The population
Every Pump.fun coin creation is enumerated from the program’s mint-authority signature history, which only appears in creation transactions. Failed transactions are counted separately. Time ranges that could not be read are recorded as gaps rather than ignored.
A uniform random sample is followed in depth. Selection uses a hash of the creation signature, so it is independent of anything about the coin, its creator or its outcome — the sample includes failures and abandoned coins in their true proportion.
What is comparable
Coins can now be quoted in SOL, USDC or other pump coins, and “mayhem mode” coins can have their curve parameters reset by the protocol. Only standard SOL-quoted curves are used for price-based research; the others are counted and shown, but not mixed in.
Point-in-time features
Features at 1, 5, 15, 30 and 60 minutes only use observations taken at or before that moment. Snapshots are scheduled slightly ahead of each boundary for that reason. Creator history only counts earlier launches, and only outcomes that were already known at the time.
First-hour transaction counts are reconstructed from the full signature history after the hour has closed. They are correct as of each minute, but a live system would not have had them in real time; that difference is kept in mind when judging any alert.
Outcomes
Outcomes are measured from an entry point (for example 5 minutes) to horizons of 1 hour, 6 hours, 24 hours and 7 days, and stored separately from features. Peaks come from sparse snapshots, so they are lower bounds. A “net after costs” figure models a 1 SOL position with fees and constant-product slippage on the curve.
Testing
- Chronological train/test splits with an embargo gap, so nothing from the test period informs the model.
- Every feature and model is compared with simple baselines (the base rate and the best single feature).
- Results carry bootstrap confidence intervals, and false positives and misses are reported.
- No result is described as predictive unless it beats the baseline out of sample with its uncertainty taken into account.
Known limits
- Social attention and news are not yet measured — no authorised source is connected.
- Trader-level data (unique buyers, creator selling) needs a larger RPC budget and is not yet collected.
- Prices after a coin leaves its bonding curve come from DEX Screener snapshots.
- Associations are not causes, and market behaviour changes over time.