🤖 machine learning

Predicting Channel Closures in the Lightning Network with Machine Learning

This paper demonstrates that predicting Lightning Network channel closures is most effectively achieved using simple machine learning models based on temporal and behavioral node features, as network topology offers no additional predictive value and the inherent privacy of the network fundamentally limits accuracy when relying solely on public gossip data.

Original authors: Simone Antonelli, Vincent Davis, Harrison Rush, Anthony Potdevin, Jesse Shrader, Vikash Singh, Emanuele Rossi

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Simone Antonelli, Vincent Davis, Harrison Rush, Anthony Potdevin, Jesse Shrader, Vikash Singh, Emanuele Rossi

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

Imagine the Lightning Network as a massive, bustling city of digital highways. In this city, people (nodes) build temporary bridges (channels) between each other to trade Bitcoin quickly and cheaply without waiting for the slow, expensive main road (the Bitcoin blockchain).

Most of the time, these bridges are stable. But sometimes, a bridge collapses. It can happen in two ways:

Mutual Closure: The two people on the bridge agree to take it down together, pack up their money, and walk away happily.
Forced Closure: One person gets angry, unresponsive, or suspicious, so they unilaterally blow up the bridge. This is bad news: it freezes their money for weeks and causes traffic jams for everyone else.

The authors of this paper asked a simple question: Can we predict when a bridge is about to collapse, and how, just by watching the public news feeds?

The "News Feed" (Gossip Data)

In this digital city, there is a public bulletin board called the "gossip protocol." It broadcasts updates like: "Bridge A opened," "Bridge B updated its toll fees," or "Bridge C is closed."

The researchers collected two years of these public updates. They treated this data like a temporal graph—a movie of the city evolving over time, rather than just a single photograph. Their goal was to build a machine learning "detective" that watches this movie and predicts the fate of every open bridge.

The Detective's Toolkit

The team tried out many different types of "detectives" (Machine Learning models):

The Simple Observer (MLP): A basic model that just looks at the facts of a specific bridge and the history of the two people using it.
The Neighborhood Watch (Graph Neural Networks): Complex models that try to learn from the entire city map, thinking that if your neighbors are in trouble, you might be too.
The Tree-Planter (Gradient Boosted Trees): Models that make decisions by asking a series of "yes/no" questions, like a flowchart.

The Big Surprise: The Neighborhood Doesn't Matter

The most interesting finding was that the complex neighborhood models failed to beat the simple observer.

Think of it like this: If you want to know if a specific couple is going to break up, looking at their friends' relationships (the graph topology) doesn't help much. What matters is their own behavior.

Did they just stop talking to each other recently?
Have they had breakups with other partners before?
How long has this relationship lasted?

The paper found that the simplest model (the MLP) was the best detective. It only needed to know:

Recency: When was the last time this person sent a message?
History: How many times has this person forced a bridge to close in the past?
Age: How old is the bridge?

The "map" of the whole network (who is connected to whom) turned out to be useless for this specific prediction. The signals were entirely local to the two people on the bridge.

The "Privacy Wall"

Why couldn't the models predict perfectly? Why was the best score only "moderate"?

The authors explain that the Lightning Network is designed with privacy as a core feature. The public bulletin board (gossip) hides the most important clues:

The Bank Balance: How much money is actually in the bridge?
The Traffic: How many payments are failing?
The Uptime: Is the person's computer turned off?

These are the real reasons a bridge collapses. But because the network keeps these secrets to protect users, the "detective" is forced to guess based on shadows and silhouettes. It's like trying to predict a car crash by only watching the road signs, without being able to see the driver's speed or if they are asleep.

The Takeaway

The paper concludes that while we can build a decent predictor using public data, we hit a "glass ceiling." The best we can do is look at who is active and who has a history of drama.

The researchers released their data and code to the public, hoping others can try to crack the code further. They suggest that to get better predictions, we might need to peek behind the privacy curtain (using private data that only the bridge owners see), but for now, the public news feed only tells us so much.

In short: To predict if a digital bridge will break, don't look at the whole city map. Just look at the two people on the bridge, see if they've been quiet lately, and check if they've been troublemakers before. But remember, the most important clues are hidden in a locked box.

Technical Summary: Predicting Channel Closures in the Lightning Network with Machine Learning

1. Problem Statement

The Lightning Network (LN) is a second-layer protocol for Bitcoin that enables off-chain transactions via payment channels. These channels can terminate via mutual closure (cooperative, immediate fund release) or forced closure (unilateral, resulting in funds being locked for a timelock period of days to weeks). Forced closures are detrimental to network efficiency, consuming on-chain fees and freezing liquidity.

The core problem addressed is predicting the future state of an open channel based solely on publicly available gossip data. The task is formalized as a temporal link classification problem: given the state of the LN graph at time $t$ , predict whether each currently open channel will remain open, close mutually, or be force-closed within a future window $\Delta t$ (set to 180 days in this study).

The challenge lies in the inherent privacy of the LN. Critical factors influencing closure decisions—such as channel balances, payment failure rates, and node uptime—are not broadcast via the gossip protocol. The study investigates whether observable temporal patterns and behavioral history can compensate for this missing information.

