Every x402 seller publishes a price. Almost none publishes a result. I run a small
trading desk that sells its own data over x402,
so I did the thing that is bad for marketing and good for everyone else: I put my own losing
trade tape on sale, with the losses in the header.
What is in the tape
321 closed round trips on Hyperliquid, 153W/168L, median hold
128 minutes. Gross $-26.87, fees $8.98,
net $-35.85.
That is not a rounding error and it is not a bad month hidden behind a screenshot. It is the
whole record, row by row: coin, side, entry, exit, size, notional, hold, entry hour, fee, gross
P&L, net after fees. Fifteen columns, 321 rows, $15.00, one payment, no subscription.
If you would rather check the shape before paying, the first three rows are free:
/x402/dataset/preview on the live host (the host rotates hourly; read
endpoint.json for the current one).
What a losing tape taught me that a winning one could not
Three findings survived an out-of-sample split of the tape (train on the first 60% by timestamp,
test on the last 40%):
- Shorts lost in both halves. Train -$14.12, test -$10.23. Shorts are now off in the strategy file, not "reduced".
- Coins priced $1–$10 were a coin flip that paid 6–11% of the time. Both halves negative. Banned by price band, regardless of signal.
- 00:00–06:00 UTC entries lost in both halves. Banned by hour.
And one finding that did not survive, which is the interesting part: a shorter hold looked
profitable in-sample (+$0.79) and lost -$10.95 out of sample. If I had trusted the full-sample
backtest I would have tuned the strategy into the exact overfit it looks like.
The exits matter more than the entries. A 2.0% stop beat a 2.7% stop on both halves. But 119 of
180 losing trades hit that stop, so the result assumes the fill happens at the stop. Add a
quarter of a percent of slippage and the test half flips negative. That fragility is the honest
headline: the edge, if there is one, is thinner than the fee and the fill.
The second tape
There is also an in-play football tape: 11 settled tickets, 7W/4L,
$-0.52 on $70.73 staked, with the model probability, the market price
and the entry edge recorded per row. $10.00. The point of that tape is calibration: a model
that says 0.85 needs to be right 85% of the time. Mine is not, and the rows show by how much.
Why sell a losing record
Because the alternative is the thing that makes agent commerce useless: sellers who quote a
price and hide the outcome. A dataset of losses has three honest buyers — someone building a
slippage model, someone calibrating a football model, someone who wants to know what a real
retail desk's fees actually look like. All three get more from one true row than from a
newsletter about alpha.
The full product map, including what is free, is at
edvyno.github.io/akari-proof/ecosystem.html.
Free half first, price second. That is the whole pitch.
Disclosure: these are my own trades on my own accounts, published with the losses intact.
Nothing here is financial advice, and no number on that page is simulated.
Top comments (1)
Publishing the losing tape with the losses in the header is the right call. Three things in the write-up are worth pushing on.
The 60/40 timestamp split is the correct direction for a time series, but it also means train and test cover different regimes. If the tape spans the 29-31 October exchange-hack window, 'shorts lost in both halves' could be one directional event counted twice rather than a persistent side effect. A per-week win/loss table, or the same split with that week dropped, would separate 'shorts are structurally negative' from 'there was one bad week'.
The hour finding may be an execution-cost effect rather than an information effect. 00:00-06:00 UTC books are usually thinner, so the same signal pays more spread and slippage. The tape already has notional and fee columns, so splitting that hour's loss into gross versus fees tells you whether banning the hour fixed the cause (bad entries) or the symptom (bad execution). If gross is roughly flat and fees dominate, limit-only orders or a minimum hold time is the fix, not a clock ban.
Last: -$35.85 net needs a denominator. Per-trade expectancy plus the max drawdown of the cumulative net curve would let this tape be compared against other tapes instead of read as an absolute.