Forecast Track Record
A probability is only worth reading if the person publishing it keeps score. Every closed question appears below with the number that was published before resolution, scored the same way professional forecasters are scored.
Calibration by probability band
The marker on each bar is the band midpoint. A well calibrated board has the filled bar landing close to its marker as the sample grows.
Every closed question
Will SpaceX complete an IPO in the first half of 2026?
Priced within the indicated range; the model correctly weighted the filing timetable over sentiment.
Will a concentrated AI-thesis fund disclose a forced unwind before September 2026?
Liquidity-terms mismatch was the identified driver and was the proximate cause.
Will the FOMC raise the target range in the first half of 2026?
Correctly low, though the board was slower than markets to discount the hike scenario.
Will XRP outperform Bitcoin in the second quarter of 2026?
Depth and distribution drivers held; the magnitude of the gap exceeded the model.
Will tokenized Treasury products pass $30B outstanding by mid-2026?
Custody approvals were the correct leading indicator rather than front-end yields.
Will a top-five hyperscaler cut AI capex guidance in the first half of 2026?
Miss. The board underweighted how much lease-based financing defers reported capex pressure.
Will US foreclosure starts rise more than 20% year over year by mid-2026?
Loss-mitigation programmes deferred starts further than the lag assumption allowed.
Will a US regional bank fail in the first half of 2026?
Correctly low. Funding stress showed up in margin compression rather than resolution.
How scoring works
What is a Brier score?
The squared difference between the forecast probability and the outcome, averaged across all closed questions. Zero is perfect, 0.25 is what you get by forecasting 50% on everything, and anything above 0.25 is worse than that baseline.
Why publish the misses?
A forecast record that only shows wins is marketing. Calibration is only meaningful when the denominator is complete, so every closed question appears here with the probability that was published at the time.
What does well calibrated mean?
Questions forecast at 70% should resolve YES about seventy per cent of the time. The calibration table compares each probability band with the observed frequency, which is a stricter test than a simple hit rate.