Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

DeXposure-FM claims to be the first time-series graph foundation model for DeFi that forecasts inter-protocol credit exposure — edge existence, edge weights, and protocol TVL changes — up to 12 weeks ahead, and turns the forecast graphs int

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 04:47 UTC pith:NIOL72F3

load-bearing objection A genuinely useful open forecasting stack for DeFi exposure networks, undermined by an unvalidated flow proxy that counts pure price moves as flows and by overclaimed validation; core benchmark results are real, but the financial-tool claims need reframing. the 4 major comments →

arxiv 2602.03981 v2 pith:NIOL72F3 submitted 2026-02-03 cs.LG cs.AIecon.EM

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

Aijie Shu , Wenbin Wu , Gbenga Ibikunle , Fengxiang He This is my paper
classification cs.LG cs.AIecon.EM
keywords decentralized financecredit exposuregraph foundation modeltime-series forecastingsystemic riskstress testingnetwork contagiontotal value locked
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a single large pre-trained graph-tabular model, fine-tuned on 43.7 million weekly records of inter-protocol token holdings across 4,300+ protocols, can forecast the future credit-exposure network of decentralized finance — which protocols will be linked, with what weights, and how their total value locked will change — at horizons of 1, 4, 8, and 12 weeks. It argues that these forecasts beat strong baselines and, crucially, beat a persistence baseline in exactly the tail events where the network changes most, with win rates of 83–100% in the worst 20% of stress-test cases. If right, the model gives regulators and risk managers a forward-looking, open-source tool for macroprudential monitoring and scenario-based stress testing of DeFi. The load-bearing premise is that the constructed exposure measure, built from changes in dollar-valued token holdings, reflects true economic exposure rather than price-driven noise.

Core claim

On the paper's own terms, the central discovery is that a multi-task neural model — fine-tuning a pre-trained graph-tabular transformer on weekly DeFi exposure snapshots — jointly predicts edge existence, edge weights, and node-level TVL changes at 1, 4, 8, and 12 weeks, and that measuring network statistics on the predicted graph (the forecast-then-measure pipeline) yields forward-looking systemic-importance scores, sector spillover indices, and simulated stress-test losses. The empirical claim is that the model consistently outperforms state-of-the-art baselines, including a frozen graph foundation model, a temporal graph neural network, and the persistence baseline, on edge-level forecast

What carries the argument

The central object is the weekly weighted directed exposure graph: nodes are protocols with weights equal to total value locked, and a directed edge p→q exists when protocol p holds a token issued by protocol q, with edge weight given by the sum over tokens of inferred positive flows from changes in holdings (Eq. 3). The model is a pre-trained graph-tabular transformer (GraphPFN), defined here as a transformer-based foundation model that fuses graph structure with tabular node features, which produces protocol embeddings; three task heads predict edge existence (binary cross-entropy), edge-weight residuals (Smooth L1 on log(1+w) with residual learning), and node TVL log-changes. The forecast

Load-bearing premise

The exposure measure defined by Eq. (3) counts any opposite-signed change in two protocols' dollar holdings of a shared token as a physical flow, so a pure price move with no transfer at all is recorded as credit exposure.

What would settle it

Build a second version of the DeXposure dataset using only actual on-chain token transfers (quantities, holding prices constant) instead of the ΔS-based value flows of Eq. (3), then rerun the same stress-test benchmark; if the model's tail-of-persistence improvement disappears, its predictive edge is an artifact of the price-driven construction rather than a genuine measure of inter-protocol credit flows.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Regulators get forward-looking watchlists: predicted systemic-importance rankings and spillover indices computed on forecast graphs, not just current snapshots.
  • Scenario-based DeFi stress tests can be run on predicted future networks, giving loss estimates that beat persistence exactly when the network is rewiring.
  • The same forecast-then-measure pipeline extends to any deterministic graph functional, enabling dashboards of density, concentration, and sector connectivity as fragility indicators.
  • Open model weights and code make the measurement stack reproducible and extensible by other researchers and supervisors.
  • Edge-existence forecasts (AUROC 0.995 vs. 0.763 persistence at h=1) show the model learns genuine network rewiring rather than merely repeating last week's topology.

