Forecast Track Record

Brier context

35/50 terminal rows; effective independent sample: 8 theses. No public accuracy claim is made below the gate.

• Forecast mean Brier: 0.1428.
• Constant 50/50 portfolio baseline: mean Brier 0.2429; Brier Skill Score 0.4118.
• In-sample frequency baseline (p=0.9): mean Brier 0.0829; Brier Skill Score -0.724.

The in-sample frequency baseline uses this sample's own outcomes. It is a deliberately hard comparison, not an out-of-sample climatology. Forecast-weighted mean Brier is 0.1428; thesis-weighted mean Brier is 0.1428.

Контекст: 35/50 закрытых прогнозов, эффективная выборка — 8 независимых тезисов. Частотная baseline использует исходы этой же выборки и не является внешней исторической частотой.

How to read this

• Probability. Every forecast is registered in advance with a numeric probability: 0.70 means the system gave the outcome a seven-in-ten chance. The number is locked before the outcome and never edited.
• Terminal review. On the check date the forecast is scored once against its pre-registered failure condition. That score is permanent.
• Hit / partial / miss. Hit: the claim held. Miss: the failure condition fired. Partial: the outcome landed in between — it is used rarely and shown like any other result.
• Brier score. The squared distance between the stated probability and the scored outcome: 0.0000 is perfect and 1.0000 is the worst possible. For one binary hit/miss scored at p=0.5 the score is 0.2500; that is not automatically the portfolio baseline when partial outcomes are scored as 0.5. Lower is better.
• Resolution. Each claim states in advance what evidence counts and what would falsify it, so scoring is mechanical, not retrospective judgement.

Кратко: вероятность фиксируется заранее; прогноз получает исход hit / partial / miss по заранее записанному условию. Brier — квадрат расстояния между вероятностью и оценённым исходом. Для одного бинарного hit/miss при p=0.5 результат равен 0.25, но portfolio baseline меняется, если partial оценивается как 0.5.