Allen vs Duncan: what the archived predictions show

The card and the question

This is a retrospective review of UFC Fight Night Allen vs Duncan, dated October 10, 2026. We ask how the published selections performed in each market, and what those results can actually establish. The source is Verdikt AI's public event archive; this is an analysis of our own recorded outputs, not an independent validation.

We keep the market definitions separate. Picking a winner, picking a finish method and predicting a round total are different questions with different denominators. Several selections can concern the same fight, so the table must not be treated as a collection of independent experiments.

Inspect the complete five-market event record

Results and probability error

Only correct and incorrect grades enter the graded denominator. Draws, no contests, cancellations and ungraded entries are excluded and counted under Other statuses. The Brier score is the mean squared difference between the probability attached to the selected outcome and its binary grade: mean((p − outcome)²). A correct grade uses outcome 1; an incorrect grade uses outcome 0. Lower is better for otherwise comparable predictions.

The probability calculation excludes any graded row without a usable confidence estimate; Scored probabilities gives its exact denominator. Scores across different markets should not be read as a ranking of models: the outcomes, available lines and base rates differ. In particular, winning method is a selected-outcome score, not a full multiclass Brier score.

A confident selection can lose

One concrete moneyline error was Felipe Franco versus Brendson Ribeiro. The archived pick was Felipe Franco at 76.86% confidence; its recorded grade is incorrect. That is a useful counterexample to treating the confidence label as a guarantee.

An individual loss does not by itself establish poor calibration. To examine a claimed 65% probability, group many forecasts made under comparable conditions and compare their probabilities with observed frequencies. This card is too small to support a stable calibration conclusion, and fights from the same card can share conditions.

What this review does not establish

This table describes archived selections and recorded grades. It does not establish that every file was publicly available at a particular pre-fight timestamp. It is not a clean held-out experiment, an independently audited betting record, or proof of a sportsbook advantage.

Accuracy also does not equal return. A return calculation needs the exact selection, matching bookmaker price, capture time, staking rule and exclusions. This review does not substitute a model probability for a missing price or claim that a closing snapshot was executable before the event. For a disputed grade, provide the event, market, line and supporting result so that the record can be checked.

Grading and evaluation methodology · Interactive archive · Report an incorrect grade