Methodology

Every prediction on YearlyFuture is based on a real institutional report, peer-reviewed research or an official data source. This page explains how probabilities are assigned and updated.

1. How Predictions Are Built

• The topic is researched and only directly relevant sources are selected.
• Each source is verified to be real and reachable (dead links are not counted as sources).
• Videos, social media posts and marketing pages are never used as sources.
• The condition under which a prediction counts as fulfilled is written in advance.

2. How Probabilities Are Assigned

Probabilities combine the strength of the evidence, the time remaining until the target date and historical base rates for similar claims. Values are kept between 1% and 99%; no prediction is given 0% or 100%. 50% is used only when the evidence is genuinely balanced, never merely to avoid a decision.

3. Updates

Probabilities are not static. As new developments are reviewed, values move up or down, and every change is logged with its date. The "last probability update" shown on each prediction page is the output of this process.

4. Transparency and Limits

• AI assistance is used in the analysis; every prediction is published together with its sources and reasoning.
• A probability expresses uncertainty, not a guarantee.
• Nothing here is investment, legal or medical advice.
• For corrections and feedback: Contact.

5. Worked Example (Based On A Real Prediction)

The steps below use the real data of a prediction that is currently published on the site; the result matches the probability shown on that prediction's page exactly. No number has been produced after the fact.

• Prediction: By the end of 2027 a leading AI lab will announce that it has reached at least 50% accuracy on the hardest tier (Tier 4) of the FrontierMath benchmark.
• 1) Base rate: Accuracy on this tier was below 2% in 2024 and reached roughly 20-25% by early 2026. The historical base rate is therefore not high and keeps the starting value low.
• 2) Evidence direction (upward): The official evaluation run by Epoch AI and two independent arXiv papers verify that the evidence is real and reachable. The evidence is directly on topic, so the value is adjusted upward.
• 3) Time remaining (downward): The target date is the end of 2027, so the time left is short. Moving from the 20-25% band to 50% in two years requires an aggressive leap, so the adjustment is capped downward.
• 4) Result: Combining the strength of the evidence, the time remaining until the target date and historical base rates for similar claims, the probability was set to 30% with a Low confidence level. That is the value shown on the prediction page; it is updated with its date as sources change.

The Record line on that prediction page shows the sha256 digest of the content as published, so it can be verified that the text was not changed afterwards.