Oracle AI Women's Bantamweight Accuracy Report

The Numbers

The Oracle AI model has recorded 5 predictions in the Women's Bantamweight division, with 2 correct outcomes for a raw accuracy of 40%. This falls below the performance threshold typically associated with profitable or reliable fight forecasting.

The Brier Score of 0.2776 quantifies calibration error across probability estimates. With elite performance defined below 0.15, good below 0.20, and average below 0.25, this score registers as below average — indicating the model's probability assignments have been poorly calibrated in this weight class. The model has expressed confidence levels that systematically overstated actual win rates.

Method accuracy of 80% (4 of 5 fights) shows the model has identified how fights end with reasonable precision, even when the who has proven elusive. This divergence — strong method prediction alongside weak winner prediction — suggests the model captures fight dynamics but struggles with fighter-specific assessment in this division.

For context: established MMA prediction systems typically operate between 60-70% accuracy on main-card matchups. At 40% over five fights, this sample performs worse than random guessing would over extended trials.

Confidence Calibration

The tier breakdown reveals uniform negative edge across all confidence levels — the model has underperformed its stated probability in every tier where predictions exist.

High confidence (70-84%): 1 correct from 2 picks (50% actual vs. 77% expected). The -27% edge here is concerning; these were the model's most trusted predictions, and they failed at a rate far above expectation.

Medium confidence (60-69%): 1 correct from 2 picks (50% actual vs. 64.5% expected). The -14.5% edge represents the smallest deviation, suggesting this tier has been least miscalibrated — though still unprofitable.

Low confidence (50-59%): 0 correct from 1 pick (0% actual vs. 54.5% expected). The -54.5% edge is severe but derives from a single prediction; no robust conclusion is possible.

Lock tier (85%+): No predictions issued. The model has not identified any Women's Bantamweight fights as high-certainty opportunities — a prudent restraint given overall performance, though this also means no positive edge exists to offset losses elsewhere.

No tier has generated positive expected value. The model has been overconfident at every operational level in this division.

What the D Grade Means

A D grade signals performance that demands structural skepticism. This is not a "rough patch" or variance within normal bounds — over five predictions, 40% accuracy with poor calibration indicates the model's current configuration lacks predictive validity for Women's Bantamweight specifically.

Honest limitations: The sample size of five predictions is small. A single additional correct pick would raise accuracy to 60%, altering superficial interpretation. However, the Brier Score — which incorporates confidence calibration across all predictions — provides independent evidence of poor performance that sample size alone does not excuse. The 0.2776 score reflects systematic overconfidence, not merely bad luck.

Scope constraint: This report covers a cross-promotion weight-class slice, not the model's full public record. Performance in Women's Bantamweight does not predict performance in other divisions, weight classes, or promotional environments. Do not generalize this grade to the Oracle AI system as a whole.

The 80% method accuracy offers a genuine bright spot, but bettors or analysts relying on winner prediction should treat this division as a known model weakness until sample size and accuracy improve substantially.

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. Draws and No Contests are excluded from accuracy calculations; this report contains zero such exclusions.

This analysis covers Women's Bantamweight across all included promotions — a mixed promotional environment, not UFC-specific or single-promotion data. For complete methodology documentation, see the full technical breakdown.

Report generated from verified prediction log. All statistics derived from actual recorded outcomes.