Where Pith is reading between the lines

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

  • Beyond the paper: because exposures are computed from dollar-valuation changes, the model's strong link forecasts may partially reflect co-moving prices rather than actual transfers; a decisive test would rebuild exposures from on-chain token quantities alone.
  • Beyond the paper: with 98.5% week-to-week edge persistence, the model's practical value is in early-warning around structural breaks, not routine nowcasting — a deployment pattern that the paper's own tail-stratified results support.
  • Beyond the paper: the same architecture of a graph-tabular encoder plus hand-built exposure mapping could transfer to other reconstructed financial networks, such as interbank or supply-chain credit links, where the underlying bilateral exposures are similarly implicit.
  • Beyond the paper: the authors' planned move from raw TVL to risk-weighted exposures would likely change the SIS rankings materially, so current rankings should be read as relative, provisional signals rather than authoritative importance scores.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces DeXposure-FM, a time-series graph foundation model obtained by fine-tuning GraphPFN on the DeXposure dataset (43.7M weekly observations, 4,300+ protocols, 602 blockchains). It forecasts edge existence, edge weights, and node TVL changes at horizons h=1,4,8,12 weeks, and then applies deterministic functionals to predicted graphs to construct systemic-importance scores, sector spillover/concentration measures, and DebtRank-style stress-test losses. The model is benchmarked against frozen GraphPFN, ROLAND, and a persistence baseline on a strict 2025 hold-out set. The paper claims consistent outperformance over strong competitors and states that the financial-economics tools are fully supported by empirical verification.

Significance. If the exposure construction is valid, the paper would provide a valuable open and reproducible measurement stack for DeFi systemic risk: code and model weights are public, the evaluation is a genuine out-of-time split, and the edge-existence AUROC (0.995 vs 0.763 at h=1) and edge-weight RMSE (3.388 vs 4.296 at h=1) are notable improvements over persistence. The multi-task design and forecast-then-measure framing are potentially useful for macroprudential monitoring. However, the significance is conditional on the central measurement premise: the edge weights used as 'credit exposure' are inferred from value changes in token holdings, not from verified transfers, and the paper itself concedes that TVL can move mechanically with prices. As submitted, this premise is not validated, and several headline claims are stronger than the reported evidence.

major comments (4)
  1. [§3.3, Eq. (3); §7.2] The exposure edge weights are defined from ΔS, the change in USD value of token holdings, rather than from actual token transfers. Eq. (3) counts a flow whenever one protocol's holding value of a token falls and another's rises. Because ΔS mixes quantity and price, a pure price move with no transfer can be recorded as exposure flow; the paper concedes in §7.2 that TVL 'can also move mechanically with prices even when underlying positions are unchanged.' Any two-sided value change need not be a bilateral flow, since tokens can move through external markets or CEXs. Thus every downstream result — edge-level AUROC/RMSE, DebtRank stress losses, SIS, and spillover forecasts — is for a constructed quantity whose economic interpretation as credit exposure is unvalidated. The authors should either validate Eq. (3) against actual on-chain transfers (or a quantity-based reconstruction) or explicit
  2. [§5.3.2, Table 4, Figure 3] The positive stress-test result rests entirely on a post-hoc worst-20% subset selected by the largest persistence-baseline errors on the same test set. Table 4 shows ΔMAE(all) is negative at every horizon (−0.95, −1.71, −0.17, −0.55), while ΔMAE(worst20%) is positive. Selecting the subset on the test outcome biases the reported gains and win rates; this is not a valid out-of-sample evaluation of a 'tail regime' advantage. The claim that the model is most valuable exactly when persistence breaks down is not established without a pre-specified conditional/tail evaluation protocol or a separate validation set for regime selection. At minimum, the post-hoc nature should be stated prominently and the result treated as exploratory.
  3. [§6.3.2, Figure 4; Abstract; §8] The forecast-then-measure validation is contradicted by the paper's own Figure 4: predicted vs. realized cross-sector spillover HHI has R²=0.000 and TVL concentration HHI has R²=0.001. The text describes these plots as a 'calibration and monitoring check,' but the abstract and conclusion assert that the financial-economics tools are 'fully supported' by empirical verification. An R² of essentially zero for two of the headline aggregate risk metrics does not support the claim. The authors should either remove the 'fully supported' language, report which tools are actually supported, or provide additional validation metrics (e.g., rank correlation, directional accuracy) that genuinely support the monitoring interpretation.
  4. [Table 3; Abstract; §5.2.2] The claim that DeXposure-FM 'consistently outperforms' competitors is too strong when compared against persistence. In Table 3, persistence has lower edge-weight MAE at h=4 (2.372 vs 2.489) and h=12 (2.541 vs 2.648), and lower node ΔTVL MAE at h=4 (0.138 vs 0.140), h=8 (0.213 vs 0.229), and h=12 (0.272 vs 0.286). The consistent gains are on edge existence (AUROC/AUPRC) and edge-weight RMSE. The abstract and conclusion should be qualified to reflect the actual pattern of results.
minor comments (5)
  1. [Eq. (3)] The piecewise definition of F is ambiguous: the two cases are written as independent conditions, so when both ΔS_p<0 and ΔS_q≥0 hold the formula could assign two different values. The intended flow is presumably min(-ΔS_p, ΔS_q) under a joint condition; please correct the notation.
  2. [Table 4 and Figures 2–3] The stress-test comparison reports point estimates and win rates without confidence intervals or sample sizes. Given the post-hoc subsetting, it would be important to show uncertainty, e.g., bootstrap CIs for ΔMAE and tests for the win rates.
  3. [Figure 4] The R² values are useful, but the plots would benefit from regression lines, confidence bands, and per-horizon breakdowns. The current pooling of h=1,4,8,12 in one panel obscures whether longer horizons are systematically worse.
  4. [§6.1.1, Eq. (20)] The normalization of PageRank and log(1+TVL) is described only as 'to comparable scales.' Please specify the exact normalization (e.g., min-max, z-score) used in the reported SIS values, since SIS rankings are a stated policy output.
  5. [General] The paper repeatedly uses 'first' and 'state-of-the-art' in the abstract and introduction. Please soften to match the evidence presented, and consider citing prior DeFi network measurement work beyond [47] to support the novelty claim.

