Oracles and prediction engines: what BrainPredict borrows
Prediction markets often mix two jobs. The oracle settles after the fact, when the contract resolves under the Rules. The prediction engine estimates the probability beforehand, from facts already on file. That gap is the distortion: price sometimes quotes emotion or wordplay, not the real chance.
Oracle vs prediction engine
An on-chain oracle states whether the event written in the contract happened. UMA Optimistic Oracle treats the report as innocent until someone disputes it. Token holders then vote (DVM). On Polymarket, that path closes the contract.
Chainlink aggregates several sources and keeps the median. That is a structured data feed, not a second bookmaker. The engine must publish a probability before the market has the last word.
BrainPredict publishes that estimate blind to the price, then reads the gap. Buying or selling the contract names a pricing error, not a political camp. Metaculus and Tetlock remain an intellectual bar, not a live fair-value feed.
What we borrow
Silver
Nate Silver never enters a raw poll. Each institute has a weight, then ageing over time (W × decay), already in our Ifop, Datafolha, Quaest weighting. Axios, 16 July 2024, cites him as a Polymarket adviser since June 2024. Adviser, not « principal ».
Tetlock
Philip Tetlock (Good Judgment Project) sets the bar: stay calibrated, update in small steps, measure honesty (Brier score). The public figure stays locked until 15 public files are closed. We do not print an invented rate.
UMA / Polymarket
UMA and Polymarket resolve the contract text. The Resolution Rules say what counts as a strike, talks or a deal. The page title is not the contract. Ground gained is not Yes if the Rules say no. That is the semantic trap, already seen on our geopolitical training markets.
Comparative matrix
| Layer | Chainlink | UMA OO + DVM | Aggregators | BrainPredict |
|---|---|---|---|---|
| Input | Multi-source, median | Rules wording | Polls, models (Silver, Tetlock, Metaculus) | Facts already on file, no language model |
| Truth | Structured feed | Innocent until disputed | Calibration over time | Fair first, price after |
| Out of scope | Not a parallel book | Artillery if Rules = air | No stored LLM votes = not run | Large wallets or bookmakers without a live feed |
| Silver / Metaculus / Pinnacle | not a score | not a score | unavailable (no stored figure) | unavailable (no public resolution count yet) |
What BrainPredict publishes
We publish the factual estimate first, without looking at the price, then the gap. File, blind estimate, then the quote, then a buy or a sell only if the gap is worth it. The football prior (league draw share) ignores the price. We do not invent a chance statistic.
We do not publish a public win rate: we do not yet publish a counted public hit rate. The indicators stay locked under 15 closed files. Picking few bets, and being able to exit before the contract ends, is not a crystal ball.
Our geopolitical training losses often come from the same hole. The ground moved, the Rules did not. UMA does not apply a headline: it applies the contract text. Bad faith or a grey zone is contract risk, not a plot. A very low outsider price is not, by itself, a signal.
There is no oracle memo PDF to download. The public predictions page is available.