Oracle risk: when both legs of a hedge lose

A cross-venue hedge is supposed to be indifferent to the outcome — one leg pays $1.00 whichever way the event goes. In our own settled record, 127 of 426 positions (29.8%) had both legs settle the same way, which a complementary pair mathematically cannot do. Most of the cause was not bad matching. It was two venues reading two different instruments.

What is oracle risk in prediction markets?

Oracle risk is the chance that two venues resolve the same question from different sources, and therefore reach different answers. The markets are identical in wording, scope and deadline; the resolution source is not. Because the disagreement lives outside the text of the market, no amount of better question-matching detects it.

The clearest example is weather. Two venues both list "will the high temperature in this city exceed 90°F?" and each names its own weather station. One station reads 90.4°F and the other 89.8°F. Both YES and NO legs settle the same way, and a position that could not lose on paper loses in fact.

How often does a cross-venue hedge fail this way?

We classified all 127 non-complementary settlements by whether the two legs named the same subject:

• 97 named the same subject — correctly matched pairs, similarity around 0.80, that still settled together. These are oracle failures, not matching failures.
• 23 named different subjects — genuine false matches, where two markets read alike and were not the same event.
• 7 were too ambiguous to classify either way.

The split matters because the two groups need opposite responses. False matches are a bug to fix. Oracle divergence is a risk to price — the matcher was right, and tightening it further would only throw away real hedges.

Which markets carry the most oracle risk?

Anything resolved by an instrument or an index rather than by an unambiguous public event. In our settled data the failure rate by category was 24% for market and weather questions, 9% for sports and 6% for entertainment. Sports and awards resolve on a result the whole world observes identically; temperature, rainfall and index levels resolve on a reading, and readings vary by source, rounding rule and timestamp.

Before committing to any cross-venue pair, read both resolution rules end to end and confirm they name the same source. If either venue leaves the source unstated, treat the pair as carrying oracle risk.

Can better matching solve oracle risk?

No, and that is the whole point. A similarity score reads the question text; oracle risk lives in the settlement rules. The pair really is the same question, so every text-based signal correctly says "match". We handle it by pricing it instead: a cross-venue pair on a weather-resolved question is automatically graded high match-risk, which removes it from the recommended list while leaving it visible on the board with its warning attached.

The false-match half is fixable, and each identified cause is now blocked by a guard — including a location guard, after two markets naming different cities scored 0.65 on similarity because the city was one token in twelve, and an opponent guard, after pairs like "Will this team win?" against "A vs B — the other side" turned out to be one directional bet placed twice rather than a hedge.

What does the record look like once the class is removed?

Scoped to the positions the board actually recommended, 87 have settled, 74 paid out as designed, averaging +2.24% each, or +$1,657 per $1,000 committed. The 13 that did not behave as hedges are named on the public results page rather than deleted from it.

That distinction is worth stating plainly, because it cuts both ways. A correctly matched hedge cannot lose — exactly one leg pays — so a track record led by losses would misdescribe the instrument. But a record with the failures removed would misdescribe the engineering. Both belong in view.

Why publishing the wins alone would be dishonest

When both legs of a mispaired position settle YES, the position pays twice. Our two largest single gains on record came from exactly the mechanism that produced our largest losses — the same defect, a different coin flip. Presenting those as arbitrage wins while burying the losses would mean reporting the lucky half of a loss-making class. That is why the failures are published alongside the returns.

Keep reading: Prediction market arbitrage: the complete guide · Polymarket vs Kalshi: fees, access & price gaps · Polymarket arbitrage: finding mispriced markets · Liquidity: why the quoted edge is not the real one