The signal arrives early
An arXiv preprint studying Solana memecoins reports that a vast majority showed the paper’s defined rug-pull characteristics within one hour of launch. The result points to a possible early screening signal: a model given the opening minutes of trading may find patterns associated with the study’s one-hour outcome. It does not show that the models can stop a rug pull or prevent investor losses.
The abstract gives no percentage for that statement. In its methods discussion, however, the paper says that over 80% of memecoins experience rug pulls. Because the study defines the outcome through specific trading and liquidity rules, that figure is a reported rate for an operational label, not a separate verification of every case as fraud.
The supplied record identifies the work as an arXiv preprint, version 1, dated 20 Aug 2026. No journal is listed in the record.
Across seven months, the researchers assembled a dataset of 6.4 million Solana memecoins. The models used tabular liquidity and trading features—structured measurements of market activity—and did not use smart-contract code analysis.
A rule-based definition
To turn a messy market event into a yes-or-no outcome, the study used an operational label. A token was marked as a rug pull when TVL—the pool value tracked by the analysis—fell below 99%, or when its idle time exceeded 80% of its lifetime without trading. That makes the outcome measurable, but it also means the headline result concerns a rule-based label rather than every possible form of fraudulent behavior.
The forecast was deliberately short-horizon. Each model received the first five minutes of trading data and predicted whether the rug-pull label would apply within one hour. The setup therefore asks a narrow question: can opening activity flag the paper’s defined one-hour outcome?
The comparison covered activity within each platform, transfer between Raydium and PumpFun, and training on fused data from the two venues. Since the feature set excluded code analysis, the experiment concentrated on liquidity and trading behavior.
Performance depended on the platform
The validation design followed token issuance time. The researchers used forward rolling time-series checks, held the last time window out as an independent test set, trained on the most recent three months, and expanded validation windows when needed to keep at least five positive rug-pull samples in each fold.
The model lineup included Random Forest, XGBoost, a multilayer perceptron, or MLP, FT-Transformer, Tab-Transformer and AutoInt. Hyperparameters—the settings selected before a model is finalized—were tuned with Optuna’s Bayesian search across the folds.
Results were summarized with positive-class F1, Matthews Correlation Coefficient, or MCC, and area under the precision-recall curve, or AUCPRC. These measures assessed how closely predictions matched the study’s positive rug-pull label, rather than directly measuring prevented losses.
The processed test set was unevenly distributed across the two platforms. It contained 43,835 PumpFun tokens carrying the rug-pull label and 9,711 without it, compared with 2,931,946 labelled rug pulls and 1,893,539 non-rug-pulls on Raydium. These counts describe the evaluation set used in the analysis.
Within-platform tests showed MCC scores around 0.25–0.36 across most models. That gave the systems a measurable signal when training and testing took place within the same general setting, although the score remained tied to the study’s operational label.
The picture changed when models crossed the platform boundary. Cross-DEX prediction produced a marked performance drop, with MCC approaching zero or turning negative across nearly all models. Patterns learned on one venue did not transfer cleanly to the other in the reported experiments.
A stronger benchmark, with a large warning
Among the model families, Random Forest and XGBoost were more robust in the reported cross-domain comparisons than the neural-network-based models. The result underlines how much the data distribution mattered when the platform changed.
Fused data produced one of the paper’s highlighted results in a PumpFun evaluation. XGBoost trained on the combined data reported a positive-class F1 of 0.7885, an MCC of 0.3947 and an AUCPRC of 0.8011. Those are the reported scores for that setup, not a guarantee of performance on another platform or time period.
The supplied analysis reports no confidence intervals for the cross-platform and fused-data results. It also reports no confidence intervals or significance tests for the within-platform range. Those omissions make the point scores useful benchmarks, but leave the reliability of the differences less certain than the numbers alone might suggest.
The analysis has a narrow technical scope. It covers Raydium and PumpFun over the seven-month observation period, and its inputs are liquidity and trading features rather than smart-contract code. The results therefore do not establish performance on other exchanges, in later periods or with code-level signals.
Nor does the operational label equal independently verified fraud by itself. It is built from the TVL and idle-time thresholds, and the study does not directly measure whether investors were spared losses. The evidence is therefore narrower than the word rug pull might suggest.
The authors describe the work as a baseline and say the reported performance is not yet sufficient for real-world deployment in high-risk financial environments. They also warn that false negatives—rug pulls the model misses—could carry substantial investor losses.
What the study leaves open
The next tests identified in the analysis include independent validation on additional Solana DEXs and later market periods, along with work combining liquidity signals with bundle-level and behavioral features. The authors also leave open whether temporal graph or sequence-based models could improve performance.
Another open question is whether prediction horizons shorter than five minutes could provide useful additional warning time, and whether deployment with DEX platforms would reduce actual investor losses. Those are questions for future testing, not findings established by this preprint.
The narrower conclusion is that early tabular trading data can produce a detectable signal for the study’s operational one-hour label, particularly in within-platform testing. Cross-platform reliability and real-world investor benefit remain unestablished.
Paper data and sources
Original title: Catching the Rug: Early Prediction of Fraudulent Memecoins on Solana via Machine Learning
Authors: Jianghai Li, Pavel Kuznetsov, Yury Yanovich et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text