Pump Anatomy

Real data · Solana

Which early, measurable features are associated with what happens next — tested out of sample, against simple baselines, with failures included. No predictive claim is made until the evidence supports one.

The study population

Since 2026-11-01 16:31:49 UTC · every stage keeps the failures in

Launch transactions enumerated115,794
every pump.fun creation signature, complete for the covered windows
Succeeded on-chain106,146
9,648 failed and are kept out of the population
Selected by the uniform random sample2,197
hash-based, so selection is independent of outcome
Decoded and followed2,184
launch, creator and quote asset verified from the transaction
Comparable standard SOL curves931
excludes 739 mayhem-mode and 526 non-SOL launches
With point-in-time features2,146
computed after the first hour closes
With measured outcomes915
measured once each horizon has passed

What usually happens next

Share of comparable launches, measured from the 5-minute mark. Base rates are the baseline any signal must beat.

Horizon	Launches	Curve completed	Reached 2x (observed)	Above 5m price	Profitable after costs	Effectively dead
1h	915	2.1%	0.6%	8.2%	0.7%	1.1%
6h	808	2.2%	0.6%	6.7%	0.7%	1.4%
24h	554	2.6%	0.9%	7.4%	0.6%	1.7%

“Reached 2x” uses sparse snapshots, so it understates peaks; “profitable after costs” models a 1 SOL entry at 5 minutes with fees and slippage. Small samples move a lot — treat early figures as provisional.

Latest association study

Chronological train/test split with an embargo, per-feature AUC with bootstrap intervals, and a simple model compared with the best single feature.

insufficient data

Not enough labelled launches to test anything yet (68 of 200 labelled launches needed; 0 of 20 positive launches needed in train; 0 of 20 positive launches needed in test; 0 of 20 negative launches needed in test). Sir FOMO reports nothing rather than guess.

Why there is no result yet

68 of 200 labelled launches needed
0 of 20 positive launches needed in train
0 of 20 positive launches needed in test
0 of 20 negative launches needed in test

Read with care

Multiple comparisons: 13 features were tested, so some apparent associations are expected by chance alone.
Association is not causation: these features describe launches; they do not explain why outcomes happened.
Results depend on the sample period (2026-11-01T16:34:50.000Z to 2026-11-01T19:35:37.000Z); market conditions change and may not repeat.
Peak-based outcomes come from sparse snapshots and under-count true peaks; at-horizon and after-cost outcomes are more conservative.
803 launches had no outcome label yet and were excluded.

Run: #12 · pump-anatomy-market-only-v1
Started: 2026-11-03 22:42:29 UTC (2.6 h ago)
Code version: not recorded
Config hash: 601e6f453a20ce1a…