What a 70% prediction-market probability actually means

Why these numbers need a reading guide

Prediction markets have attracted regulatory attention in 2026, including the CFTC's March 2026 advance notice on prediction markets. A consultation is not a final rule, and rules vary by jurisdiction. The useful educational question here is how to interpret a probability appearing in a headline, rather than which platform to use.

Start by refusing to round a probability into certainty. A forecast of 70% leaves substantial room for the event not to happen. If the event fails to occur once, that does not by itself show that the forecast was foolish. To evaluate a forecasting method, you need many recorded forecasts and their outcomes.

Read the event definition before the price

What exactly must happen, by what deadline, according to which source, for the contract to resolve? 'A policy will be announced' is different from 'a policy will take effect'. 'A candidate will win a nomination' is different from 'a candidate will hold office'. A screenshot can hide those distinctions while retaining the dramatic percentage.

Also identify the quote. A last trade, an executable bid, an executable ask, and a displayed midpoint are not identical. Thin activity or a wide bid–ask spread can make one price a poor summary of what people can actually transact at. Treat the number as dated information from a particular mechanism, not as a timeless probability handed down by the crowd.

When many participants help—and when they do not

The wisdom of crowds experiment lets you inspect what happens when independent errors average out, and what happens when estimates share a bias. Markets can aggregate information, but participants may rely on the same report, face similar incentives, or have unequal resources. Participation is not a guarantee of diverse evidence.

Compare that with information cascades. People may infer that earlier participants knew something and follow their action. The model does not reproduce a real trading venue; it helps identify a question worth asking: is the new movement based on new information, or on other people reacting to the same old information?

A worked example: a project launch forecast

Imagine an illustrative internal forecast that a project will launch before Friday has risen from 40% to 70%. This is not a live market quote. The interpretation changes depending on why the estimate moved. A signed-off dependency is different evidence from a dozen colleagues repeating an optimistic comment.

1. Define launch precisely: available to which users, with which required functions, by what time zone and deadline?
2. Identify the new evidence and whether it is independent. Use Bayesian updating to see how the diagnostic strength of evidence matters.
3. Keep a contingency for delay. Decide whether preparing it costs less than being unprepared, rather than assuming 70% makes preparation unnecessary.

Check calibration across forecasts

A well-calibrated collection of 70% forecasts should resolve positively about 70% of the time over an appropriate collection of comparable cases. That is a long-run property, not something one event can reveal. You also need enough observations and an evaluation set chosen before inspecting which examples make the forecaster look good.

Brier's original paper on probability forecasts provides a foundation for scoring probabilistic predictions. For a simple binary event, the commonly used score is the squared difference between the probability and the outcome coded as zero or one. Smaller is better. Evaluation should also compare against a sensible baseline, not only against another confident commentator.

Convert a forecast into a decision carefully

The same probability can imply different actions for different people. If a delay would be mildly inconvenient, an expensive backup may not be worthwhile. If a delay would interrupt a critical service, the same forecast could support preparing a backup. Probability is one input; consequences and the available alternatives are other inputs.

Explore value of information before chasing every update. Ask which missing fact could actually change your decision. A small movement from 70% to 72% may be irrelevant if your action remains the same. More decimal places can create the appearance of useful precision without supplying decision-relevant knowledge.

• Read the event, deadline, and resolution rule.
• Identify the quote and the observation time.
• Ask what independent information changed.
• Separate forecast evaluation from your own decision costs.

A useful number can still be an imperfect number

Fees, risk preferences, market access, liquidity, and contract design can complicate a direct price-to-probability interpretation. Treat a market estimate as one source to compare with other forecasts and evidence. Do not average several sources as though they are independent when all of them are copying the same signal.

You can practise the skill with ordinary forecasts: delivery dates, exam preparation, or whether a meeting will finish on time. Record a probability before the outcome, define the event, and review a collection later. The learning comes from making uncertainty explicit and being accountable to evidence, not from finding the most dramatic screenshot.