The Model, Explained — Song vs. Figueiredo (Macao)
Every card, our prediction engine outputs a number for each fight. This column does the opposite of hiding behind that number: it walks through what the pick is, how confident the model is, and which real career stats make the lean reasonable — in plain English, no betting advice.
One honesty note up front, because it matters. The model does not say "fighter X wins because of his takedown defense." It's a four-voice ensemble — our V14 model, the newer V15+, the independent Paperclip model, and the betting market's implied price — blended by a stacking layer. It outputs probabilities, not reasons: there are no per-fight feature attributions. So when this column points at a grappling gap or a striking-volume edge, that's our editorial read of each fighter's published career stats — "here's what the stats say about why this lean is reasonable" — not a readout of the model's internals. We think saying that out loud beats pretending the box is glass. For reference, the production winner model runs roughly 70% accuracy on the 2025-2026 holdout (the precise canonical figure is being finalized against our leak-free grading).
Main Event — Song Yadong vs. Deiveson Figueiredo
Pick: Song Yadong — but barely. Raw V14 loved Song at 78.7%. The production blend pulled all the way back to 58% — a lean, nothing more. Why the cold feet? All four voices point at Song, but disagree wildly on how much: V14 78.7%, V15+ 63.1%, market 82.2%, with Paperclip the skeptic. The stacking layer leans hardest on Paperclip (its weight is the heaviest of the four), dragging an apparent blowout down to a flip.
What the stats say about why it's close: Figueiredo is the live grappling threat — a career 1.2 submissions and 1.61 takedowns per 15 minutes, versus Song's 0.1 and 0.78. Song's edge runs the other way: 4.42 significant strikes landed per minute to Figueiredo's 2.63. That's the fight in two numbers — if it stays standing, Song's volume wins; if Figueiredo drags it down, his submission threat is real. Exactly the fight the model refuses to call confidently.
Where the model fades its own V14
Three fights where the production blend disagreed with the raw V14 number — worth flagging, because the blend is what we publish:
- Perez vs Sumudaerji: V14 leans Sumudaerji; the blend flips to Perez (50.9%). Toss-up.
- Ding Meng vs Jose Henrique: V14 likes Jose Henrique at 62.2%; the blend flips to Ding Meng (50.9%). Toss-up.
- Amorim vs Lookboonmee: the sharpest split — V14, V15+, and the market all lean Amorim, yet the published pick is Lookboonmee (50.2%), because the single heaviest-weighted voice (Paperclip) drags it across. Only one of four voices favors the published side. A genuine coin-flip the model won't pretend to know.
---
Methodology: picks/probabilities from the frozen pre-fight prediction file (cutoff 2026-05-17). Career stats are UFC averages from our fighter database. The ensemble exposes probabilities and global blend weights only — not per-fight stat attributions — so the "why" framing is editorial, grounded in real career numbers. No betting advice.
Postscript (resolved): the main event flipped roughly as the model feared — Song won, but by submission in Round 2, the grappling finish the column flagged as the live variable (with the roles reversed). Across the card the production picks went 10/12 on graded fights. Published pre-fight; results noted for transparency only.