How Pick'em Labs works
Pick'em Labs is a sports analytics tool that combines publicly available fighter statistics, live sportsbook odds, and independent analysis from multiple AI models. This page explains exactly how predictions are generated, what data is used, and where the current limitations are.
The analysis process
For every fight on an upcoming card, Pick'em Labs pulls three independent inputs and combines them into a single analysis view.
- 01Fighter data is collected
Career records, physical attributes, detailed fight statistics, and recent fight-by-fight results are sourced from the Cito API and Sherdog. This includes striking accuracy, takedown rates, submission attempts, and finish percentages.
- 02Sportsbook odds are fetched
Live moneyline odds are pulled from every major sportsbook available via The Odds API — typically 8-11 books per fight. Rather than relying on any single book's line, the market probability shown is the median across all of them, so one outlier price can't skew the picture. Implied probabilities are derived from that consensus price.
- 03AI models analyze the matchup
Three AI models — Claude, GPT-4, and Gemini — independently evaluate the matchup using fighter statistics, each fighter's recent fight-by-fight results, and odds context. Each model produces a predicted winner, a confidence score, a predicted method and round, and a structured breakdown covering key advantages, risks, and a likely fight script.
- 04Results are surfaced
The three model outputs are compared to produce a consensus pick and an average confidence score. The AI win probability is then shown alongside the market-implied probability so you can see where they agree and where they diverge.
AI models
Pick'em Labs currently uses three large language models. Each model receives the same structured prompt containing fighter statistics and odds data. They analyze the matchup independently — no model sees another's output before producing its prediction.
Developed by Anthropic. Known for careful, structured reasoning and nuanced risk assessment. Claude tends to produce detailed fight scripts and explicit breakdowns of counter-arguments.
Developed by OpenAI. Broadly capable across sports analysis tasks and draws on extensive training data covering MMA history, fighter profiles, and fighting styles.
Developed by Google DeepMind. Provides a third independent perspective on each matchup, helping to identify cases where model agreement is strong versus cases where predictions diverge significantly.
Data sources
Pick'em Labs currently draws from three primary data sources.
Career and per-fight metrics are sourced from the Cito API and Sherdog. This includes significant strike rates, accuracy, takedown data, submission attempts, and win/loss records. Statistics reflect career averages and may not capture very recent fights until data is updated.
Each fighter's most recent completed UFC bouts — opponent, result, method, and round — are pulled from the same sources and included alongside career stats, so the AI models have fight-by-fight context, not just career averages.
Moneyline odds are fetched in real time from The Odds API, which aggregates lines from multiple major sportsbooks. Odds are displayed in American format. A median implied probability across all available books is calculated and used as an input to the AI analysis prompt.
Confidence scores
Each AI model returns a confidence score between 0 and 100 alongside its predicted winner. This score represents the model's self-assessed certainty in its own prediction given the data it was provided.
The consensus confidence shown in the Model Consensus panel is the arithmetic average of the three individual model scores. It is not a probability of the predicted fighter winning — it is a measure of how certain the models are collectively in their pick.
A high consensus confidence combined with a pick that disagrees with the betting market is typically the most analytically interesting signal Pick'em Labs surfaces.
AI vs market
The AI vs. Market panel compares the AI consensus confidence against the market-implied probability derived from sportsbook odds. The difference between these two numbers is displayed as the AI–Market Gap.
A positive gap means the AI models collectively assign a higher win probability to a fighter than the market does. A negative gap means the market is more confident in that fighter than the AI models are. The gap is also labeled with a tier — from "Market-aligned" to "High-risk contrarian pick" — based on how big an underdog the AI's pick is per the market; a large enough gap on a genuine underdog is flagged with a Contrarian pick badge.
This comparison is informational. It is intended to surface disagreements between AI analysis and market consensus, not to imply that either the AI or the market is correct.
Current limitations
Pick'em Labs is an early-stage tool. There are several known limitations to be aware of when interpreting its output.
- Statistical recencyFighter statistics reflect career averages. Recent form, injuries, coaching changes, and camp reports are not systematically incorporated into the data layer.
- AI knowledge cutoffsLarge language models have training cutoffs. Fights, results, and fighter developments that occurred after a model's cutoff date may not be reflected in its analysis.
- No round-by-round granularityEach fighter's recent fights are included at the result level (opponent, method, round, time) — not round-by-round strike or grappling breakdowns. Career stat comparisons still rely on aggregate averages, not fight-by-fight splits.
- Model variabilityAI model outputs are not deterministic. The same prompt may produce slightly different predictions across separate requests. Predictions should be treated as one analytical input, not a definitive forecast.