DeXposure-FM

DeXposure-FM forecasts weekly credit-exposure graphs between DeFi protocols. The model predicts edge existence, edge exposure weight, and node TVL change.

Paper: arXiv:2602.03981v2

Code: EVIEHub/DeXposure-FM

Latest four-horizon release

Status: reproduction candidate. An independent full run of the new evaluation-only entrypoint has not yet completed. Saved September training-run metrics must not be confused with a new acceptance-run result.

New September checkpoints are published separately under reconstructions/20260905. Use the versioned reproduction guide to download, verify, evaluate or retrain them. The September metrics are reported separately from paper metrics. Historical root weights below remain unchanged.

Historical v2 artifact status

The v2 manuscript was revised on 30 June 2026. metrics-v2-paper.json records the manuscript values. The older metrics-h1.json belongs to a different h=1 run. The hosted checkpoints are not a complete archive of all four original v2 fine-tuned weights.

Hosted file	Documented horizon association	v2 evidence status
dexposure-fm-h1.pt	1	Earlier run. Its recorded AUPRC 0.978 and AUROC 0.996 do not equal the v2 h=1 row.
dexposure-fm-h4.pt	4	Accompanying metrics agree with the v2 h=4 row; this does not establish a verified weight-to-result mapping.
dexposure-fm-h8-h12.pt	Unresolved	Contains one model state; the filename does not prove which horizon was saved.
graphpfn-frozen-all-horizons.pt	1, 4, 8, 12	Frozen baseline checkpoint.

The original four-horizon weight-to-result mapping is incomplete. The old training script overwrote best_model.pt between horizons. See the historical audit for the evidence and remaining ambiguity. The new September release uses four explicitly identified files and does not replace the old weights.

v2 reported Task I results

These are the values printed in arXiv v2. They describe the paper experiment, not every checkpoint currently present in this repository.

Model	h	AUROC	AUPRC	Edge MAE	Edge RMSE	Node MAE	Node RMSE
DeXposure-FM	1	.995	.972	2.465	3.388	.056	.400
DeXposure-FM	4	.995	.973	2.489	3.424	.140	.680
DeXposure-FM	8	.994	.967	2.554	3.509	.229	.890
DeXposure-FM	12	.993	.967	2.648	3.606	.286	1.046
Frozen GraphPFN	1	.988	.938	3.260	4.383	.059	.401
Frozen GraphPFN	4	.988	.940	3.189	4.049	.142	.682
Frozen GraphPFN	8	.987	.938	3.169	4.064	.215	.896
Frozen GraphPFN	12	.986	.936	3.136	4.130	.324	1.056

ROLAND and Persistence baseline rows are recorded in docs/V2_REPRODUCIBILITY.md.

Download

huggingface-cli download EVIEHub/DeXposure-FM --local-dir checkpoints

The repository is gated. Accept its access conditions and authenticate before downloading.

Reproduce the v2 experiment

git clone https://github.com/EVIEHub/DeXposure-FM.git
cd DeXposure-FM
uv sync --frozen
uv run python bin/download_dataset.py
bash run_v2_reproduction.sh

The fixed experiment uses horizons 1, 4, 8, and 12, holdout start 2025-01-01, validation window 24 weeks, 20 epochs, and seed 42.

A run matches the paper only if every metric rounds to the same displayed value and all four reconstructed checkpoints are present and nonempty.

Intended use

• Research on weekly DeFi exposure-network forecasting.
• Reproduction of the paper's graph forecasting and macroprudential experiments.
• Offline risk monitoring and stress-test research with human review.

Out of scope

• Real-time trading or automated financial decisions.
• Price prediction.
• Claims about non-DeFi networks without a separate validation study.

Limitations

• Weekly snapshots can miss intra-week changes.
• On-chain data can be delayed, incomplete, or missing protocol context.
• Forecast performance can change under a new market regime.
• Current hosted fine-tuned weights do not provide all four original v2 checkpoints.

License

Apache-2.0. See the repository NOTICE and LICENSES/ files for third-party components.