Understanding Quantitative Confidence Scores
An agreement label describes a comparison between methods. It is not automatically a confidence interval or a verified probability of being correct.
Three different concepts
An event probability describes how likely a specified outcome is under a model. A confidence interval describes an estimator’s uncertainty under a stated statistical procedure. An agreement label describes how closely selected estimates align. Those concepts should not share an unexplained ‘confidence’ label.
Worked example: shared-data agreement
Suppose 20-session realised volatility, 90-session realised volatility and an ATR-derived proxy are 18%, 19% and 20%. Their median is 19%; the maximum-to-minimum ratio is about 1.11. That is evidence of close numerical agreement for those inputs.
It is not three independent confirmations of accuracy. The methods use overlapping observations from the same asset and can all miss a new regime. A future realised volatility of 35% would remain possible despite the close clustering.
Aiovel’s actual labels
The Probability Map distinguishes agreement among volatility inputs from agreement between theoretical touch estimates and descriptive historical frequencies. The historical windows overlap and use the same history. Neither comparison establishes calibrated forecast confidence.
A historical count of 46 hits in 500 overlapping windows is a descriptive 9.2% frequency. It is not 500 independent experiments and does not justify an independent-binomial uncertainty interval without further analysis.
What a stronger validation would require
Freeze a model version, record estimates before outcomes, define the event and horizon, reserve a genuinely forward sample, and report calibration and error by regime. Explain how overlapping outcomes affect uncertainty estimates. Until such a test exists, an agreement score should remain an agreement score.
Compare the label with the dated Gold sample, not with a claim of guaranteed trading performance.
Sources and checks
Definitions checked against the references below on September 17, 2026. Worked examples are illustrative unless explicitly dated. These references do not validate Aiovel forecasts.
Explore the dated public sample and check its source timestamp before using it.
Explore the probability cone →Quick answers
Does High agreement mean a high event probability?
No. Agreement and event probability describe different quantities.
Does shared-data agreement establish calibration?
No. Forward outcome testing and an appropriate uncertainty analysis are needed.