Methodology

Imagine you want to know how many strikeouts a pitcher will record tonight. You could look at his season average and guess — but that gives you a single number with no sense of the range.

Monte Carlo simulation takes a different approach. Instead of one answer, you simulate the full game thousands of times, introducing realistic randomness at every plate appearance. After 2,500 runs, you have a complete distribution — not just "he'll probably get 6 Ks" but "there's a 61% chance he gets 7 or more."

The name comes from the famous casino in Monaco. Like a casino, the method relies on the law of large numbers: run enough trials and the pattern converges on the real underlying probabilities.

FullCountProps is built in two layers that work together:

The matchup model answers: "What happens when this pitcher faces this batter in this context?" The simulation engine chains those answers together across an entire game — tracking innings, outs, baserunners, pitch count, and fatigue — to produce full stat distributions.

At the heart of everything is a LightGBM gradient-boosted tree classifier trained on over 6 million plate appearances from 5 seasons of Statcast data (2021–2025).

For every upcoming plate appearance, the model takes 33 features and outputs the probability of 8 possible outcomes:

Percentages shown are 2024 league averages. The model adjusts these for every specific pitcher-batter matchup.

The 33 Features

Our feature set is organized into five categories:

• 10
  Pitcher Statcast — K rate, BB rate, HR rate, whiff%, called strike + whiff%, zone%, swinging-strike%, avg fastball velocity, chase rate, in-zone contact%
• 10
  Batter Statcast — K rate, BB rate, HR rate, xBA, xSLG, barrel%, hard-hit%, chase rate, whiff%, contact%
• 5
  Matchup Context — platoon advantage, home/away, park HR factor, park K factor, park hits factor
• 5
  Game-Day Context — umpire K-rate tendency, catcher framing score, pitcher recent 14-day K rate, batter recent 14-day BA, sportsbook game total line
• 3
  Weather — game-time temperature, wind speed, wind direction (blowing out vs. in)

Each of the 2,500 simulations plays out a complete baseball game, plate appearance by plate appearance:

1. Load the confirmed starting lineup and batting order from the MLB Stats API
2. For each plate appearance, the matchup model predicts the probability of all 8 outcomes for this specific pitcher vs. batter
3. Apply context adjustments: park factors, umpire tendency, catcher framing, weather, platoon
4. Randomly sample one outcome from the adjusted probability distribution
5. Update game state — advance runners, record outs, score runs, track individual player stats
6. When the pitcher reaches his simulated pitch count limit, hand off to the bullpen composite
7. Repeat through 9 innings (or extras if tied)

After 2,500 full games, each player has a frequency distribution of outcomes. "Corbin Burnes recorded 7+ Ks in 1,530 of 2,500 sims" gives us P(Over 6.5 Ks) = 61.2%.

Raw matchup probabilities are adjusted by five real-world factors before each PA outcome is sampled:

Once we have the simulated probability distribution, we compare it to the sportsbook's odds to find mispriced props:

A positive edge means our model believes the outcome happens more often than the sportsbook's price implies. We surface all props with edges of 3% or more — anything below that is within the noise.

Removing the Vig

Sportsbooks build in a profit margin (the "vig" or "juice") on every line. Typical MLB prop vig is 4–5%. Before comparing our probability to the market, we mathematically strip out the vig to get the book's true implied probability. This ensures we're measuring real edge, not just beating the vig.

Kelly Criterion Sizing

For each edge, we calculate an optimal bet size using the Kelly Criterion — a mathematically proven formula for maximizing long-run bankroll growth. We use quarter-Kelly (25% of the theoretical optimum) to account for model uncertainty, with a hard cap of 5% of bankroll on any single bet.

This is what makes FullCountProps different from every other prop analytics service. For every single projection, you can see exactly what drove the number:

If we say "bet the over on strikeouts" and you see the umpire factor is driving half the edge, you can decide for yourself whether you trust that signal. If the data feed pulled the wrong umpire, you can catch the mistake before placing a bet.

This kind of transparency doesn't exist at BallparkPal, THE BAT X, or any other prop service we're aware of. We think it should be the standard.

We backtest every model version using strict out-of-sample walk-forward testing. The model is trained on seasons T-3 through T-1 and tested on season T, with no future information leaking into training features.

Calibration

A well-calibrated model "hits what it says." When we predict 60% probability, the outcome should happen about 60% of the time. Our current calibration error (ECE) is 3.1% — meaning on average, our predictions are off by about 3 percentage points. That's considered good for sports prediction.

Calibration data from 12,847 graded props in the 2024 backtest.

We believe in transparency about what the model does not do well:

• ⚠Bullpen transitions are simplified — when a starter exits, we use the team's aggregate bullpen stats rather than modeling individual relievers.
• ⚠Pinch hitting and defensive substitutions are not modeled — late-game lineup changes can affect PA volume for projected starters.
• ⚠Early season data (April) is noisy — the model leans heavily on prior-season rates until ~400+ PA accumulate.
• ⚠Weather data is fetched ~75 minutes before first pitch and may not reflect conditions for late-starting or rain-delayed games.
• ⚠Stolen bases and errors are not modeled — these affect ~2-3% of base-running situations.