Circularity Check

0 steps flagged

No significant circularity: forecast benchmarks are strict out-of-time comparisons against external baselines; the main self-reference is the DeXposure dataset citation [47], which is not a derivation step.

full rationale

The claimed derivation is not circular. Eq. (3) defines edge weights from changes in USD token holdings (ΔS), and the model is trained to forecast future edge existence/weights and TVL changes. The target is not set equal to the model's output by construction; Eq. (9) uses residual learning over the lagged weight, and the persistence baseline is a separate benchmark. Evaluation is an expanding-window walk-forward split (§4.2.1) versus external competitors (GraphPFN, ROLAND, persistence) with a strict 2025 hold-out, so the central benchmark result does not reduce to a fitted parameter. The one self-referential element is the DeXposure dataset [47], whose authors overlap (Wu, He), and the paper trains and validates exclusively on it; however §3 describes the construction in detail and the dataset is publicly available, so the citation is not doing the deductive work. The measurement-validity concern — Eq. (3) counts pure price moves as flows — is acknowledged in §7.2 ('TVL can also move mechanically with prices even when underlying positions are unchanged'); that is a correctness/validity caveat about the target definition, not a circular derivation. Likewise, the low R² in Figure 4 for some aggregate metrics is an empirical performance issue, not evidence that the prediction is equivalent to its input. Hence no circular step rises above a minor self-citation; score 2.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The model predicts the authors' own DeXposure construction and nothing else: eqs. (2)–(4) define exposures as TVL-dollar changes, M(σ) maps tokens to issuers via heuristic matching, and the contagion simulator (eq. 22) adds hand-set thresholds. All evaluation targets come from the same data family, built by the same author group [47]. Practical thresholds (θ, τ, k, α/β/γ, loss weights) are chosen by hand or inherited. The SIS index is a new ad hoc composite with no distress-based validation. The central forecast claim additionally rests on standard ML assumptions (negative sampling, loss weighting, early stopping), which are reasonable.

free parameters (6)
  • Multi-task loss weights (λ_exist=2.0, λ_weight=0.5, λ_node=20.0) = 2.0 / 0.5 / 20.0
    Hand-calibrated in §4.2.2 to balance gradient contributions; the reported benchmark numbers depend on this choice.
  • Negative sampling ratio (5:1 neg:pos) = 5:1
    §4.2.1; AUROC/AUPRC values depend on the negative-sample distribution, especially under 98.5% edge persistence.
  • SIS weights α=β=γ=1/3 and top-k=5 = 1/3 each; k=5
    §6.1.1 eq. (20); arbitrary equal weights for the 'validated' systemic-importance score; no sensitivity analysis or validation against realized distress.
  • Contagion distress threshold τ = 0.1
    §6.2.1; all stress-test loss numbers scale with this hand-set threshold.
  • Data-construction thresholds θ (node pruning, TF-IDF mapping) = not reported in this paper
    §3.2–3.3; the exposure graphs — hence every training target — inherit thresholds from the DeXposure dataset paper [47].
  • Learning rates (5e-4 heads, 5e-5 backbone) = 5e-4 / 5e-5
    Table 2; standard hyperparameters, minor influence on the central result.
axioms (6)
  • domain assumption The ΔS-based flow decomposition (eq. 3) measures true token flows between protocols
    Pure price changes are counted as exposure flows; §7.2 concedes TVL moves mechanically with prices even when positions are unchanged.
  • domain assumption Token-to-protocol mapping M(σ) is correct
    §3.2 fallback: metadata lookup, manual mapping, TF-IDF similarity, primary-market fallback; a mis-mapping reverses the direction of exposure edges.
  • domain assumption DebtRank-style simulator (eq. 22) approximates real DeFi contagion
    §6.2.1 uses linear proportional loss allocation with threshold 0.1; liquidation cascades, oracle failures, and governance attacks, which the paper cites elsewhere, are not in the simulator.
  • domain assumption Weekly snapshots capture contagion-relevant dynamics
    §6.5.3 concedes flash-loan attacks [35] and rapid liquidation cascades [46] are smoothed out by weekly aggregation.
  • standard math Standard ML background results (BCE, Smooth L1, Adam, early stopping)
    §4.2; textbook results, not load-bearing beyond the evaluation protocol.
  • domain assumption DefiLlama on-chain data is a complete view of DeFi exposure
    §6.5.2: CEX balances, OTC derivatives/lending, and off-chain stablecoin reserves are unobserved, so the model cannot see DeFi–TradFi spillovers.
invented entities (2)
  • SIS (Systemic Importance Score) no independent evidence
    purpose: Protocol-level systemic-importance ranking, eq. (20) — the headline forward-looking watchlist tool.
    A transparent but arbitrary composite (PageRank + top-k tail-exposure share + log TVL with equal weights). The only evidence presented is that predicted mean-SIS tracks realized mean-SIS (R²=0.700), which says nothing about whether the ranking identifies actually systemic protocols; no validation against realized distress.
  • Forecast-then-measure dashboard (predicted watchlists, predicted spillover HHI) no independent evidence
    purpose: Forward-looking macroprudential early-warning signals from predicted graphs.
    An invented pipeline rather than an ontological entity; its headline outputs show R²=0.000 (spillover HHI) and R²=0.001 (TVL HHI) in Fig. 4, so no falsifiable external handle is demonstrated.

