An event probability is only one part of a price forecast

The probability that an event occurs differs from the probability that an asset rises after it. We propose evaluating those forecasts separately and combining them only through a declared conditional model. The design remains unrun. It is useful whenever an agent receives event odds and must decide what those odds imply for a price horizon.

The continuous record companion preserves a historical directional-edge null. That result gives a reason to test the mapping from information to direction rather than assume a richer signal creates predictive skill. The controls paper companion supplies the trace discipline to retain which source supported each forecast.

One event, two possible responses

Take an illustrative event probability of 0.70. Suppose the proposed conditional model assigns a 0.40 chance of a positive one-hour return if the event occurs and a 0.60 chance if it does not. The implied probability of a positive return is 0.70 × 0.40 + 0.30 × 0.60 = 0.46. Treating the event's 70% chance as a 70% chance of price appreciation would reverse the interpretation in this example.

The conditional estimates could differ because traders already anticipate the event, because its details matter or because other news arrives. These are candidate explanations to investigate, rather than facts established by the arithmetic. Specify the event resolution rule and the price horizon before collecting outcomes. A binary occurrence label may conceal severity, timing or revisions that actually drive the market response.

Save three forecasts at the same information cutoff: event occurrence, positive price response conditional on occurrence and positive price response conditional on nonoccurrence. Include the resulting marginal price forecast. Score occurrence after event resolution and score the marginal price forecast after the horizon ends. Conditional forecast evaluation uses only the corresponding resolved branch, with branch sample sizes disclosed.

Test whether decomposition helps

Compare the two-stage method with a direct price forecast and a historical base-rate forecast. Use the identical event packet and price snapshot. A more elaborate explanation should earn its cost through a better proper forecast score or a clearer diagnosis of error. Compare fixed-policy decisions separately, because two methods with different probabilities may still fall on the same side of an action threshold.

We would retain event-source disagreements and ambiguous resolutions as explicit statuses. Resolving them after observing the price move would contaminate both forecast tasks. Where conditional branches are sparse, report wide uncertainty and avoid pooling unrelated event types simply to improve precision.

The contribution is a decomposition that exposes where a signal fails. Accurate event odds with poor price forecasts implicate the response model, timing or market expectation. Poor event forecasts implicate the event channel itself. That distinction guides the next experiment without turning a prediction-market number into a trading recommendation.