What Auditing 832,941 Pump.fun Launches Taught Me About Memecoin “Success”

In September I withdrew it.

This is not a story about a typo. It is about a measurement trap that anyone who builds or reads on-chain statistics can fall into, and explaining it is more useful than the number ever was.

What “graduation” was supposed to measure

On pump.fun, a new token trades on a bonding curve. If enough buyers arrive, the curve completes and the token “graduates” to a regular trading pool. The share of launches that graduate is a natural headline for how hard the platform is for new tokens.

To measure it, I ran a collector that watched new launches as they appeared and recorded one of two outcomes for each token: graduated, or not graduated within 24 hours (“timeout”).

“Not graduated” was really “not seen graduating”, credit: Arati Uday Kamat

What the labels really measured

The blockchain records every graduation, but my collector did not read the chain’s full history. It watched the stream of new launches, and new launches arrive continuously, so most tokens were watched closely only in the first few minutes of their life.

That had two consequences. A “graduated” label really meant “completion seen while the collector was still watching”, which in practice meant within minutes of launch. And a “timeout” label did not mean the token failed: it mixed tokens that never completed with tokens that completed later, after the collector had stopped looking.

So 0.198% was never the 24-hour graduation rate. It described a fast lane: tokens that completed almost immediately. Every statistic built on those labels inherited the same limit, including the comparison with an earlier, higher rate and the apparent effect of advertising a Telegram channel, so I withdrew those too. The corrected dataset now treats every timeout row as censored (“not seen completing”), not as a failure.

A model that looked excellent, then did not

In the audited version of the paper I also fitted a prediction model on one 15-day period and then tested it, locked and unchanged, on the following 14 days, a comparison I had specified in advance. On the first period the model looked excellent: an AUROC of 0.86, where 0.5 is a coin flip and 1.0 is perfect. On the next 14 days it scored 0.46.

My later check found the main cause. The model contained a calendar term fitted to the dates of the first period; frozen, that term no longer matched the new dates. So the drop is mainly a lesson about model design, not evidence that the market itself changed. I have published that clarification in the dataset’s correction notice.

Five lessons for anyone reading memecoin statistics

1. Ask how the outcome was observed, not just what it was. “Didn’t graduate” and “wasn’t seen graduating” are different claims. Most viral statistics never tell you which one you are reading.

2. The chain is complete; your collector is not. On-chain data is often described as a perfect record. The ledger is, but scrapers, APIs and dashboards sample it, and every sample has blind spots.

3. A great backtest is a hypothesis, not a result. Features that quietly encode the calendar, or anything else specific to one window, can make a model look strong and then collapse on new data. If a trading signal has not been tested on data it never saw, treat it as unproven.

4. Be suspicious of a single precise number. “0.198%” sounds exact, but the precision of a decimal is not the accuracy of a measurement. Graduation figures measured with different methods and windows disagree. A large study of pump.fun by Szwajcok, Tsuchiya, Liu, Soska, Payer and Christin (“Meme Coin Factories”, arXiv 2609.10246) found that 1.02% of coins in their sample graduated. Ranges with their method attached are more honest than one viral figure.

5. Corrections should travel as far as the original claim. They rarely do. That is why I am writing this piece.

What still holds

The core lesson for traders is unchanged, and if anything stronger. Graduation on launchpads is rare, automated trading is heavy in the first minutes, and a token’s first impression tells you little about its later life. What changed is how confidently any single number can describe it.

My datasets and correction notes remain public, so anyone can check the work, including the parts I later corrected.

A separate study of mine on coordinated wallet cohorts (arXiv 2607.02795) is also being corrected, because its buyer records turned out to include sell transactions; its figures, including the 1,012 cohorts, should not be used until the corrected version is out.

This article is for information only and is not investment advice.

Research links

Auditing Collector-Generated Graduation Labels on Pump.fun: arXiv:2607.02823

RED-PUMP-2026 dataset and correction notice: doi.org/10.5281/zenodo.20633486

About the author

Arati Uday Kamat is an independent researcher working on market microstructure and on-chain measurement. ORCID 0009-0000-4781-312X.