pith-pipeline@v1.3.0-alltime-deepseek · 20179 in / 20785 out tokens · 197191 ms · 2026-08-03T04:47:49.758384+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of DeXposure-FM: A Time-series, Graph Foundation Model for Credit Exposures and Stability on Decentralized Financial Networks." pith.science (2026). https://pith.science/paper/NIOL72F3

@misc{pith2026260203981,
  author       = {Pith},
  title        = {Pith review of: DeXposure-FM: A Time-series, Graph Foundation Model for Credit Exposures and Stability on Decentralized Financial Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NIOL72F3}},
  note         = {Machine review of arXiv:2602.03981}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Credit exposure in Decentralized Finance (DeFi) is often implicit and token-mediated, creating a dense web of inter-protocol dependencies. Thus, a shock to one token may result in significant and uncontrolled contagion effects. As the DeFi ecosystem becomes increasingly linked with traditional financial infrastructure through instruments, such as stablecoins, the risk posed by this dynamic demands more powerful quantification tools. We introduce DeXposure-FM, the first time-series, graph foundation model for measuring and forecasting inter-protocol credit exposure on DeFi networks, to the best of our knowledge. Employing a graph-tabular encoder, with pre-trained weight initialization, and multiple task-specific heads, DeXposure-FM is trained on the DeXposure dataset that has 43.7 million data entries, across 4,300+ protocols on 602 blockchains, covering 24,300+ unique tokens. The training is operationalized for credit-exposure forecasting, predicting the joint dynamics of (1) protocol-level flows, and (2) the topology and weights of credit-exposure links. The DeXposure-FM is empirically validated on two machine learning benchmarks; it consistently outperforms the state-of-the-art approaches, including a graph foundation model and temporal graph neural networks. DeXposure-FM further produces financial economics tools that support macroprudential monitoring and scenario-based DeFi stress testing, by enabling protocol-level systemic-importance scores, sector-level spillover and concentration measures via a forecast-then-measure pipeline. Empirical verification fully supports our financial economics tools. The model and code have been publicly available. Model: https://huggingface.co/EVIEHub/DeXposure-FM. Code: https://github.com/EVIEHub/DeXposure-FM.

Figures

Figures reproduced from arXiv: 2602.03981 by Aijie Shu, Fengxiang He, Gbenga Ibikunle, Wenbin Wu.

Figure 1
Figure 1. Figure 1: Multi-layer credit exposure in DeFi. A user stakes ETH with Lido, receiving [PITH_FULL_IMAGE:figures/full_fig_p033_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Predictive contagion stress testing (Task II). We compare system loss (%) under [PITH_FULL_IMAGE:figures/full_fig_p034_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Advantage regimes for predictive stress testing. Bars report [PITH_FULL_IMAGE:figures/full_fig_p035_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Forward-looking risk metric forecasting on predicted exposure graphs. Each [PITH_FULL_IMAGE:figures/full_fig_p036_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Early warning event studies. For each event window, we compare predicted vs. [PITH_FULL_IMAGE:figures/full_fig_p037_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Toward a Risk Assessment Framework for Institutional DeFi: A Nine-Dimension Approach

    cs.DC 2026-05 unverdicted novelty 6.0

    A nine-dimension risk framework for institutional DeFi adds three new dimensions to prior taxonomies and shows that five of twelve 2024-2026 incidents, including the two most systemic, require at least one of the new ...

  2. DeXposure-Claw: An Agentic System for DeFi Risk Supervision

    cs.AI 2026-06 unverdicted novelty 5.0

    Presents DeXposure-Claw, an agentic supervision system that combines DeXposure-FM forecasts, deterministic monitors, and DeXposure-Bench evaluation on five years of DeFi data.

