AI and Mispriced Probabilities

Liquidity is thin, participants are mostly retail, and flow follows headlines. A contract can sit at a stale price for days because nobody with a view is looking at it, then move sharply on a story that a careful reader would have discounted. The same outcome can trade at different prices on two venues because their users differ, not because the outcome does. And contracts on scheduled data, where the distribution of outcomes is well understood, are often priced with less rigor than the data deserves. Each of these is a mispricing that a systematic reader can find before a discretionary one happens to.

Language models read the information that bears on each contract: news, filings, procedural records, polling, forecast distributions for scheduled data, and the state of code and governance for protocol events. They extract what matters, weigh it against base rates for that category of event, and produce a probability estimate with its reasoning written out. Statistical models calibrate those estimates against the historical record so that a stated probability means what it says. A second set of models reads resolution rules, the text that decides what counts as the outcome occurring, and matches contracts across venues that are truly equivalent while flagging the ones that only look it.

The estimate is compared with the market’s implied probability on every venue listing the outcome, net of fees, bid-ask spread and the time value of capital locked until resolution. Where the model’s probability is far enough from the price, there is positive expected value: a positive delta between what the contract should cost and what it does. The required size of that delta scales with the model’s confidence, with the record of that category’s calibration, and with the size of the loss if the contract resolves the other way. Small gaps on binary outcomes are noise and are not traded.

When two venues price the same outcome differently by more than the cost of trading both, buying the cheap side and selling the rich one locks in the difference whatever happens. The models make this workable at scale by comparing resolution rules, settlement sources and timing across venues before a pair is treated as one trade, and by netting the arbitrage against withdrawal times, capital lock-up and the fee on each leg. A pair whose rules differ in any material way is not an arbitrage, and the model’s job is to say so.

Models propose; they do not trade. Every candidate arrives with its probability, its reasoning and its sources, and a person reviews it before a position is sized. Model output is logged and scored against outcomes, and a category whose estimates stop calibrating is suspended until the reason is understood. The models use public information only, and the same eligibility test that applies to the rest of the mandate applies here: no contract is traded where the firm could hold anything that is not public.

The systems run on infrastructure the firm controls, using open-weight models where the workload allows and hosted models only for tasks where nothing sensitive leaves the building. Estimates, reasoning and the record of outcomes are the firm’s own data and stay inside it. This is the same approach Wesvel’s AI Systems practice builds for clients, applied to the firm’s own book first.