Oracle AI Middleweight Accuracy Report: 71.4% Raw Accuracy, B+ Grade

The Numbers

The Oracle AI Middleweight model has recorded 10 correct predictions from 14 fights, yielding a 71.4% raw accuracy rate. The Brier Score of 0.1969 falls within the "good" performance band (below 0.20), indicating that the model's probability estimates align reasonably with actual outcomes. A Brier Score in this range suggests calibrated confidence rather than wild overstatement or excessive conservatism.

Method accuracy sits at 46.2% — meaning the model correctly predicted how fights would end (KO, submission, or decision) in under half of cases. This is a known limitation: predicting fight outcomes proves substantially easier than predicting fight methods, a pattern consistent across MMA prediction literature.

Context matters for interpretation. The sample size of 14 predictions is small — roughly half a typical UFC calendar year for one division. In MMA prediction, 60-65% accuracy over large samples often separates competent models from coin-flipping. At 71.4%, this model shows promise, though the limited n means variance could shift these numbers substantially with 10-15 additional fights.

Confidence Calibration: Where Edge Lives

The tier breakdown reveals where the model earns its keep — and where it lacks sufficient conviction.

Lock picks (85%+ confidence): 0 predictions made. The model issued no highest-confidence calls in this Middleweight sample. This is notable: the expected hit rate for Locks is 92.5%, so the -92.5% "edge" reflects absence, not failure. The model simply found no Middleweight matchups compelling enough to warrant maximum confidence. This conservatism may limit upside but also avoids catastrophic Lock misses.

High confidence (70-84%): 2/2, 100% actual vs. 77% expected (+23% edge). Small sample, perfect execution. Both High calls converted.

Medium confidence (60-69%): 5/7, 71.4% actual vs. 64.5% expected (+6.9% edge). The model's most active tier performs above expectation, generating modest positive edge.

Low confidence (50-59%): 3/5, 60% actual vs. 54.5% expected (+5.5% edge). Even coin-flip territory produces slight outperformance.

Positive edge exists across all tiers where predictions were made. The model's restraint at the top — refusing to force Lock calls — appears warranted by the data.

What the B+ Grade Means

A B+ grade signals solid, above-average predictive performance with clear limitations. The 0.1969 Brier Score confirms the model makes probability estimates that respect outcome uncertainty. The 71.4% accuracy exceeds typical MMA prediction benchmarks.

Honest weaknesses: The 46.2% method accuracy is poor. The model identifies winners more reliably than it anticipates how they win. The 14-fight sample is too small for statistical confidence in long-term performance — add 20 more fights and this grade could shift to A- or C+. The absence of Lock predictions suggests either unusual Middleweight parity during this period, or the model's threshold for certainty sits appropriately high.

Scope limitation: This report covers a cross-promotion Middleweight slice, not the headline public record. Results mix across promotions with varying talent depth, rulesets, and data quality. Comparability to single-promotion models is limited. Do not treat this as equivalent to a UFC-only Middleweight report.

Methodology

The Oracle AI operates as a 57-module prediction engine with specialist sub-models voting on fight outcomes. Predictions are recorded before events and verified against official results post-fight. Draws and No Contests are excluded from accuracy calculations — none occurred in this sample.

Full technical methodology: Methodology Breakdown

Report scope: Cross-promotion Middleweight division. Mixed data environment.