Learnability of pump.fun Meme Tokens

Pump.fun launches tens of thousands of meme tokens; the natural question is whether anything observable early in a token's life predicts its fate. The project poses this as several distinct learning problems on one feature store rather than the ill-posed 'will it moon': P1 graduation (does the bonding curve complete and migrate to an AMM), P2 magnitude (peak return, a heavy-tailed regression), P3 rug/manipulation (weakly-supervised anomaly detection), and P4 archetypes (unsupervised clustering). Graduation is chosen as the clean, unambiguous, on-chain milestone to anchor everything else.

The stated bar is deliberately high: correct causal-time discipline, survival framing, regime-aware validation, calibrated probabilities, effect sizes with uncertainty, and interpretable models — explicitly not accuracy theater. Two prior attempts (a market-cap heuristic and a supervised rug classifier) are demoted to baselines, and one of them is shown to have a leaky label that partly re-learned itself.

Every model answers one question: given only what was observable in the window [0, T] after launch, predict something about (T, end]. Features come strictly from [0, T], labels strictly from after T, and a leakage linter checks this per feature. Splits are time-based walk-forward only — train on earlier launch cohorts, test on strictly later ones — because random splits leak the regime. Because graduation is a single-digit-percent event, the reported metrics are PR-AUC, lift, precision@k, and calibration, since precision@k mirrors the real decision of aping into only k tokens.

The data engineering unlocks scale. Rather than grind the rate-limited /coins API (~3/sec), the pipeline reads the on-chain bonding-curve account directly: getMultipleAccounts fetches 100 accounts per call and the graduation complete flag is byte 48 of the 81-byte layout. This resolved 98,175 mints in 7.5 minutes on a keyless public RPC (~220/sec, ~50x faster) and verified 100% (300/300) against /coins ground truth. The approach then scales to a 1.25M-token, 13-month population from a free Kaggle dataset, with 96% of graduations resolved via batched RPC.

Honest labels replace the discarded leaky one: graduated (bonding-curve migration), dud (never moved past the launch floor, 81%), and honest_rug (moved past floor then collapsed >=70% off a real peak), with the explicit caveat that honest_rug is a collapse label, not proof of malicious intent, because that needs trade/LP-level data deliberately not collected.

The headline modeling result is not a single best model but a frontier, produced by the L6 map experiment. Cheap metadata — creator prior-launch history, social links, launch timing — is available for every one of the 1.25M tokens but caps at an out-of-time ROC-AUC of roughly 0.66. Expensive on-chain microstructure — holder counts, bot scores, churn — is available only for the ~2,000-token supervised frame (the 1,969 tokens with both early features and a realized outcome) but reaches roughly 0.91. This is the real design choice: you trade sample size and coverage against signal strength.

Feature-family ablation shows creator history dominates the metadata regime while launch timing is nearly noise. On the strong side, near-launch holder count alone hits AUC 0.85 and survives age-stratification (0.79-0.90 within every snapshot-age quartile), confirming it is genuine early traction and not a snapshot-timing artifact; early transaction frequency is a strong inverse tell at AUC 0.26, flagging bot spam that churns without accreting holders. A separate 'constructing the right problem' experiment (L2) narrows the population — for instance to creators with a prior graduate — and its uncaptured runtime output indicates this simultaneously raises the base rate, the out-of-time ROC-AUC, and the top-5% lift; the population report's comparable cut shows a prior-graduate rate of 1.32% vs 0.91% (a ~1.5x lift).

Treating graduation as success turns out to be a mistake, and the second half of the project says so directly. Studying tens of thousands of graduated mints from Dune's base_trades table — with the AMM phase cleanly isolated because pool sol_reserves exceed the ~115-SOL bonding-curve completion cap — a hand-implemented, censoring-aware Kaplan-Meier estimator (product-limit, Greenwood variance bands) finds median post-graduation lifetime of about 1.2 minutes. 88.1% of graduates are dead within an hour and 97.3% within a week; only ~8% 'sustained' tokens trade for hours to days. The dominant failure mode is death by silence at a modest drawdown, not a dramatic LP pull — a strict rug flag (>=90% drawdown and dead) fires on only 2.3%.

The magnitude story is a power law. Post-graduation peak-reserve multiple has a Hill-estimator exponent of alpha ~2.31, with median 1.35x, p99 14.9x, and 89% of peaks arriving within the first hour of migrating. Yet the run is predictable out-of-time from reserve-free behavioral features: a >=2x classifier reaches ROC ~0.92 with ~4.9x top-decile lift, and the rarer >=5x reaches ROC ~0.94 with ~10x lift in the top 5%. Counter-intuitively, the narrower and more concentrated the graduation, the bigger the post-grad run — the same 'concentrated beats organic' pattern the graduation half surfaced.

The honesty checks are the backbone of the work, not an afterthought. The on-chain graduation decoder was verified 300/300 against the API ground truth before being trusted at scale, and independently cross-checked against a mined prior codebase that decodes the identical bonding-curve layout and confirms the same byte-48 complete flag. The prior work also independently flags the old is_rug label as leaky — it thresholded zero-balance ratio, which was simultaneously a model input — validating the decision to discard it.

In the post-graduation return study, a specific leak is caught and disclosed: reserve-level early features like max SOL can mechanically read the target for fast-graduating tokens whose first-5-minute window overlaps the post-grad peak (that single feature scores AUC 0.95 for >=5x, a mechanical leak). The reported models therefore use only reserve-free behavioral features. Calibration is checked out-of-time on decile bins to confirm the predicted probabilities are honestly ordered, and every headline caveat — reserve semantics as a liquidity proxy rather than price, survivorship conditioning on graduation, right-censoring of recent cohorts — is stated plainly.

The contribution is methodological as much as empirical. It demonstrates that the single discipline where most meme-coin models die — temporal leakage and regime leakage — can be enforced end to end, and that doing so overturns a headline 'success rate' by more than an order of magnitude (~52% to under 1%). The learnability map reframes the modeling question from 'what is the best model' to 'what feature-cost point can you afford,' which is the decision an actual operator faces.

The limits are owned rather than hidden. The honest_rug and reserve-based magnitude measures are collapse and liquidity proxies, not proof of intent or tradeable returns; the strong-feature regime rests on only ~2,000 tokens; the population is one-month-dominated; and the P&L model is explicit that only its win rate is real (out-of-time model precision) while the payoff multiples are documented assumptions. The through-line is that honest negative results are results: an early signal that cannot beat base rate at a given horizon is reported as such, and the frontier is quantified rather than papered over.