Notes · 2026-10-11
My model's 139 favourite memecoins: 137 were the developer buying his own launch
I trained a model to pick new Solana memecoins worth buying in their first minutes. It scored each coin a few minutes after launch, and the best scores became calls. On paper it looked like it was finding something.
Then I looked at what it had picked. Of the 139 coins over its call line, 137 were the same thing.
The template
A coin on pump.fun starts on a "bonding curve": a fixed sale that anyone can buy into. When the sale is full the coin "graduates" into a normal trading pool. That usually takes hours, if it happens at all.
In these 137 coins it took one transaction.
- The creator bought the whole sale at launch: 793.1 million tokens for about 85 SOL (roughly $9,900).
- The coin graduated at once, with the creator holding 79% of it.
- Bots then traded it about 430 times a minute, and within five minutes the price stood near nine times the graduation price.
To a model that looks at activity, buyers and price, this is the best coin it has ever seen. To a person, it is one wallet holding four fifths of the supply and a crowd of bots making it look alive.
Why it can't be traded
I measured every trade on 1,687 coins, timed to the millisecond, against the price of the pool's first trade.
| 1 second | 10 seconds | 1 minute | 4 minutes | |
|---|---|---|---|---|
| Buyout coins | 8.0x | 8.1x | 8.1x | 8.8x |
| Every other coin | 1.02x | 1.01x | 0.88x | 0.54x |
The whole eightfold jump happens inside the first second, in practice the same block as the launch. Nobody outside that block gets it. After that the price is flat until the creator sells.
So buying one of these is a race against that sale. On 893 of them bought eight minutes in:
- a third reached +30% before the creator pulled out; those averaged +22%;
- 37% were pulled about six minutes after the buy, for a loss of 45% to 67%.
I tried to tell the two groups apart with everything I had recorded: the creator's share, the top holders, bundles of same-size buys, buyer counts, order flow, the socials. Each one separated them barely better than a coin flip. On 60 pulls I studied one by one, no signal came first.
What the ordinary coins do
Once the buyouts were set aside, the model had almost nothing left to call: about one coin a day. And ordinary new coins, bought eight minutes after they graduate, lost 22.5% a trade across 391 of them. Only 27% ended above zero.
What I changed
- Buyout coins are recognised at launch (97% of them graduate within three seconds of launching; 1% of other coins do) and are never called or traded.
- The model was retired by a rule written before the result: its picks on coins it had never seen averaged -17% over 74 trades.
- Every later test is judged only on coins the model has not seen, with fills at prices that were really there.
The uncomfortable part is not that the model failed. It is that it failed while its numbers looked good, and the only thing that showed it was reading what it had actually picked, coin by coin.
This is software research on public data, not financial advice.