Many markets in, one reading out.
Every index follows the same four steps, and each published value can be traced back to the prices behind it.
An example contract: the traded price is the chance the market is pricing in.
AI finds the markets relevant to each index, using criteria set in advance.
Is it about the theme? Does a higher price move the index up or down? How much should it count?
More money behind a market, more weight. Seat counts are the exception: one seat is one seat.
Every value traces back to the markets, the prices at that hour, and how each was classified.
More exchanges, better forecasts
Each exchange is a crowd putting money behind its best guess. A well-weighted blend of exchange prices beats any single exchange most of the time.
simulated scenarios where the blend beat the best single exchange
exchanges outperformed on 3,000 questions, 2.4% lower error than the strongest
Each exchange misses in its own way. Blended, the misses partly cancel (Bates and Granger, 1969).
Exchanges that called past questions well count for more; consistently biased ones count for less.
Weights lean toward an even split, so a short track record can’t chase noise.
Results come from a simulation of five exchanges of similar quality, not from Micah’s index data, scored by Brier score. Today’s published indexes weight by money (step 03); weighting by track record is where the research points next. It can’t remove an error every exchange shares, such as an ambiguous resolution rule.
The full paper, with the derivation, simulation design and a worked example.
Request the paper