UFC Fight Predictor debuts with 71.6% accuracy

The interesting part here is less the flashy AI wrapper and more the modeling discipline: leakage control, differential features, and SHAP-driven explanations make this feel like a serious learning project rather than a toy demo.

• –The 71.6% number is respectable, but UFC is noisy enough that calibration and live performance will matter as much as raw accuracy
• –Using fighter_A minus fighter_B features is a smart choice because it makes feature attributions easier to interpret
• –The division-style prior is a reasonable domain tweak, but it also risks baking in historical bias if the underlying style labels are coarse
• –Claude is being used as a narrator, not a predictor, which is the right architecture if the deterministic model is doing the actual scoring
• –The biggest value here is educational: it shows a full pipeline from data hygiene to inference to explanation, not just a leaderboard score