How PropsMath generates forecasts

1. Inputs

Each MMA forecast draws on three inputs:

• Sportsbook market consensus (via BestFightOdds and an aggregated odds feed) — the market implied probability, devigged (the book's built-in margin, the "vig", removed so the number reads as a fair chance) across all books offering the outcome.
• UFC Stats fighter history — age, recent form (results over the last three fights), career win rate, finish rate, and recency-weighted striking output per fighter. These feed a fundamentals adjustment whose weights are hand-set starting points, not yet machine-fitted — so we keep its influence deliberately bounded.
• Pre-fight research signals and X/Twitter social sentiment — evidence and chatter gathered before the fight, graded by source quality and used as soft context, never as a primary signal.

2. Blending

The forecast starts from the devigged market price as its anchor. A bounded fundamentals adjustment then nudges that anchor up or down from the fighter-history features above — and that nudge is scaled downon thin or illiquid markets, where the line itself is less trustworthy. When our model and the market clearly disagree, we say so plainly instead of averaging them into one number that sounds more certain than it really is. We'd rather show you an honest disagreement than a tidy number that hides one.

On top of that market-and-fundamentals number, we fold in research signals— evidence and social chatter gathered before a fight, weighted by how reliable that kind of signal has proven to be. These adjustments are deliberately small: a signal can't move a forecast much until its type has earned a track record of being right, and during the live event we hold them back entirely until they've measurably improved accuracy. They're a finishing adjustment, never the foundation.

3. Model-market divergence

For each outcome we compute the gap between our model probability and the market-implied probability. We call this model-market divergence — the raw signal that, once it clears the confidence and per-market bars described below, becomes a surfaced forecast.

4. Calibration

Probabilistic forecasts are only useful if they're calibrated — i.e., the events we predict at 60% actually happen about 60% of the time. Every forecast is written to a permanent calibration log before the event starts. After the event resolves, we attach the actual outcome and update the public calibration chart on the track record page. We never rewrite or remove a past forecast.

We also feed the resolved log back into futureforecasts. Before each new probability is published, we look up what the historical hit rate was for predictions in the same confidence bin and pull the new estimate toward that empirical rate.

When the sportsbook consensus and the prediction-exchange consensus diverge meaningfully on a fighter, our model probability for that side is shrunk toward 50/50 in proportion to the gap. The market disagreement itself is statistical uncertainty, and confidence claims that ignore it would be dishonest.

5. Accuracy vs the market (the public benchmark)

Anyone can claim to be the most accurate forecaster. The honest test is the same one used in quantitative finance: compute the Brier score (how far our percentages sat from what actually happened) and log-loss (a stricter version that punishes confident misses harder) of our published probabilities against the actual outcomes, then run the identical computation on the de-vigged sportsbook consensus — the same margin-removed market probability from section 1 — for the same fights. Lower scores are better. The difference — positive when PropsMath outperforms the market — is published live on the track record page and updates automatically whenever a fight resolves. We never remove a forecast, even one that scored badly.

6. Confidence and tiers

Two checks decide whether a divergence is worth surfacing, and at what tier. First, every forecast is graded for confidence from how much model probability and edge it carries:

• High — the model gives the side at least a 65% chance and sits at least 8 percentage points above the market.
• Medium — at least a 55% chance and 5 percentage points of edge.
• Low — anything below that.

Each market then sets its own bar: tighter, noisier markets (a specific method-and-round, say) demand more edge and higher confidence than a clean moneyline. An outcome that clears both its edge and confidence bar — and that the model rates above 50% — is surfaced. The single highest expected-value qualifier becomes the Top Forecast; the rest are Forecast Opportunities; everything else is filed below threshold and still written to the public record.

7. What this isn't

PropsMath is not a sportsbook, a tout service, or an investment adviser. We do not handle wagers, escrow funds, or guarantee outcomes. Forecasts are published for informational and analytical purposes only. Past forecast performance does not predict future results.

If you choose to act on a forecast, that decision is yours alone. 21+. If sports betting is affecting your life, help is available — see our responsible gambling resources.