  3. DeXposure-Claw: An Agentic System for DeFi Risk Supervision

    cs.AI 2026-06 unverdicted novelty 5.0

    DeXposure-Claw combines a graph time-series foundation model for forecasting DeFi networks with rule-based monitors and data-health gates to emit regulator-aligned risk tickets, evaluated via a new six-axis benchmark ...

Reference graph

Works this paper leans on

53 extracted references · 1 canonical work pages · cited by 2 Pith papers

  1. [1]

    Systemic risk and stability in financial networks

    Acemoglu, D., Ozdaglar, A., Tahbaz-Salehi, A., 2015. Systemic risk and stability in financial networks. American Economic Re- view 105, 564–608. URL:https://www.aeaweb.org/articles?id=10. 1257/aer.20130456, doi:10.1257/aer.20130456

    doi:10.1257/aer.20130456 2015

  2. [2]

    Chronos: Learning the language of time series

    Ansari, A.F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S.S., Pineda Arango, S., Kapoor, S., Zschieg- ner, J., Maddix, D.C., Wang, H., Mahoney, M.W., Torkkola, K., Wilson, A.G., Bohlke-Schneider, M., Wang, Y., 2024. Chronos: Learning the language of time series. URL:https://arxiv.org/abs/2403.07815, arXiv:2403.07815

    Pith/arXiv arXiv 2024

  3. [3]

    Oracles in decentralized finance: Attack costs, profits and mitigation measures

    Aspembitova, A.T., Bentley, M.A., 2023. Oracles in decentralized finance: Attack costs, profits and mitigation measures. Entropy

    2023

  4. [4]

    Towards Verifiability of Total Value Locked (TVL) in Decentralized Finance

    Auer, R., Faragò, M., Turi, D., 2025. Towards Verifiability of Total Value Locked (TVL) in Decentralized Finance. BIS Working Papers 1268.BankforInternationalSettlements. URL:https://www.bis.org/ publ/work1268.htm

    2025

  5. [5]

    Mapping micro- scopic and systemic risks in tradfi and defi: a literature review

    Aufiero, S., Bartolucci, S., Caccioli, F., Vivo, P., 2025. Mapping micro- scopic and systemic risks in tradfi and defi: a literature review. URL: https://arxiv.org/abs/2508.12007,arXiv:2508.12007

    Pith/arXiv arXiv 2025

  6. [6]

    Pre- dicting systemic risk in financial systems using deep graph learning

    Balmaseda, V., Coronado, M., de Cadenas-Santiago, G., 2023. Pre- dicting systemic risk in financial systems using deep graph learning. Intelligent Systems with Applications 19, 200240. URL:https://www. sciencedirect.com/science/article/pii/S2667305323000650, doi:10.1016/j.iswa.2023.200240. 24

    arXiv 2023

  7. [7]

    Debtrank: Too central to fail? financial networks, the fed and sys- temic risk

    Battiston, S., Puliga, M., Kaushik, R., Tasca, P., Caldarelli, G., 2012. Debtrank: Too central to fail? financial networks, the fed and sys- temic risk. Scientific Reports 2, 541. URL:https://www.nature.com/ articles/srep00541, doi:10.1038/srep00541

    doi:10.1038/srep00541 2012

  8. [8]

    On the use of cross-validation for time series predictor evaluation

    Bergmeir, C., Benítez, J.M., 2012. On the use of cross-validation for time series predictor evaluation. Information Sciences 191, 192–

    2012

  9. [9]

    Measuring defi risk

    Bertomeu, J., Martin, X., Sall, I., 2024. Measuring defi risk. Finance Research Letters 63, 105321. URL:https://www.sciencedirect. com/science/article/pii/S1544612324003519, doi:10.1016/j.frl. 2024.105321

    arXiv 2024

  10. [10]

    On the opportunities and risks of foundation models

    Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., Bryn- jolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J.Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei-Fei, L., Finn, C., Ga...

    Pith/arXiv arXiv 2021

  11. [11]

    Enhancing GDP nowcasts with ChatGPT: a novel application of PMI news releases

    de Bondt, G.J., Sun, Y., 2025. Enhancing GDP nowcasts with ChatGPT: a novel application of PMI news releases. Working Paper Series 3063. European Central Bank. Frankfurt am Main, Germany. URL:https://www.ecb.europa.eu/pub/pdf/scpwps/ecb. wp3063~f88c1b73fc.en.pdf, doi:10.2866/2788332. eCB Working Pa- per Series No. 3063; released 2025-06-30

    doi:10.2866/2788332 2025

  12. [12]

    The anatomy of a large-scale hypertextual web search engine

    Brin, S., Page, L., 1998. The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems 30, 107–

    1998

  13. [13]

    Macroeconomic fore- casting with large language models

    Carriero, A., Pettenuzzo, D., Shekhar, S., 2025. Macroeconomic fore- casting with large language models. URL:https://arxiv.org/abs/ 2407.00890,arXiv:2407.00890

    arXiv 2025

  14. [14]

