Copying coordinated pump.fun wallets loses money even though the wallets themselves extract cash from the curve. All 32 tested configurations fail out of sample. Gafarov's paper gives a careful account of why an outsider cannot capture the same returns.

Where the early wallets get paid

A new pump.fun token trades from its first second against a constant-product bonding curve with virtual reserves. There is no liquidity provider. Buyers supply the real SOL held in the curve, so early profits come from buyers who arrive later. Wallets that consistently buy first and sell into the subsequent demand collect that transfer.

Gafarov built a collector that follows Solana block by block. It recorded 88.3 hours (26 to 30 September 2026), covering 200,443 launches and 15,854,121 curve trades. Each trade includes the post-execution reserves reported by the protocol. The paper then constructs wallet groups prospectively: a decision at time t draws only on events before t. Wallet pairs build up evidence through co-buying, shared funding, joint signing, direct transfers and consolidation of proceeds. This produces 23,710 groups.

Group members buy about 11 s after creation; other early buyers arrive at 15 s. The measured gap is 4.3 s (interval 4.2 to 4.4). Across 51,912 tokens, members' net cash flow exceeds a placebo set matched on entry time and buy size by 0.17 SOL per token [0.12, 0.23]. Late buyers average -2.34 SOL per token, leaving unsold inventory unvalued.

The evidence for a pump-like price path is weaker. Among comparable controls, 2.7% of group tokens double after the first minute, against 2.1% of controls. The gap is +0.7 pp [-0.3, +1.6]. One 126-wallet team buys about 95 tokens an hour, and 53% of its tokens reach 20 SOL, compared with 9.0% overall. Inverse-probability weighting puts the gap at +32.2 pp. At graduation, the weighted gap falls to +0.6 pp [-0.0, +1.3], suggesting a push to 20 SOL followed by departure. Nearest-neighbour matching and the within-creator contrast give +0.9 and +3.2 pp.

Can a follower get there in time?

The follower's rules are precise enough to rebuild. It buys 0.05 SOL at the first curve observation at or after the later of the signal and 4 s after creation, with latency added at 1, 2, 3 or 5 s. The paper tests two cost conventions. Under the project convention, each side pays the larger of the event fee and 2%, plus 1.5% slippage. Under the event convention, the charge is the recorded fee plus 0.3%. Exit rules are selected on the first 70% of signals, then tested on the last 30%.

Every configuration loses out of sample: returns range from -7% to -4% per trade under event costs and from -11% to -8% under project costs. At 1 s, the buyers-only follower earns -7.8% under project costs, versus -7.9% for random entry. Event costs give it -4.0%, versus -3.6%. Comparing identical trades across all signals, increasing latency from 1 s to 5 s costs 2.6 points under project costs and 2.9 under event costs. The buyers-only signals lose about one point.

For anyone trading launches, the cost decomposition is the table to read. A fixed momentum rule enters at 10 SOL within 120 s, takes profit at 1.5x, stops at -20% and holds for 60 s. Across 2,014 positions it earns +5.4% gross. The 2% fee floor reduces that to +3.7%, and slippage leaves +0.6%. One second of latency takes the return to -9.9%. Changing position size from 0.01 to 0.5 SOL shifts the result by under one point, locating the loss in timing.

Gafarov keeps the conclusion conditional: "under these cost and latency assumptions no strategy we tested earns a positive out-of-sample return." The paper explicitly leaves sub-second, colocated or bundled execution outside its one-second resolution.

Curve SOL at 60 seconds does as well as the groups

Gafarov acknowledges in the abstract that naive co-buy clustering ranks tokens at least as well as the discovery groups. The paper therefore assigns the structure a descriptive role. Its strictly prospective outcome asks whether a token below 20 SOL at 60 s reaches that level later. Against a base rate of 1.8%, naive components achieve 29.3% precision in the top 500; discovery groups achieve 10.2%. The team layer's agreement score gets to 18.5%, still trailing the naive components.

The frozen final block contains 36,532 tokens, of which 8.6% ever reach 20 SOL. Three team rules flag 2,359, 904 and 615 tokens, with precision of 50.4%, 70.6% and 73.0%. Yet 26% to 40% of those signals arrive after the curve already holds 20 SOL. Restricting the signals to earlier arrivals brings precision down to 33.3%, 53.0% and 54.9%, against base rates of 4.4% to 4.9%. A single team accounts for 67% to 79% of the flags.

The decisive comparison uses curve SOL at 60 s alone: AUC is 0.939. A first-minute model without group information essentially matches that at AUC 0.938. Adding group features changes AP by -0.015 [-0.030, +0.004] and AUC by 0.000. Gafarov defends the team result as an association that "holds for the final block". I accept that. His sentence continues, "as a statement about forecasting it does not", and forecasting is the half a trader can use.

Accounting for your own deposit

The paper's strongest contribution corrects the portfolio accounting. A time-consistent evaluator begins with 1 SOL and stakes 5% on each trade. Requiring a later observed quote for an exit leaves 88% of exits unresolved; the book finishes at 1.198 SOL across 43 trades. Using reserves for the exit while ignoring the trader's own deposit leaves 60% unresolved and a final balance of 1.121.

Including the trader's SOL deposit and token withdrawal in the exit reserves changes the result sharply. Unresolved exits fall to 2.8%, and the evaluator produces 332 positions. The final balance is 0.504 SOL, with a 54.9% drawdown. Elsewhere, a label leak generated a +300% mean return and a 100% win rate before it was caught.

Implementation still requires judgment. The evaluator includes the trader's position in reserves only at first order; it does not re-simulate how other traders react to the follower's purchase. The quote formula reproduces 99.63% of buys and 98.07% of sells within 1e-4. Gafarov did not investigate the sell mismatches further. Discovery thresholds appear in the replication package rather than the paper. We did not find the candidate exit grid in the text, only the selected rule: 1.3x take profit, 0.7x stop and a 60 s maximum hold. Parameter sweeps use 3-hour and 12-hour slices instead of the full window. A separate check repeats three headline results on each of the four days.

We could not rerun any of this.

Our crypto data does not contain the pump.fun launch universe or wallet identities. One-minute bars also cannot recover the per-event post-trade reserves on which the execution model depends.

A follower achieving positive net returns on the same signals with sub-second inclusion would change my view. The paper's one-second clock cannot test that possibility. The fast teams' own positive cash flow says an edge exists for those already inside the first seconds.