THE FUTURE IS PROBABILISTIC. EXPLORE THE POSSIBILITIES.Experimental platform

METHODOLOGY NOTE

Confidence needs a track record.

Why one correct prediction is not enough to establish the quality of a forecasting model.

A single outcome has limited information

An event assigned a small probability can still happen. A model can get one result wrong while expressing reasonable uncertainty. Conversely, an overconfident model may appear correct by chance.

Evaluate the distribution

Proper scoring rules evaluate probabilities rather than only the most likely label. Our demo uses the multiclass Brier score and log loss. Their definitions, ranges, and display transformations are documented in the methodology.

Check calibration across many events

For similar, comparable events, forecasts around 70% should resolve positively roughly seven times in ten over a sufficiently large sample. Small samples and changing event populations make that relationship hard to estimate.

Keep evaluation independent

Record forecasts before close, resolve with external evidence, and keep model versions visible. The 12 synthetic matches in this release demonstrate the interface, not the performance of any real model.

Sources and scope

This note describes our product design and sample framework; it reports no empirical findings. Read the full methodology for definitions, assumptions, and primary metric references. All numerical examples are illustrative.