2. Methodology

Dataset Construction

The authors constructed a dataset spanning June 9, 2022, to October 14, 2024, derived from daily snapshots of LN gossip messages.

Scale: 693,277 directed events (openings/closures) across 874 timestamps involving 36,170 unique nodes.
Preprocessing:
- Warm-start: The initial snapshot (June 2022) was used to initialize graph state but excluded from training to prevent data leakage from accumulated history.
- Multigraph handling: Parallel channels between node pairs were removed to create a simple graph, eliminating ambiguity in feature selection.
- Splitting: Data was split chronologically (70% train, 15% validation, 15% test) to prevent temporal leakage.
Labels: Open edges at time $t$ were labeled based on their status at $t + 180$ days: Open, Mutual, or Forced. The dataset is highly imbalanced, with ~83% of open edges remaining open at prediction time.

Feature Engineering

Features were extracted directly from gossip messages at both the edge (channel) and node (endpoint) levels:

Edge Features: On-chain timestamps, locked capacity, funding block fee rates, fee policies (base/proportional), HTLC bounds, time locks, and disabled flags.
Node Features: For each endpoint, the model tracks running counts of events (openings, mutual closures, forced closures) and recency metrics (time since last update).
Temporal Encodings: Channel age and the recency of the last interaction for both source and destination nodes.

Models Evaluated

The study benchmarked a range of approaches to determine the value of graph structure versus local features:

Baselines: Random (uniform/stratified) and Majority class predictors.
MLP (Multi-Layer Perceptron): A non-graph model classifying edges independently using concatenated edge, node, and temporal features.
Gradient-Boosted Trees: XGBoost and LightGBM using the same input vectors as the MLP.
Graph Neural Networks (GNNs):
- Static GNN (GraphSAGE): Operates on the snapshot of open edges without temporal dynamics.
- Temporal GNN (TGN): Incorporates message passing over the dynamic graph and learns node embeddings via attention.
Spectral Encodings: Augmented MLPs using eigenvectors of the normalized Laplacian to capture structural roles without iterative aggregation.

Evaluation Pipeline

The evaluation was adapted from the Temporal Graph Network (TGN) framework. Unlike standard link prediction which samples edges, this pipeline predicts the status of all currently open edges at each timestamp. Performance was measured using Macro-average F1-score to address class imbalance.

3. Key Results

Performance Comparison

MLP Superiority: The simple MLP achieved the highest performance with a Macro F1 of 0.38.
Graph Models Underperform: The TGN (0.36), Static GNN (0.35), and Spectral MLP (0.35) all performed worse than or equal to the non-graph MLP.
Baseline Comparison: All learned models outperformed the stratified random baseline (0.32), but the margin was modest. The MLP significantly outperformed the majority baseline (0.30) only when class weighting was applied.

Feature Importance and Ablation

Dominant Signals: Feature ablation and SHAP analysis revealed that temporal and behavioral signals are the primary drivers of prediction. Specifically:
- Recency of endpoint activity ( $src\_recency$ , $dst\_recency$ ).
- Per-node history of past closures ( $count\_forced$ , $count\_mutual$ ).
- Channel age.
Insignificant Signals: Static channel metadata (fees, capacity, timelocks) and graph topology provided negligible additional predictive power.
Graph Aggregation: Adding GNN message passing or spectral encodings to the MLP did not improve performance; in some cases, it degraded it. This suggests that neighborhood structure does not reveal information about a specific channel's closure risk that isn't already captured by the endpoints' individual histories.

Confusion Matrix Analysis

The model successfully identified channels that would remain open (high recall for the Open class). However, it struggled to distinguish between Mutual and Forced closures, often confusing them with each other or with the Open class. Performance for closure types dropped significantly for older channels, where closures are rare events.

4. Key Contributions

Formalization: Defined channel closure prediction as a temporal link classification task over an evolving graph.
Dataset Release: Publicly released a two-year dataset of LN gossip snapshots with labeled closure outcomes.
Empirical Findings: Demonstrated that for this specific task, local temporal and behavioral features outperform graph-based representations.
Privacy Analysis: Provided evidence that the fundamental limit of predictability is the privacy design of the LN, where critical closure triggers (balances, failures) are invisible to the gossip protocol.

5. Significance and Claims

The paper claims that the moderate predictive performance (Macro F1 ~0.38) is not a failure of machine learning architectures but a reflection of a fundamental information gap.

Privacy Limitation: The LN is designed to hide channel balances, payment flows, and node uptime. Since these are the primary triggers for forced closures, their absence from gossip data makes high-accuracy prediction impossible using public data alone.
Behavioral Proxies: The model relies on the closest observable proxies: how recently a node has been active and its historical tendency to close channels.
Graph Topology Irrelevance: Unlike social networks where a node's neighbors strongly predict its behavior, the LN's topology offers little insight into a specific channel's fate because closure decisions are bilateral and private.
Practical Utility: While perfect prediction is unattainable with public data, temporal modeling remains a valuable tool for node operators to optimize routing reliability and fee earnings, provided they understand the inherent limitations.

The authors conclude that significantly improving prediction beyond current levels would likely require access to private, node-local data (e.g., local balance histories or payment logs), which contradicts the LN's privacy-preserving architecture.