UFC 333

Every fighter carries a profile built only from fights that had already happened when the prediction was made — never their career averages, which would contain the result being predicted. The models are refitted for each card against everything before that date.

“Will it finish?” is the claim worth reading. Tested on a held-out window it reached 60.7% against a 52% baseline, and its probabilities are well calibrated: fights called at 60–70% finished 68.5% of the time.

“Who wins?” is weak, and we would rather say so. On the same test it reached 62.6% — which is not statistically distinguishable from simply picking the fighter with the better UFC win rate. It is published for the record, not because it is clever.

Knockout or submission, we say only sometimes. Across all finishes the model is barely better than each division’s own knockout share and is worse on accuracy, so a probability printed on every bout would be meaningless. Above 75% confidence it earns its place: on the finishes it flags, the real knockout rate ran 8 to 14 points clear of what the division alone predicts, in every one of five walk-forward test windows, covering about a quarter of finishes. The call appears on those bouts and nowhere else.

It almost never calls a submission, and that is the finding rather than a gap. Not one bout in the held-out window reached 75% that way. A knockout is written into two fighters’ careers long before the cage door shuts; a submission tends to come out of a scramble no record predicts. Every bout still shows its division’s historical mix, which is context and not a claim.

It has no concept of who a fighter has beaten. A dominant champion and a fighter with the same record against far weaker opposition look identical to it, which is the single biggest thing missing and the next thing we will add.

Where a fighter is making their UFC debut, is a late replacement, or has too few fights with statistics, we publish no prediction at all rather than guess. Those bouts are shown as abstentions.

This exists to answer a question about data, not to help anyone gamble. We publish no odds, make no comparison to bookmakers’ lines, and none of it should be used to stake money. The insight stories explain the analysis these models are built from.