In July, a preprint analysis of 832,941 token launches on the pump.fun platform reported that only 0.198% of them graduated within 24 hours. This figure quickly gained traction, appearing on exchange blogs, news sites, and social media in multiple languages. However, in September, the author withdrew the claim.
The issue wasn't a simple typo but a deeper problem with measuring on-chain statistics. The study had defined 'graduation' as a token completing its bonding curve and moving to regular trading. The collector monitoring these launches observed outcomes only within the first few minutes of a token's life, leading to mislabeling. Tokens labeled as 'not graduated' might have simply completed later, after the collector stopped watching.
The corrected dataset now treats 'timeout' labels as censored data rather than failures. The author also found flaws in a prediction model that initially appeared strong but failed on new data. The model had included a calendar term that didn't match subsequent dates, highlighting the importance of robust model design.
The author shared five key lessons for interpreting memecoin statistics, emphasizing the need to question how data is observed, the limitations of collectors, and the dangers of relying on single precise numbers. They also noted that corrections rarely travel as widely as the original claims, prompting this article to address the issue transparently.