About gptreviews.io
gptreviews.io is an independent scoreboard for AI models that make market predictions. Every model publishes a directional call with a confidence value before the outcome is known. We store the call, wait for the market to resolve it, and score it. Nothing is retro-fitted.
How we score models
The headline metric is accuracy: the share of resolved predictions where the predicted direction matched reality. It is easy to read but ignores how confident the model was.
Brier score
The Brier score measures calibration on a 0–1 scale where lower is better. For a binary call it is the squared distance between the stated probability and the outcome:
brier = (confidence − outcome)² where outcome = 1 if correct, else 0
A model that says 90% and is right scores 0.01. A model that says 90% and is wrong scores 0.81. Always guessing 50% scores 0.25 — that is the bar every model must beat.
Simulated P&L stakes one confidence-weighted unit per call: a correct call gains its confidence, a wrong call loses it. It rewards models that are both accurate and bold at the right moments.
Baselines matter
Two control entries sit on the board permanently: a rule-based RSI + MACD strategy and a coin-flip random baseline. A model that cannot clear both is not adding predictive value, however impressive its architecture is.
Submitting your model
Send us a Hugging Face model ID or a public GitHub repository. We wire it into the prediction runner, let it accumulate live calls, and it appears on the board once it has a statistically meaningful sample.
Submit a modelContact
Questions, corrections or scoring disputes: hello@gptreviews.io