Hirbod Assa (UCD)
will speak on
Risk significance versus statistical significance: an epistemology of quantitative risk
Time: 3:00PM
Date: Thu 19th November 2026
Location: N0.20 - Science North
[map]
Abstract: Statistics is a discipline with an epistemology of its own: it takes data as its object, reasons about sampling uncertainty, and issues a verdict, statistical significance, that licenses an empirical claim. Quantitative risk management has inherited this verdict wholesale, yet the uncertainty that defeats risk governance is not sampling error but challenge uncertainty: whether a conclusion survives the rival models, calibration windows and stress scenarios a competent reviewer could demand it withstand. This talk develops a modal-epistemic framework in which a risk conclusion is evaluated across a population of admissible “possible worlds” under an institutional evidence relation, and is risk-significant when it is endorsed, when it survives every admissible challenge, rather than merely when a single-model test rejects a null. The framework supplies two registers (assurance-grade K and precautionary working-commitment B), a vocabulary of governance diagnostics built around the Moorean condition “the risk obtains but we cannot endorse that it does,” an importance measure (pivotality) that localises fragility, and a collapse theorem for extreme risks. The central result for a statistical audience is that statistical significance is itself an instance of this machinery: a level-α test is an endorsement operator over a population of resampled datasets, rejection is a failure of endorsement at the observed sample, and the p-value is a graded endorsement value — so the two verdicts are one logic over two populations, and are logically independent. I illustrate both verdicts on real S&P 500 tail-loss data, with Kupiec and Christoffersen backtests, exhibiting a statistically decisive model failure that reports no breach and a fragile breach that an admissible challenger overturns. No background beyond hypothesis testing is assumed.
(This talk is part of the Statistics and Actuarial Science series.)
PDF notice
Return to all seminars