metadata
license: cc-by-4.0
language:
- en
pretty_name: Lightning Network Gossip Channel Closure Dataset
size_categories:
- 100K<n<1M
tags:
- lightning-network
- bitcoin
- temporal-graph
- link-classification
configs:
- config_name: default
data_files:
- split: full
path: tgbl-ln_edgelist.csv
Lightning Network Gossip — Channel Closure Dataset
Daily snapshots of the Lightning Network (LN) collected from gossip messages between 2022-06-09 and 2024-10-14. Used in the paper Predicting Channel Closures in the Lightning Network with Machine Learning (Antonelli et al., 2026; arXiv:2605.12759) and consumed by the ln-channel-closure-prediction codebase.
Schema (one row per gossip event)
| Column | Description |
|---|---|
chan_id |
Lightning channel identifier (block:tx_idx:vout). |
transaction_id, transaction_vout |
Funding transaction reference. |
last_update |
Last gossip update timestamp (ms). |
capacity |
Channel capacity (sat). |
src, dst |
Endpoint public keys (hex). |
src_*, dst_* |
Per-direction routing policy: time-lock delta, htlc bounds, fee base/rate, disabled flag, max-htlc, last-update timestamp. |
channel_status |
OPEN / CLOSED. |
closing_info |
OPEN / MUTUAL / FORCED / LOCAL_CLOSED / PENALTY_CLOSED. |
src_alias, dst_alias, src_implementation, dst_implementation |
Endpoint metadata. |
gossip_ts |
Event observation timestamp (ms). |
ts, height, block_avg_fee_rate |
On-chain funding metadata. |
Statistics
- Events: 693 277.
- Unique nodes: 36 170.
- Class distribution at prediction time (over open edges): open ≈83 %, mutual ≈9 %, forced ≈8 %.
Notes on the initial snapshot
The first day (2022-06-09) carries the entire pre-existing LN state as a
single batch of artificial "openings" — about 38 % of the rows. The
benchmark's warm_start mode initialises from these events without using
them for training/eval.
License
Dataset is released under CC-BY-4.0. The underlying gossip messages are public by design.
Citation
@misc{antonelli2026predicting,
title={Predicting Channel Closures in the Lightning Network with Machine Learning},
author={Simone Antonelli and Vincent Davis and Harrison Rush and Anthony Potdevin and Jesse Shrader and Vikash Singh and Emanuele Rossi},
year={2026},
eprint={2605.12759},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2605.12759}
}