Win probability models are everywhere in baseball now. Your sportsbook app shows them. Analytics sites post them. Even broadcast overlays show a percentage ticking up and down throughout the game.
But most people using these numbers don't know what's actually inside them — and that leads to either over-trusting the model or dismissing it entirely. Both are mistakes.
Here's what you actually need to know.
What a win probability model does
At its core, a win probability model answers one question: given everything we know about these two teams right now, how often does Team A win?
"Right now" is the key phrase. There are two kinds of win probability:
Pre-game win probability — Based on season stats, pitching matchups, home field, rest. Calculated before first pitch.
In-game win probability — Updates live as the game unfolds: score, inning, base/out situation.
For betting purposes, you almost always care about pre-game probability. That's the number that tells you whether the line is fair before money is committed.
What goes into the model
Different models use different inputs, but the core factors are consistent:
- Winning percentage — The most obvious signal. Teams that win more should win more.
- Run differential — Often more predictive than W-L record, especially early. A team that's 8-8 but outscoring opponents by 2 runs per game is probably better than their record shows.
- ERA and pitching — Starting pitcher matchups matter in baseball more than any other major sport. The same team can swing 10–15% in win probability depending on who's starting.
- Home field advantage — Real but often overpriced. In MLB, home field is worth roughly 4–6% in win probability historically. Our model uses a fixed bump.
- Rest and travel — Secondary factors. Usually embedded in team performance implicitly.
What the number actually means
If a model says a team has a 58% win probability, that means: over many games with similar matchup conditions, that team wins about 58% of the time.
It does not mean:
- They will definitely win tonight
- The model is right about this specific game
- You should bet them without checking the price
A 58% win probability has a fair moneyline of roughly -138. If the book is showing -120, there's edge. If the book is showing -160, you're overpaying — the line implies a higher probability than the model assigns.
This is the entire point of the edge calculation.
The problem with early-season models
Early in a baseball season — the first 3–6 weeks — win probability models are operating on thin samples. A team that goes 8-2 in their first 10 games is probably not a 80% win-rate team. Regression to the mean is real and it hits hard.
There are two ways models handle this:
1. Ignore it (common, bad) — Just use current-season W-L and run differential as inputs. Gets noisy fast in April.
2. Shrinkage toward a prior (better) — Weight current-season results by sample size and blend with preseason expectations. As more games play out, current performance gets more weight.
Our model uses the second approach. Early in the season, we blend current stats with a league-average baseline. By June, the current stats dominate. This means early-season edge readings are more conservative — we're less likely to post a +20% edge in April because we're not fully trusting 15-game samples.
Practical implication: be skeptical of any model (ours or anyone else's) that shows wild edges in April. Either the model doesn't apply shrinkage, or there's a pricing inefficiency that won't last.
What win probability can't tell you
- Injury news after the model ran — If a lineup change or bullpen news drops, the number is stale. Always check injury reports.
- Weather — Wind and temperature affect run scoring. Models usually don't incorporate game-day weather.
- Umpire tendencies — Some analysts incorporate ump data. Most don't.
- Motivation and pressure — Completely outside the model. Treat it as noise unless you have a strong reason to believe otherwise.
How to use it in practice
Step 1: Get the model's win probability. We post ours on every pick with a rationale explaining where the number is coming from.
Step 2: Calculate the fair moneyline. For a favorite: `-(probability / (1 - probability)) × 100`. For an underdog: `((1 - probability) / probability) × 100`.
Step 3: Compare to the book's line. If the book's implied probability is meaningfully lower than the model's, that's edge. If it's higher, you're paying a premium.
Step 4: Decide whether the edge is big enough to act on. We post the edge percentage on every pick. Generally, edges under 3% are noise. Edges above 8–10% are either a real opportunity or a data problem worth checking.
Step 5: Check the model's assumptions. Is the starting pitcher confirmed? Is there recent news that would change the inputs? A good model is a starting point, not a final answer.
Win probability is a tool, not an oracle. The teams that win more on average win more games. The models that track real performance inputs beat gut feelings over large samples. But no model wins every game, and no edge calculation is guaranteed.
The goal is to find spots where you're getting paid more than the risk warrants — then do that consistently over hundreds of games.
For informational use only. Past results don't guarantee future performance. Bet responsibly.