The prediction market trades I keep coming back to are the boring ones, where two contracts about the same underlying event imply probabilities that cannot both be true at once, and you buy the cheap side of the contradiction. A market on a candidate winning both the nomination and the general election trading a couple of cents above the market on the nomination alone. A contract on something happening by March priced above the identical event by June. These show up more often than you would expect, and the interesting question is how much of the gap is actually yours to take.
The arithmetic markets have to obey
Any set of related markets has to respect a few inequalities whatever anyone believes about the world. A joint event can never be more likely than either of its components, so the YES price on A-and-B should never exceed the YES on A alone. A subset can never be more likely than its superset, which in practice usually means deadline chains: by March is contained inside by June, so the earlier deadline has to trade at or below the later one. A set of mutually exclusive outcomes can sum to at most 1, and if the set is also exhaustive it has to sum to about 1, give or take fees.
There is a floor as well as a ceiling. The joint event cannot be less likely than the probability of A plus the probability of B minus 1, so if two things are each priced at 90 cents, a market saying both happen cannot rationally sit at 60. These bounds hold no matter what your model of the underlying event is. You do not need an opinion on the election or the protocol upgrade. You only need the prices to disagree with each other.
When they do, the constructions are mechanical. Joint priced above a component: buy YES on the component, buy NO on the joint, and you collect in every state of the world. Mutually exclusive outcomes summing well over 1: buy NO across the set, and since at most one outcome can happen, all but one of your NOs pay out. An exhaustive set summing well under 1: buy every YES and one of them has to pay. On paper these are locks, which is exactly why the ones that survive on a live order book deserve suspicion.
Screening for violations
Doing this by eye across platforms does not scale, but the screen itself is simple. Pull every active market, convert prices into implied probabilities, and group anything that references the same entity and event. Titles and tags get you most of the way there, the same election, the same company, the same deadline family. Then apply the inequalities with a buffer. I would not flag anything under three or four cents of gross edge, because fees, spread, and slippage will eat a violation that only exists at the midpoint. And use executable prices, the ask on whatever side you would actually buy, not last trade and not the mid, since thin books generate phantom violations constantly.
Deadline ladders are the most productive place to look. Platforms list the same milestone across several dates, the legs are tended by different market makers with different attention spans, and the far-dated ones go stale. Cross-platform pairs on the same event are the second place, though those come with their own problems, which is most of the rest of this post.
Why the apparent free money usually is not
Start with resolution criteria, because this is where most of these trades die. Two markets that look like the same event are often subtly different bets. One resolves on an official announcement, the other on credible reporting. One uses UTC deadlines, the other local time. One counts an interim appointment, the other requires a confirmed one. If your joint-versus-component trade actually spans two slightly different definitions of the event, the inequality does not bind and you can lose both legs. I read both rulebooks word by word before sizing anything, and if the resolution sources differ at all I treat the position as correlated relative value rather than arbitrage and size it like a bet that can go wrong.
Then there is legging risk. A coherence trade needs two or more fills, and the books on mispriced legs are usually thin, which is often why the mispricing exists in the first place. You lift one leg, the other side moves or disappears, and you are left holding a directional position in a market you never had a view on. The old discipline applies here: fill the thinner book first, use limits, and size each leg so being stuck with half the structure is annoying rather than dangerous.
Capital cost is the quieter killer. Your money is locked until resolution, and a four-cent edge that resolves in nine months is a mediocre annualized return after fees, before you even count platform risk. Some venues also let anyone redeem a complete bundle of mutually exclusive outcomes instantly, which is why those sets rarely drift far above 1 for long. The gaps you can actually capture tend to live in the harder cases, cross-platform or cross-family, where no redemption mechanism exists and resolution risk is real. And resolution itself can be contested. Outcomes have historically been disputed and occasionally flipped after the event looked settled, which converts a locked profit into a coin flip at the worst possible moment.
A workflow that keeps you honest
- Confirm the violation on executable prices at your intended size, not on midpoints or last trades.
- Put the resolution rulebooks side by side and hunt specifically for differences in source, deadline, timezone, and edge cases. Any difference downgrades the trade from lock to relative value.
- Compute net edge after fees on every leg, then annualize it against the latest plausible resolution date. If the annualized figure would embarrass a savings account, pass.
- Check depth on all legs and plan the fill order, thinnest first, with a written answer to what you will do if a leg breaks.
- Size so the worst leg-break leaves a directional position you are willing to hold or exit at a known cost.
- Keep watching after you are filled. Rule clarifications and disputes move supposedly locked trades, and exiting early at a smaller profit often beats holding through resolution drama.
Most of what the screen surfaces will fail on the rulebooks or the annualized math, and that is fine, because the gaps are useful even when they are not tradable. An inverted deadline ladder tells you where the stale attention is. A cross-platform gap that refuses to close usually means the two venues define the event differently in some way you have not spotted yet, which is worth knowing before you trade either market for any reason. We pull prediction market prices into Blockcircle alongside our other signal feeds partly for this kind of relative read, and the incoherent pairs are reliably more informative than the well-arbed ones.
If you want a starting point, pick one entity you already follow, list every market that references it across the venues you use, and hand-check the inequalities once. The first pass takes an hour and will probably find nothing you can trade. It will also teach you exactly how each platform defines its events, and that knowledge is what eventually lets you tell a genuine pricing error from a bet you did not know you were making.