Published forecasts

Any forecast you run can be published to a permanent public URL. These are ours. They are the same output you get from forecast(), the MCP tool, or the app, so they are the fastest way to see what a FutureSearch forecast actually looks like before you run one.

Each page shows the answer, the full rationale, and every source the research agents read. Nothing is redacted except the input data of the person who ran it.

Publish your own

Publishing happens in the web app. There is no publish method in the Python SDK, the public API, or the MCP server, so if you ran the forecast from code, open the session in the app to publish the result.

Open the finished forecast in the app. The control only appears once the task has completed.
Click the globe icon in the results toolbar. The popover is headed Publish results.
Type a slug. It becomes the URL, so pick something readable. Slugs are 3 to 80 characters, lowercase letters, digits and hyphens, and globally unique, so a common word may already be taken.
Publish, then copy the URL from the field that replaces the input.

Unpublishing is one click in the same popover, and it frees the slug for reuse.

A good slug names the subject and the resolution point, which is what someone searching the question would type: anthropic-ipo-pricing-date, spacex-positive-free-cash-flow-by-2027, gemini-3-5-pro-context-window-size. Avoid dates of publication, run numbers, and internal project names.

The published page is the full result, not a screenshot of it, plus a generated preview card sized for social embeds, a canonical URL, and article structured data.

What stays private

Publishing exposes the result and its reasoning. It does not expose your input data. The published bundle carries only the parameters needed to render the result, such as forecast_type, output_field and units, and internal fields including trace mappings and analysis metadata are stripped from the artifact and all its children.

The question text is part of the page. If you published a forecast whose question contains something confidential, unpublish it.

Published forecasts feed the world model

Publishing adds the forecast to the shared world model, the set of drivers reconciled across every forecast the system has made. A published forecast can then help calibrate other people's answers on related questions, and theirs help calibrate yours.

The effect is visible on the page. Published forecasts show what the answer was weighed against, phrased as a note like "weighed against 2 related questions here and 1 public forecast", which expands to show the siblings and the public forecasts consulted. Your own past private forecasts are never shown to visitors, even where they informed the result.

Binary: will it happen?

Binary forecasts return a probability from 0 to 100 and a rationale. See forecast() reference and the Turn Claude into an Accurate Forecaster guide.

Date: when will it happen?

Date forecasts return a p10 through p90 spread of dates rather than a single guess, which is usually the honest answer for timing questions.

Numeric: how much?

Numeric forecasts return percentiles with units. Compare the p10 and p90 to see how much the agents actually know.

Categorical and thresholded: which one, and how far?

Categorical forecasts spread probability across mutually exclusive outcomes. Thresholded forecasts give one probability per threshold on the same underlying quantity. Both are covered in Forecast Categorical and Threshold Questions.

Conditional: what if?

Conditional forecasts run the outcome in the world where a condition holds and the world where it does not, jointly, so the two branches stay coherent. See Forecast Conditional Scenarios.

This set forecasts the same date under three different assumed causes, next to the unconditional forecast, which shows how much the answer hinges on the cause:

The full list of everything we have published is in the app sitemap.