A real-looking edge
Prediction markets now list contracts like "Will Bitcoin be above $X at 5pm today?" Each one pays a dollar if yes, nothing if no, so its price is the crowd's probability.
The options market prices exactly the same question, and it does so with a far larger crowd, far more money, and professional volatility traders on both sides. So we wrote down a simple test before looking at any results: take the probability implied by Bitcoin and Ether option volatility, compare it with what the prediction market charges, and buy whichever side is at least five cents too cheap.
On the first exchange it worked. On data it had never seen, the cheap side paid about seven cents per dollar. The edge sat almost entirely on one side: people overpay for "yes, Bitcoin will be above X", the way people overpay for lottery tickets. It was the third confirmation in our research ledger, out of more than a hundred pre-registered tests.
The exchange we could not use
The first exchange is not legally tradable from Argentina, and routing through a server abroad would break its terms. We don't do that. The finding stayed on paper, and a few days later we retired the paper monitor too: a measurement nobody can act on is not worth running.
So we went to the regulated competitor, which is open to us and lists the same contracts. We repeated the test on its own history: 9.6 million trades over two and a half years. This time we were stricter about fills. We used only prices a buyer actually paid on that side, and charged the exchange's real fee.
It confirmed again. About six and a half cents per dollar in the exploration years, two and a half cents on the held-out period. Both coins, both sides, near the money and far from it.
Two independent exchanges, two confirmations. That is usually where people stop checking.
The ruler was bent
The test priced each contract once an hour. For the market's price it used the last trade in the past hour. For the model's price it used where Bitcoin was at the end of that hour.
Read that again slowly, because it took us a while. The last trade could be fifty-nine minutes old. In fifty-nine minutes Bitcoin moves. The model knew where it had moved to; the old trade price did not. Compare a fresh number with a stale one and the stale one will look "wrong" on a schedule, in the direction the market went. That isn't a mispricing. That is a clock.
So we repriced every one of the 9.6 million trades against Bitcoin at the minute of that trade, so neither side knew anything the other didn't:
| exploration | held out | |
|---|---|---|
| As first measured | +6.5¢ | +2.4¢ |
| Model priced at the trade's own minute | +2.0¢ | +2.3¢ |
| Buying at the next trade instead | −1.6¢ | −0.8¢ |
About two-thirds of the confirmed edge was the measuring stick. What remains, about two cents, exists only at the instant someone else's trade prints. Wait for the next trade, a median five and a half minutes later, and the price has moved two cents against you. The edge becomes a loss.
What is left, and what we're doing with it
Two cents that vanish in minutes is not nothing, but it is a race, and races are won by whoever is fastest. Whether a small bot that reads the order book every ten seconds can win it is a question the trade history cannot answer. It only shows the trade that happened, not what else was on offer.
So we built exactly that, on paper: a poller that re-reads the live book every ten seconds, prices every strike, and simulates a fill only against what is really there. It does not trade.
Its first full day: nine moments where a quote looked five cents cheap. All nine were gone by the time the book was read again, with one dollar in total available at the good price. One day proves nothing either way. The verdict date is written down: December 22nd, after at least 150 settled paper trades.
If it works, it is a small strategy: a couple of thousand dollars of capital for a couple of thousand a year. We have decided that small is fine. A living built from strategies is a sum, and a capital-light sleeve that returns its own size is worth more than a big one that returns four percent.
The lesson that goes on the checklist
Every test that compares a market price with a model now prices the model at the moment of the trade, never at the end of a bar. And every result gets a second line: what if you were one trade late? An edge that only exists at the exact instant of someone else's fill is not yours yet.
The first exchange's number almost certainly had the same flaw. We can't trade there, so we won't spend another day finding out.