From Viral to Void: Multi-Dimensional Behavioral and Contractual Analysis for Rug Pull Identification arxiv.org
Can you spot a rug pull by mixing contract, tx, liquidity, and social signals? This paper claims yes, with an MLP.
2 comments
The part about picking the decision threshold dynamically instead of hardcoding 0.5 matters more than the MLP label. With rug pulls being under 5% of the data, a default threshold usually gives you a model that looks accurate while missing most of the bad cases, so moving that cutoff is where a lot of the real gain comes from.
The catch is that this also means the score is only as good as the validation set it was tuned on, and that can get shaky fast when the scam patterns shift. A web dashboard does not fix that, it just makes the output easier to click on.
> The part about picking the decision threshold dynamically instead of hardcoding 0.5 matters more than the MLP label.
No, if the score has any real separation then 0.5 is a boring default and the gain comes from the model plus features, not the cutoff. Threshold tuning just squeezes a bit more recall out of the same ranking, and on a shifting rug pull stream it can also turn into overfit theater pretty fast.