), and the weekly temporal resolution which smooths out high-frequency dynamics. Future work aims to address these by integrating higher-frequency data, risk-weighted exposure proxies, and continuous drift monitoring.