I sat down with my own prediction market book one weekend and asked a question of every open position, one at a time. What single real-world event decides whether this contract pays out? There were about a dozen positions across two venues, and on the surface they looked pleasantly varied. A few election markets, two on rate decisions, one on an inflation print, one on whether a specific bill would pass, a couple of stranger ones. Different questions, different categories, different settlement dates. By the time I finished the exercise, more than half the book traced back to one election. Counted properly I was holding maybe three bets, and one of them was sized far larger than I would ever knowingly make it.
The structure of prediction markets makes this almost inevitable. Every contract is priced and settled on its own. Your venue dashboard shows line items, your spreadsheet shows line items, and nothing in any interface tells you that four of those line items resolve off the same underlying event. A market on who wins an election, a market on chamber control, a market on a policy passing next year, and a market on a cabinet appointment can all be downstream of the same relatively small set of voters in the same handful of places. If the night goes one way, all four settle in your favor. If it goes the other way, all four settle against you within the same hour. On that night your book behaves like one large leveraged position with four different tickers stapled to it.
Why the overlap hides so well
Equity traders have decades of tooling for exactly this problem. Factor models, sector reports, correlation matrices. Nobody needs to be warned that ten oil producers are one bet on crude. Prediction markets have almost none of that scaffolding, and the market structure actively disguises the overlap. Venues file markets under different categories, so your election position sits under politics while your recession-by-year-end position sits under economics, even though the same outcome feeds both. Question wording differs enough that no string match will ever catch it. Settlement dates spread across months, which feels like time diversification but usually is not, because the information that decides all of them arrives on one night or in one ruling.
Chained markets are the sneakiest version. A contract on someone winning a nomination and a contract on the same person winning the general election are mechanically linked, since one is roughly a conditional version of the other. Hold YES on both and the downside stacks in a way no per-contract limit will flag. The macro cluster works the same way. Every individual rate-decision market, every inflation-print market, and most hard-landing markets load on one underlying path of inflation data. Ten of those in a book is one view on inflation expressed ten times, with ten sets of fees.
Map every position to a driver
The fix is boring and takes about twenty minutes. For each open position, write one sentence: this pays out if X happens. Then ask what upstream event or actor actually decides X, and keep pushing until you hit something that cannot be decomposed further, an election night, a central bank reacting to inflation data, a specific court ruling, a named person deciding to run, resign, approve, or sign. That upstream thing is the driver. The test for whether two positions share one is simple. If the driver resolves one way, do both positions move in the same direction? If yes, they are the same bet for risk purposes, whatever the tickers say.
Two rules keep the exercise honest. First, the driver list should be short. If you have twelve positions and eleven drivers, you are describing your book back to yourself rather than grouping it. Most books I have looked at collapse into three to six real drivers. Second, resist building a factor model. Some positions genuinely load on two drivers, and the napkin answer is to assign each one to whichever driver dominates and move on. False precision here costs more time than it saves in risk.
The worksheet
The version I actually run lives in a plain spreadsheet:
- List every open position with venue, side, size, and current price.
- Write the payout sentence for each and name the driver in two or three words. Reuse driver names aggressively.
- Record worst-case loss per position. For binary contracts held to resolution this is simply what you paid, so the column is usually just cost.
- Sum worst-case loss per driver rather than per contract. That sum is your real exposure.
- Divide each driver total by the capital you have allocated to prediction markets. Anything over your cap gets trimmed.
- Rerun the sheet every time you add a position. New positions have a way of landing on your biggest existing driver, because the market that looks most attractively priced is usually the one your existing view already agrees with, and that is exactly how concentration compounds.
On the cap itself, I use something in the region of a quarter of risk capital per driver, and I am not especially attached to the number. The exact figure matters less than the existence of the line. Without one, the driver you feel most confident about will quietly absorb the whole book, since conviction is precisely the feeling that makes each additional correlated position seem reasonable at the time.
The hedge that fails in the middle
One failure mode deserves its own warning. When a driver runs over the cap, it is tempting to buy NO on a related market instead of trimming. Sometimes that works. Often the resolution criteria differ in ways that only bite in the middle scenarios. One market settles on the official result while the other settles on a call from a specific source by a specific date. One asks about the overall outcome, the other about a margin. One has a deadline the other lacks. In the clean scenarios the hedge behaves. In the messy ones, a delayed result, a contested count, a data revision, both legs can lose at once, and the messy scenarios are exactly what you were trying to hedge. Read the resolution text of both markets side by side before counting anything as a hedge, and if the words differ, assume it is partial at best.
Timing is the other argument for trimming early rather than hedging late. Liquidity in event markets thins and spreads widen as resolution approaches, so cutting driver exposure in the final week costs real money. The worksheet is cheap months out and expensive near the end.
We watch a lot of public prediction market activity at Blockcircle, and the same shape appears over and over: an account spread across fifteen or twenty markets that is, in driver terms, running one heavily leveraged election bet or one oversized rates bet. The spread of tickers makes a book look sophisticated from the outside. The driver map tells you whether it actually is. Run one on your own positions before the next big resolution date, while there is still time to act on whatever it shows.