TL;DR for operators

Five minutes of on-chain trading activity contain enough signal to support one-hour rug-pull screening on Solana, but the signal is not reliably portable across venues. A model that performs reasonably on PumpFun can lose much of its separation when moved to Raydium, and vice versa. Training on both venues helps—especially for XGBoost and Random Forest—but does not remove the platform effect.

For wallets, DEXs, and risk platforms, the practical unit of deployment is therefore not “Solana memecoin fraud detection” in the abstract. It is a venue-conditioned screening system with separate validation, monitoring, and escalation rules. The paper’s target is also narrower than confirmed fraud: it flags tokens whose liquidity collapses sharply or whose trading becomes inactive. That is useful for risk triage, not proof of fraudulent intent.

A shared Solana model can fail at the venue boundary

A risk team has a familiar incentive: train one detector, then reuse it wherever the same assets and broadly similar transaction fields appear. PumpFun and Raydium both operate on Solana, so a shared model may look like a reasonable way to avoid duplicated training and maintenance.

The paper by Jianghai Li and colleagues tests where that assumption breaks.1 It uses first-five-minute transaction and liquidity features to predict whether a token will meet a rug-pull condition within the following hour. Within a venue, the models retain measurable predictive separation. Move the trained model to the other venue, however, and several results collapse.

The clearest measure is Matthews Correlation Coefficient, or MCC: a balanced classification score where values near zero indicate little useful separation and negative values indicate predictions moving in the wrong direction. XGBoost records MCC 0.3562 when trained and tested on PumpFun and 0.2914 on Raydium. Under direct transfer, Raydium-to-PumpFun falls to -0.0026, while PumpFun-to-Raydium falls to -0.0664. Random Forest reaches -0.2433 in the PumpFun-to-Raydium transfer.

The chain did not change. The target did not change. The feature categories did not change. The venue did.

Five minutes are enough to make prediction plausible

The study covers roughly 6.4 million reported token addresses from November 30, 2024 through June 30, 2025. Model inputs summarize the first five minutes of activity: transaction counts, buy/sell shares and values, unique buyers and sellers, price changes and dispersion, timing variables, and price extrema. Tokens that had already met the rug-pull condition during those five minutes were excluded from the prediction evaluation.

The target combines two operational failure modes. A token is marked high risk if liquidity suffers an extreme TVL drawdown or trading becomes sufficiently inactive. In the reported experiments, this is described as a 99% TVL-drop condition OR an 80% idle-time condition.

This is a measurement boundary, not a footnote. The paper is not predicting independently adjudicated criminal intent. It predicts a market outcome associated with severe liquidity withdrawal or trading abandonment.

The validation design also matters. Models are trained and validated in issuance-time order using recent historical windows, with the final period held out for one-time testing. The results therefore speak to out-of-time prediction rather than performance on a randomly shuffled sample.

The model is learning a venue as well as a risk pattern

The descriptive statistics point to materially different early-trading microstructures. PumpFun is characterized as a many-pool, smaller-trade environment; Raydium has fewer pools with much heavier early activity. In the five-minute window, average transaction counts are about 39.3 for PumpFun versus 588 for Raydium, with average buyers around 18.3 versus 154 and sellers around 13.0 versus 84.2.

Those differences change the statistical meaning of otherwise familiar inputs. A transaction count, buy/sell ratio, price move, or timing pattern learned as suspicious on one venue may occupy a very different distribution on the other. This is domain shift: the variables have the same names, but the operating regime has changed.

For a production team, that changes the diagnostic question. Weak transfer does not automatically imply a need for a more elaborate neural architecture. In this benchmark, MLP, FT-Transformer, TabTransformer, and AutoInt do not consistently solve the transfer problem. Random Forest and XGBoost are generally more robust in cross-domain and fused-data settings.

Fused training helps, but unevenly

Exposing the model to both venues during training produces the paper’s most constructive result.

Training / test setting XGBoost MCC XGBoost AUCPRC Interpretation
PumpFun → PumpFun 0.3562 0.7564 Same-venue baseline
Raydium → PumpFun -0.0026 0.5771 Direct transfer largely loses separation
PumpFun → Raydium -0.0664 0.5792 Transfer also degrades in the other direction
Fused → PumpFun 0.3947 0.8011 Best reported PumpFun XGBoost result
Fused → Raydium 0.1543 0.6426 Improvement over failed transfer, but below same-domain Raydium

For PumpFun, fused training raises XGBoost MCC from 0.3562 to 0.3947 and AUCPRC from 0.7564 to 0.8011. Random Forest shows a similar pattern, reaching MCC 0.3942 and AUCPRC 0.7978.

The gain is not symmetric. Fused-to-Raydium XGBoost MCC is 0.1543, below the 0.2914 same-domain Raydium result. Multi-platform training therefore broadens robustness without establishing a venue-independent model.

What risk teams should change

The paper directly supports three operating choices. Five-minute transaction and liquidity summaries are plausible inputs for an early screening layer. Validation should be segmented by venue rather than reported only as an aggregate Solana score. And whenever a model is reused across market structures, teams should track cross-domain MCC, AUCPRC, and false negatives rather than infer portability from within-domain performance.

Cognaptus extends that evidence into a narrower deployment rule: use the model for prioritization rather than an automated fraud verdict. A wallet could trigger a warning or additional review. A DEX or monitoring provider could combine the score with other evidence before taking a high-cost action.

The paper also identifies what the current feature set does not capture. Bundle-level entity resolution, wash-trading patterns, transaction sequences, and graph relationships could expose coordinated behavior hidden by token-level aggregates. Their value remains a research hypothesis here; the benchmark does not test those additions.

Screening evidence is not yet deployment evidence

Two boundaries materially affect practical use.

First, the label is a proxy. A 99% TVL collapse or prolonged inactivity can identify a dangerous outcome without establishing developer intent or legal fraud. Second, the reporting contains unresolved accounting tensions: the paper describes rug pulls as a minority class even though its processed test-set counts show Rug Pull labels exceeding Non Rug Pull labels, and the platform token totals are difficult to reconcile directly with those test-set counts. PumpFun-to-Raydium migration may explain part of the mismatch, but the source package does not resolve it.

The authors characterize the work as a baseline and state that current performance is insufficient for real-world deployment in high-risk financial settings. The transfer results reinforce that boundary.

The narrower conclusion is more useful: early market behavior can support rapid Solana memecoin screening, but portability is an empirical property to be tested venue by venue, not an assumption inherited from the shared blockchain.

Cognaptus: Automate the Present, Incubate the Future.

  1. Jianghai Li and Pavel Kuznetsov and Yury Yanovich and Konstantin Nott-Whaley and Igor Vodolazov (2026). Catching the Rug: Early Prediction of Fraudulent Memecoins on Solana via Machine Learning. arXiv:2608.20271. https://arxiv.org/abs/2608.20271 ↩︎