Find where the model disagrees with the market.

FIGHTintelligence is an independent UFC analytics project that produces pre-fight win probabilities using machine learning.

Each prediction is generated before the event, evaluated through a monthly walk-forward process, and benchmarked against Pinnacle's margin-adjusted market probabilities.

The goal is not to replace judgment or claim certainty. It is to provide a transparent, data-driven view of each matchup and highlight where the model disagrees with the market.

I'm Vincent Lugat, a data scientist with over 12 years of experience in machine learning, predictive analytics, and the deployment of production-grade models.

I've followed MMA closely for more than 15 years and have trained in Brazilian Jiu-Jitsu for 5 years. FIGHTintelligence grew from the intersection of those two interests: quantitative modelling and technical fight analysis.

LinkedIn ↗

The model is trained on historical UFC fights from 2012 onward and retrained monthly using only information available before each fight.

More than 1,500 engineered features are generated for every matchup, covering:

• Technical performance: striking accuracy, output, impact, grappling, wrestling, control, defensive efficiency, and round-by-round pace.
• Recent form: time-weighted performances that give more importance to recent fights without ignoring longer-term history.
• Relative matchup fit: direct comparisons between both fighters' strengths, weaknesses, physical attributes, and style-related advantages.
• Competition level: opponent quality, ranking trajectory, main-event experience, title-fight experience, and five-round exposure.
• Physical and contextual factors: age, reach, activity, layoffs, weight-class context, and other measurable pre-fight information.

Raw fight statistics are not used in isolation. Many features are opponent-adjusted, time-weighted, or expressed as direct comparisons between the two fighters.

The final model is an XGBoost classifier. No post-fight information is allowed into training, feature generation, or prediction.

The written matchup analysis is generated separately from the win probability.

A large language model receives structured pre-fight information prepared by the analytics pipeline, including selected technical advantages, recent form, physical differences, competition level, fight format, and matchup context.

Its behavior is guided by a dedicated system prompt designed to keep the analysis concise, technically grounded, and consistent across fights. The generation process also includes structured output constraints and automatic validation to reduce unsupported claims and formatting errors.

The language model does not browse the web, does not change the model's probability, and does not make an independent prediction. Its role is to turn structured matchup information into natural MMA analysis.

Performance is evaluated through a monthly walk-forward process.

For each period, the model is trained only on fights that occurred before that month, then tested on future fights. This better reflects how the system behaves in production than a random train-test split.

Results are compared with Pinnacle's margin-adjusted closing probabilities using classification accuracy, Brier score, calibration, and betting-strategy performance.

Historical results are not rewritten after events.

The betting strategy is fixed. A fight qualifies only when:

• the model probability is at least 50%;
• the model probability exceeds Pinnacle's margin-adjusted market probability;
• the resulting market edge is positive.

Model-favored does not mean market favorite. A fighter can have a 51% model probability while still being priced as an underdog by the market. The 50% threshold refers to the model's own assessment, not the bookmaker's.