The Ledger
Solana study finds no clear early advantage for graph-based rug detection
A preprint compares graph-based and other models for detecting potential Solana rug pulls shortly after token creation. The researchers found no statistically supported early advantage for graph models, but said the tests could not rule out smaller gains.

Graph-based models showed no statistically supported advantage over a non-graph model for earlier detection of potential Solana rug pulls in a new research preprint. The paper, submitted on November 9, compares predictions made from the first minutes of a token’s activity. The authors said the result was inconclusive rather than evidence that the model types perform equally.
Source arxiv.org
Testing short histories
The researchers started with 2,700 tokens sampled from the SolRPDS dataset, collecting one-minute market histories from Birdeye and token transfers from Solana RPC. Most sampled tokens were created in 2024, while the final modeling cohort spanned September 2022 to October 2024.
The team examined observation windows of about 12.5, 35.5 and 93 minutes after token creation, fixed before data collection. For each window, the models used only information available at or before the prediction cutoff.
The comparison included graph neural networks, which use relationships between transfer endpoints, and non-graph models given the same fourteen token-level and graph-summary features. Researchers also kept a panel of tokens observed across all windows to compare observation times without changing the token population.
How tokens were labeled
The study used a nominal 24-hour horizon after creation to determine whether a token met its operational definition of a rug pull. To receive a positive label, a token had to meet conditions for a liquidity drop, a price collapse and persistent liquidity depletion.
The rule required liquidity to fall more than 80% below its earlier peak. The lowest price after the drop also had to be more than 90% below the highest price before it. The liquidity depletion also had to persist under the study’s criteria during the 30 minutes after the drop.
A positive label is an operational outcome, not proof of intent. The definition does not capture slower rug pulls or schemes without a liquidity drain, including freeze-authority abuse.
Mixed model results
The authors found differences in the size, density, reciprocity and receiver-side centralization of transfer graphs for rug-pull and non-rug tokens within the first 12.5 minutes. In their main analysis, however, none of the graph neural networks at the earliest cutoff was significantly better than XGBoost on the same tokens at any of the three cutoffs.
On that fixed panel, XGBoost’s test average precision fell from 0.464 at the earliest cutoff to 0.393 at the latest, the paper reported. In a separate exploratory comparison using models selected for each setting, the chosen graph model had higher test average precision in 5 of 6 settings. But none of the differences remained significant after correction for multiple testing.
Limits on the findings
The fixed panel excluded tokens whose rug-pull event occurred before the latest cutoff, so its findings do not extend to those faster cases. The recovered transfer graphs were partial and did not encode transfer amounts. In the original cohort, 46.4% of samples had a recovered graph at the earliest window.
The tests used a single chronological split, with test sets of 136–159 tokens. The researchers estimated that within-window comparisons needed average-precision differences of about 0.10–0.18 for 80% statistical power. They said larger test sets were needed to establish whether the observed pattern of graph-model gains was reliable.
Sources
- Original source · Preprint Early Rug Pull Detection on Solana arXiv:2611.13342
Related stories
- Study says useful proof-of-work could raise electricity use
- Researchers propose splitting Bitcoin mining rewards in fee-only era
- Study finds LLMs often misdiagnose Solana bridge flaws
- Ledger puts Solana clear-signing engine in app version 1.17.0
- Solana Validators Face Linux Flaw Rated 9.8 as Ubuntu Patch Waits