Four strategies we killed. Six errors we caught in our own maths — every one of which made us
look better than reality. And a live bot that lost 79.7% of its money, diagnosed to the cent.
You have been burned before. So don't trust us. Check us.
On 8 August 2026 a real bot traded real money on Solana. Eight round trips. Zero winners. Here is every trade, read from the chain, not reported by the bot.
| token | hold | in (SOL) | out (SOL) | result |
|---|---|---|---|---|
| DQuzj9z4 | 14m | 0.00268 | 0.00007 | −97.2% |
| HMN4VYbP | 1m | 0.00268 | 0.00044 | −83.5% |
| 2TUyVhTF | 2m | 0.00259 | 0.00036 | −86.0% |
| 36ReTtqY | 1m | 0.00318 | 0.00087 | −72.6% |
| 6CgtnoAn | 1m | 0.00318 | 0.00087 | −72.6% |
| CvNUQwwW | 0m | 0.00318 | 0.00087 | −72.6% |
| GG88qwH3 | 1m | 0.00318 | 0.00087 | −72.6% |
| EygsXrc8 | 1m | 0.00268 | 0.00038 | −85.7% |
Two theories were proposed. We tested both and both were wrong.
✗ “The money was stolen.” No. Every “Sent to…” entry is the token leg of a sell going to the liquidity pool — SOL came back in the same transaction, checked one at a time.
✗ “MEV bots sandwiched it.” No. We pulled the block each swap landed in and looked for an address that bought the same token before and sold after — the definitional signature. 5 of 5 conclusive, zero sandwiches. In every block there were zero other trades on that token at all.
network fees 8% of the loss slippage 92% of the loss cost per round trip: 80% of position size
Six of the eight tokens now show $0 liquidity. The pools were empty, so his own buy pushed the price up and his own sell pushed it back down. He was both sides of his own sandwich. At 80% cost a token must rise 4.92× just to break even — no signal survives that, however good.
The gate built from those eight trades, replayed against them: 8 of 8 blocked. 100% of the loss preventable.
Not by predicting better. By refusing to trade pools you cannot exit.
This is the section no vendor publishes. Every one of these was found in a single day, in our own work, and every single one flattered us. Not one made us look worse than reality.
That is not bad luck. It is what uncorrected measurement always does — which is precisely why the numbers you are shown elsewhere are usually too good.
| the error | what it claimed | the truth |
|---|---|---|
| Correlation aligned by array position, not date | ρ = 0.03 “independent!” | ρ = 0.562 |
| Leverage table compared a fraction to a percent | 10× safe, −1.1% DD | −107% — liquidated |
| Slippage measured against a stale price | −0.44 bps (free money) | alpha booked as a rebate |
| Cumulative return printed as “annual” | +108.6% | 10%/yr |
| Rate-limit read as “no data” | “0 fees, nothing happened” | we never looked |
| Ledger dedupe deleted 15 buried papers | N = 10 | N = 44 |
If a leverage number and a drawdown number look too good together, they are wrong.
Strategies we tested, killed, and published anyway.
| hypothesis | cause of death |
|---|---|
| Residual momentum | Sharpe −0.626 |
| Vol-scaled TSMOM | ρ 0.562 — redundant |
| Forced-flow reversal, daily | 5 events in 819 days |
| Forced-flow reversal, 5-minute | gross 0.509 → net −0.035 |
| Leveraged directional perps | no out-of-sample edge |
| Cross-sectional momentum | PBO 97% |
The fourth one is worth reading twice. The mechanism was real — a gross Sharpe of 0.509 across 171 independent events. Forced sellers really do push price past fair value. But the premium is paid to whoever is already resting on the bid, not to someone crossing the spread to get in. Measured costs took it to −0.035.
A standard flat-cost assumption would have shown that strategy printing money. We measured instead, and buried it.
One surviving edge out of everything tested. We are not going to tell you it is 500%. The leverage figure comes from the worst five-minute tick across 690,014 observations, not a friendly daily close — which is exactly the difference between 1.87× and the 10× our own broken table originally reported.
Every number ships with its window. The same strategy shows Sharpe 1.111 across its full history and 0.112 over the last three years. Same code — only the window changed. A number without its window is not a number.
Gross and net, always both. A bot showing “+X%” on price movement while paying most of it in costs is reporting fiction.
The venue is the truth, not the bot. An agent once reported +570% while the chain showed −42%.
“Could not check” is never “it is fine.” Any tool that cannot fully read its data refuses to draw a conclusion.
We log the trades we refuse, with the price at refusal — so in thirty days anyone can ask the question that could embarrass us: of the trades you blocked, how many would have won?
A rule that cannot be proven wrong is superstition, not discipline.
None of the above requires trusting us. All of it is public:
wallet 7gXQR15saMBtc6hzJKeyAY7tL1rJet5r7rNLxKyJxxTt
every trade, on Solana mainnet, forever
carry Binance Vision — funding, basis, spot
2,396 days · 2020-01 → 2026-07 · free, no key
leverage 690,014 five-minute basis observations
worst-tick drawdown, not daily close
calibration 1,579 resolved prediction markets
Beta-Binomial with credible intervals
Run the same queries. If we are wrong, you will find it — and we would rather you found it than a customer did.