    A decoder-only founda- tion model for time-series forecasting

    Das, A., Kong, W., Sen, R., Zhou, Y., 2024. A decoder-only founda- tion model for time-series forecasting. URL:https://arxiv.org/abs/ 2310.10688,arXiv:2310.10688. accepted at ICML 2024; model com- monly referenced as TimesFM

    Pith/arXiv arXiv 2024

  15. [15]

    Davis, J., Goadrich, M., 2006. The relationship between precision-recall and roc curves, in: Proceedings of the 23rd International Conference on Machine Learning, Association for Computing Machinery, New York, NY, USA. p. 233–240. URL:doi.org/10.1145/1143844.1143874, doi:10.1145/1143844.1143874

    arXiv 2006

  16. [16]

    DefiLlama: DeFi TVL aggregator and analytics

    DefiLlama, 2026. DefiLlama: DeFi TVL aggregator and analytics. URL: https://defillama.com. data used: 2025; accessed: 2026-02-02

    2026

  17. [17]

    Better to give than to receive: Predictive directional measurement of volatility spillovers

    Diebold, F.X., Yilmaz, K., 2012. Better to give than to receive: Predictive directional measurement of volatility spillovers. Inter- national Journal of Forecasting 28, 57–66. URL:https://www. sciencedirect.com/science/article/pii/S016920701100032X, doi:10.1016/j.ijforecast.2011.02.006. special Section 1: The Predictability of Financial Markets Special Se...

    doi:10.1016/j.ijforecast.2011.02.006 2012

  18. [18]

    Credit risk contagion and systemic risk on networks

    Dolfin, M., Knopoff, D., Limosani, M., Xibilia, M.G., 2019. Credit risk contagion and systemic risk on networks. Mathematics 7. URL:https: //www.mdpi.com/2227-7390/7/8/713, doi:10.3390/math7080713

    work page doi:10.3390/math7080713 2019

  19. [19]

    Systemic risk in financial sys- tems

    Eisenberg, L., Noe, T.H., 2001. Systemic risk in financial sys- tems. Management Science 47, 236–249. URL:https://pubsonline. informs.org/doi/10.1287/mnsc.47.2.236.9835, doi:10.1287/mnsc. 47.2.236.9835

    doi:10.1287/mnsc.47.2.236.9835 2001

  20. [20]

    Turningtabularfoundationmodelsintographfoundationmod- els

    Eremeev, D., Bazhenov, G., Platonov, O., Babenko, A., Prokhorenkova, L., 2025. Turningtabularfoundationmodelsintographfoundationmod- els. URL:https://arxiv.org/abs/2508.20906,arXiv:2508.20906

    Pith/arXiv arXiv 2025

  21. [21]

    Crypto-assets and decen- tralised finance: Report on stablecoins, crypto-investment products and multi-function groups

    European Systemic Risk Board, 2025. Crypto-assets and decen- tralised finance: Report on stablecoins, crypto-investment products and multi-function groups. Technical Report. European Systemic Risk Board. URL:https://www.esrb.europa.eu/pub/pdf/reports/esrb. report202510_cryptoassets.en.pdf. accessed 27 December 2025

    2025

  22. [22]

    An introduction to roc analysis

    Fawcett, T., 2006. An introduction to roc analysis. Pattern Recog- nition Letters 27, 861–874. URL:https://www.sciencedirect.com/ science/article/pii/S016786550500303X, doi:10.1016/j.patrec. 2005.10.010. rOC Analysis in Pattern Recognition

    doi:10.1016/j.patrec 2006

  23. [23]

    Sharpness- aware minimization for efficiently improving generalization, in: Interna- tional Conference on Learning Representations

    Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B., 2021. Sharpness- aware minimization for efficiently improving generalization, in: Interna- tional Conference on Learning Representations

    2021

  24. [24]

    Contagion in financial networks

    Glasserman, P., Young, H.P., 2016. Contagion in financial networks. Journal of Economic Literature 54, 779–831. URL:https://www. aeaweb.org/articles?id=10.1257/jel.20151228, doi:10.1257/jel. 20151228

    doi:10.1257/jel.20151228 2016

  25. [25]

    URL:https://www.mdpi.com/1099-4300/25/1/60, doi:10.3390/ e25010060

  26. [26]

    Computing systemic risk measures with graph neural networks

    Gonon, L., Meyer-Brandis, T., Weber, N., 2025. Computing systemic risk measures with graph neural networks. URL:https://arxiv.org/ abs/2410.07222,arXiv:2410.07222. 27

    arXiv 2025

  27. [27]

    Adam: A method for stochastic optimiza- tion, in: Proceedings of the 3rd International Conference on Learning Representations (ICLR)

    Kingma, D.P., Ba, J., 2015. Adam: A method for stochastic optimiza- tion, in: Proceedings of the 3rd International Conference on Learning Representations (ICLR). URL:https://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2015

  28. [28]

