In July I launched a preprint that analysed 832,941 token launches on pump.enjoyable and reported that solely 0.198% of them “graduated” inside 24 hours. The quantity travelled additional than something I’ve revealed: change blogs, information websites and social media posts repeated it, in lots of languages.
In September I withdrew it.
This isn’t a narrative a couple of typo. It’s a couple of measurement entice that anybody who builds or reads on-chain statistics can fall into, and explaining it’s extra helpful than the quantity ever was.
What “commencement” was presupposed to measure
On pump.fun, a brand new token trades on a bonding curve. If sufficient patrons arrive, the curve completes and the token “graduates” to a daily buying and selling pool. The share of launches that graduate is a pure headline for the way exhausting the platform is for brand spanking new tokens.
To measure it, I ran a collector that watched new launches as they appeared and recorded certainly one of two outcomes for every token: graduated, or not graduated inside 24 hours (“timeout”).
“Not graduated” was actually “not seen graduating”, credit score: Arati Uday Kamat
What the labels actually measured
The blockchain data each commencement, however my collector didn’t learn the chain’s full historical past. It watched the stream of recent launches, and new launches arrive repeatedly, so most tokens have been watched intently solely within the first jiffy of their life.
That had two penalties. A “graduated” label actually meant “completion seen whereas the collector was nonetheless watching”, which in observe meant inside minutes of launch. And a “timeout” label didn’t imply the token failed: it blended tokens that by no means accomplished with tokens that accomplished later, after the collector had stopped trying.
So 0.198% was by no means the 24-hour commencement price. It described a quick lane: tokens that accomplished nearly instantly. Each statistic constructed on these labels inherited the identical restrict, together with the comparability with an earlier, increased price and the obvious impact of promoting a Telegram channel, so I withdrew these too. The corrected dataset now treats each timeout row as censored (“not seen finishing”), not as a failure.
A mannequin that regarded glorious, then didn’t
Within the audited model of the paper I additionally fitted a prediction mannequin on one 15-day interval after which examined it, locked and unchanged, on the next 14 days, a comparability I had specified prematurely. On the primary interval the mannequin regarded glorious: an AUROC of 0.86, the place 0.5 is a coin flip and 1.Zero is ideal. On the following 14 days it scored 0.46.
My later verify discovered the principle trigger. The mannequin contained a calendar time period fitted to the dates of the primary interval; frozen, that time period not matched the brand new dates. So the drop is principally a lesson about mannequin design, not proof that the market itself modified. I’ve revealed that clarification within the dataset’s correction discover.
5 classes for anybody studying memecoin statistics
1. Ask how the end result was noticed, not simply what it was. “Didn’t graduate” and “wasn’t seen graduating” are completely different claims. Most viral statistics by no means inform you which one you might be studying.
2. The chain is full; your collector will not be. On-chain knowledge is commonly described as an ideal report. The ledger is, however scrapers, APIs and dashboards pattern it, and each pattern has blind spots.
3. An incredible backtest is a speculation, not a consequence. Options that quietly encode the calendar, or anything particular to 1 window, could make a mannequin look robust after which collapse on new knowledge. If a buying and selling sign has not been examined on knowledge it by no means noticed, deal with it as unproven.
4. Be suspicious of a single exact quantity. “0.198%” sounds actual, however the precision of a decimal will not be the accuracy of a measurement. Commencement figures measured with completely different strategies and home windows disagree. A big examine of pump.enjoyable by Szwajcok, Tsuchiya, Liu, Soska, Payer and Christin (“Meme Coin Factories”, arXiv 2609.10246) discovered that 1.02% of cash of their pattern graduated. Ranges with their technique connected are extra trustworthy than one viral determine.
5. Corrections ought to journey so far as the unique declare. They not often do. That’s the reason I’m scripting this piece.
What nonetheless holds
The core lesson for merchants is unchanged, and if something stronger. Commencement on launchpads is uncommon, automated buying and selling is heavy within the first minutes, and a token’s first impression tells you little about its later life. What modified is how confidently any single quantity can describe it.
My datasets and correction notes stay public, so anybody can verify the work, together with the elements I later corrected.
A separate examine of mine on coordinated pockets cohorts (arXiv 2607.02795) can also be being corrected, as a result of its purchaser data turned out to incorporate promote transactions; its figures, together with the 1,012 cohorts, shouldn’t be used till the corrected model is out.
This text is for info solely and isn’t funding recommendation.
Analysis hyperlinks
Auditing Collector-Generated Commencement Labels on Pump.enjoyable: arXiv:2607.02823
RED-PUMP-2026 dataset and correction discover: doi.org/10.5281/zenodo.20633486
Concerning the creator
Arati Uday Kamat is an impartial researcher engaged on market microstructure and on-chain measurement. ORCID 0009-0000-4781-312X.
Arati Uday Kamat Arati Uday Kamat Read More
