ln-gossip-benchmark

Code and benchmarking pipeline for the paper Predicting Channel Closures in the Lightning Network with Machine Learning (Antonelli et al., 2026, arXiv:2605.12759).

The task is temporal link classification: at each timestamp $t$, for every currently-open channel, predict whether it will remain open, close cooperatively (mutual), or be force-closed within the next $\Delta_t$ days, using only public LN gossip data up to $t$.

Installation

uv sync

After uv sync, install PyG's C-extension wheels (replace +cpu with +cu121 etc. on GPU):

uv pip install --no-deps \
  torch-scatter torch-cluster torch-sparse torch-spline-conv \
  --find-links https://data.pyg.org/whl/torch-2.5.0+cpu.html

Dataset

The gossip dataset (LN events, 2022-06-09 → 2024-10-14, ~293MB CSV) is hosted publicly on HuggingFace Datasets at Amboss-Tech/ln-gossip and the pipeline downloads it on first run — no authentication required.

The processed tensors are cached at ~/.cache/ln-gossip-benchmark/tgbl-ln_processed.pkl so subsequent runs skip the parse step. Override the cache location with LN_GOSSIP_CACHE_DIR.

Quick start

uv run python scripts/train.py \
  model=mlp \
  data.warm_start=true \
  model.seed=0

Reproducing the paper

All paper results use the warm-start configuration, $\Delta_t = 180$ days (180-day forward closure window) and 3 seeds (0, 1, 2). For a single seed, swap model.seed=0,1,2 with model.seed=0.

The MLP and TGN headline configs are baked into conf/model/mlp.yaml and conf/model/temporal_gnn.yaml respectively (architecture flags are not specified on the command line — that proved error-prone earlier).

Main results (Table I)

# Random baselines
uv run python scripts/train.py \
  model=random model.strategy=uniform    data.warm_start=true model.seed=0,1,2 -m
uv run python scripts/train.py \
  model=random model.strategy=majority   data.warm_start=true model.seed=0,1,2 -m
uv run python scripts/train.py \
  model=random model.strategy=stratified data.warm_start=true model.seed=0,1,2 -m

# Static (snapshot) GNN
uv run python scripts/train.py \
  model=snapshot_gnn data.warm_start=true model.seed=0,1,2 -m

# TGN
uv run python scripts/train.py \
  model=temporal_gnn data.warm_start=true model.seed=0,1,2 -m

# MLP + Spectral
uv run python scripts/train.py \
  model=edge_predictor_spectral data.warm_start=true model.seed=0,1,2 -m

# Tabular baselines
uv run python scripts/train.py \
  model=xgboost  data.warm_start=true model.seed=0,1,2 -m
uv run python scripts/train.py \
  model=lightgbm data.warm_start=true model.seed=0,1,2 -m

# MLP (best, paper headline — 0.38)
uv run python scripts/train.py \
  model=mlp data.warm_start=true model.seed=0,1,2 -m

Feature ablation (Table II)

# Time only
uv run python scripts/train.py \
  model=mlp model.use_edge_features=false model.use_node_history=false \
  data.warm_start=true model.seed=0,1,2 -m
# + Edge feat.
uv run python scripts/train.py \
  model=mlp model.use_node_history=false \
  data.warm_start=true model.seed=0,1,2 -m
# + Node feat. (== best MLP)
uv run python scripts/train.py \
  model=mlp data.warm_start=true model.seed=0,1,2 -m
# Edge + Time + GNN
uv run python scripts/train.py \
  model=temporal_gnn model.use_node_history=false \
  data.warm_start=true model.seed=0,1,2 -m
# All + GNN (== best TGN)
uv run python scripts/train.py \
  model=temporal_gnn data.warm_start=true model.seed=0,1,2 -m

Layer ablation (Figure 5)

Vary model.mlp_num_hidden_layers ∈ {0, 1, 2} for model=mlp and model=temporal_gnn.

Prediction-window ablation (Figure 6)

for dt in 30 90 180 365; do
  uv run python scripts/train.py \
    model=mlp \
    model.ground_truth_oracle.time_window_days=$dt \
    data.warm_start=true model.seed=0,1,2 -m
done

Class-imbalance ablation (Table III)

Vary model.criterion.weight.data (e.g. [1,5,5], [1,1,1], [1,3,6], ...) and/or downsampling ratio r per the paper.

SHAP feature importance (Figure 4 right)

uv run python scripts/train.py \
  model=mlp data.warm_start=true model.seed=0 \
  +model.save_model_path=checkpoints/best_mlp.pt

uv run python scripts/feature_importance.py \
  +checkpoint_path=checkpoints/best_mlp.pt

Logging

Weights & Biases is installed but disabled by default (we set WANDB_MODE=disabled in the example commands). To send metrics to wandb instead:

wandb login
unset WANDB_MODE     # or `export WANDB_MODE=online`

License

MIT.

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}
}