    Sok: Decentralized finance (defi) – fundamentals, taxonomy and risks

    Gogol, K., Killer, C., Schlosser, M., Bocek, T., Stiller, B., Tes- sone, C., 2024. Sok: Decentralized finance (defi) – fundamentals, taxonomy and risks. URL:https://arxiv.org/abs/2404.11281, arXiv:2404.11281

    Pith/arXiv arXiv 2024

  29. [29]

    Will the Real Stable- coin Please Stand Up? BIS Papers 141

    Kosse, A., Glowka, M., Mattei, I., Rice, T., 2023. Will the Real Stable- coin Please Stand Up? BIS Papers 141. Bank for International Settle- ments. URL:https://www.bis.org/publ/bppdf/bispap141.pdf. bIS Papers No. 141; accessed 2025-12-27

    2023

  30. [30]

    Large language models: a primer for economists

    Kwon, B., Park, T., Perez-Cruz, F., Rungcharoenkitkul, P., 2024. Large language models: a primer for economists. BIS Quarterly Review URL: https://www.bis.org/publ/qtrpdf/r_qt2412b.htm

    2024

  31. [31]

    Semi-supervised classification with graph convolutional networks

    Kipf, T.N., Welling, M., 2017. Semi-supervised classification with graph convolutional networks. URL:https://arxiv.org/abs/1609.02907, arXiv:1609.02907

    Pith/arXiv arXiv 2017

  32. [32]

    The rise of diffusion models in time-series forecasting

    Meijer, C., Chen, L.Y., 2024. The rise of diffusion models in time-series forecasting. URL:https://arxiv.org/abs/2401.03006. arXiv:2401.03006

    Pith/arXiv arXiv 2024

  33. [33]

    On the difficulty of train- ing recurrent neural networks, in: Dasgupta, S., McAllester, D

    Pascanu, R., Mikolov, T., Bengio, Y., 2013. On the difficulty of train- ing recurrent neural networks, in: Dasgupta, S., McAllester, D. (Eds.), Proceedings of the 30th International Conference on Machine Learn- ing, PMLR, Atlanta, Georgia, USA. pp. 1310–1318. URL:https: //proceedings.mlr.press/v28/pascanu13.html

    2013

  34. [34]

    Diffusion-LM improves controllable text generation, in: Advances in Neural Information Processing Systems, pp

    Li, X.L., Thickstun, J., Gulrajani, I., Liang, P., Hashimoto, T.B., 2022. Diffusion-LM improves controllable text generation, in: Advances in Neural Information Processing Systems, pp. 4328–4343. URL:https: //arxiv.org/abs/2205.14217

    Pith/arXiv arXiv 2022

  35. [35]

    Qin, K., Zhou, L., Livshits, B., Gervais, A., 2021. Attacking the defi ecosystem with flash loans for fun and profit, in: Financial Cryptogra- phy and Data Security: 25th International Conference, FC 2021, Vir- tual Event, March 1–5, 2021, Revised Selected Papers, Part I, Springer- 28 Verlag, Berlin, Heidelberg. p. 3–32. URL:https://doi.org/10.1007/ 978-3...

    doi:10.1007/978-3-662-64322-8_1 2021

  36. [36]

    Lag-llama: To- wards foundation models for probabilistic time series forecasting

    Rasul, K., Ashok, A., Williams, A.R., Ghonia, H., Bhagwatkar, R., Khorasani, A., Bayazi, M.J.D., Adamopoulos, G., Riachi, R., Has- sen, N., Biloš, M., Garg, S., Schneider, A., Chapados, N., Drouin, A., Zantedeschi, V., Nevmyvaka, Y., Rish, I., 2024. Lag-llama: To- wards foundation models for probabilistic time series forecasting. URL: https://arxiv.org/ab...

    Pith/arXiv arXiv 2024

  37. [37]

    Early Stopping - But When?

    Prechelt, L., 1998. Early Stopping - But When?. Springer Berlin Heidel- berg, Berlin, Heidelberg. pp. 55–69. URL:0.1007/3-540-49430-8_3, doi:10.1007/3-540-49430-8_3

    doi:10.1007/3-540-49430-8_3 1998

  38. [38]

    Temporal graph networks for deep learn- ing on dynamic graphs

    Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., Bronstein, M., 2020. Temporal graph networks for deep learn- ing on dynamic graphs. URL:https://arxiv.org/abs/2006.10637, arXiv:2006.10637

    Pith/arXiv arXiv 2020

  39. [39]

    Decentralized finance: On blockchain- and smart contract-based financial markets

    Schär, F., 2021. Decentralized finance: On blockchain- and smart contract-based financial markets. Federal Re- serve Bank of St. Louis Review 103, 153–174. URL:https: //www.stlouisfed.org/publications/review/2021/02/05/ decentralized-finance-on-blockchain-and-smart-contract-based-financial-markets, doi:10.20955/r.103.153-74

    doi:10.20955/r.103.153-74 2021

  40. [40]

    The herfindahl-hirschman index

    Rhoades, S.A., 1993. The herfindahl-hirschman index. Federal Re- serve Bulletin 79, 188–189. URL:https://ideas.repec.org/a/fip/ fedgrb/y1993imarp188-189nv.79no.3.html

    1993

  41. [41]

    Macroeconomics and reality

    Sims, C.A., 1980. Macroeconomics and reality. Econometrica 48, 1–48. URL:http://www.jstor.org/stable/1912017

    arXiv 1980

  42. [42]

    Forecasting using principal com- ponents from a large number of predictors

    Stock, J.H., Watson, M.W., 2002. Forecasting using principal com- ponents from a large number of predictors. Journal of the American Statistical Association 97, 1167–1179. URL:https://doi.org/ 10.1198/016214502388618960, doi:10.1198/016214502388618960, arXiv:10.1198/016214502388618960. 29

    doi:10.1198/016214502388618960 2002

  43. [43]

    Text mining: Use of tf-idf to examine the relevance of words to documents

    Shahzad Qaiser, R.A., 2018. Text mining: Use of tf-idf to examine the relevance of words to documents. Interna- tional Journal of Computer Applications 181, 25–29. URL: https://ijcaonline.org/archives/volume181/number1/ 29681-2018917395/, doi:10.5120/ijca2018917395

    doi:10.5120/ijca2018917395 2018

  44. [44]

    Graph attention networks

    Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., Bengio, Y., 2018. Graph attention networks. URL:https://arxiv.org/abs/ 1710.10903,arXiv:1710.10903

    Pith/arXiv arXiv 2018

  45. [45]

    Ftx’s downfall and binance’s consolidation: The fragility of centralised digital fi- nance

    Vidal-Tomás, D., Briola, A., Aste, T., 2023. Ftx’s downfall and binance’s consolidation: The fragility of centralised digital fi- nance. Physica A: Statistical Mechanics and its Applications 625, 129044. URL:https://www.sciencedirect.com/science/article/ pii/S037843712300599X, doi:10.1016/j.physa.2023.129044

    arXiv 2023

  46. [46]

    Attention is all you need, in: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I., 2017. Attention is all you need, in: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc. URL:https://proceedings.neurips.cc/pa...

    2017

  47. [47]

    Dexposure: A dataset and benchmarks for inter- protocol credit exposure in decentralized financial networks

    Wu, W., Qian, K., Lui, A., Jack, C., Wu, Y., McBurney, P., He, F., Zhang, B., 2025. Dexposure: A dataset and benchmarks for inter- protocol credit exposure in decentralized financial networks. URL: https://arxiv.org/abs/2511.22314,arXiv:2511.22314

    arXiv 2025

  48. [48]

    Do transformers really perform badly for graph representation?, in: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., Liu, T.Y., 2021. Do transformers really perform badly for graph representation?, in: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (Eds.), Advances in Neural Informa- tion Processing Systems, Curran Associates, Inc.. pp. 28877–28888. URL:https://proceedings.neurips.cc/p...

    2021

  49. [49]

    Toxic liquidation spirals

    Warmuz, J., Chaudhary, A., Pinna, D., 2023. Toxic liquidation spirals. URL:https://arxiv.org/abs/2212.07306, doi:10.48550/ arXiv.2212.07306,arXiv:2212.07306

    Pith/arXiv arXiv doi:10.48550/arxiv.2212.07306 2023

  50. [50]

    Overall” is the full test set; “Worst 20%

    Zhang, S., Wang, Z., Zheng, J., Cartlidge, J., 2026. Systemic risk in DeFi: A network-based fragility analysis of TVL dynamics. arXiv preprint arXiv:2601.08540 URL:https://arxiv.org/abs/2601. 08540. 31 User Lido wstETH Wrapper Pendle Lido Assets Staked ETH $100 Liabilities stETH $100 wstETH W rapper Assets stETH $100 Liabilities wstETH $100 Pendle Assets ...

    arXiv 2026

  51. [52]

    ROLAND: Graph learning framework for dynamic graphs, in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp

    You, J., Du, T., Leskovec, J., 2022. ROLAND: Graph learning framework for dynamic graphs, in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 2358–2366. URL:https://arxiv.org/abs/2208.07239, doi:10.1145/ 3534678.3539300. 30

    Pith/arXiv arXiv 2022

  52. [117]

    proceed- ings of the Seventh International World Wide Web Conference

    URL:https://www.sciencedirect.com/science/article/pii/ S016975529800110X, doi:10.1016/S0169-7552(98)00110-X. proceed- ings of the Seventh International World Wide Web Conference

    doi:10.1016/s0169-7552(98)00110-x

  53. [213]

    data Mining for Software Trustworthiness

    URL:https://www.sciencedirect.com/science/article/pii/ S0020025511006773, doi:10.1016/j.ins.2011.12.028. data Mining for Software Trustworthiness

    doi:10.1016/j.ins.2011.12.028 2011