A Bayesian Check-In On Our Playoff Odds at the Quarter-Season Mark

As the resident FanGraphs playoff odds watcher, I spend a lot of time looking at our playoff odds and trying to figure out both what they’re seeing and what they’re missing. Over the years, I’ve written many audits of how well our odds perform. Last fall, I described a Bayesian method that does slightly better than any of our existing models at predicting playoff teams. It’s particularly useful early in the season, when the headline FanGraphs mode (using projections) can be slow to pick up on new information and the season-to-date mode is prone to overreaction. A Bayesian filter does a good job balancing those two – or so I found last year.

If you’re looking for a detailed technical description of the way that I’m blending up our existing playoff projections to churn out different odds projections, you can find it at the bottom of the article. But first, let’s take a Bayesian trip through the league and highlight the divisions where reconsidering our odds in light of how much the results so far have diverged from preseason expectations matters the most.

AL East

If all the results looked like this, I wouldn’t have written this article. The Bayesian version of the playoff odds sit somewhere between the FanGraphs-style odds and the season-to-date odds, with exactly where it sits in between determined by how closely the team’s performance matches the FanGraphs model’s prior expectation. But pretty much across the board here, both sets of odds are in broad agreement, which means that my Bayesian averaging settles pretty much right in the middle. The Rays have banked enough wins that the FanGraphs model is very high on them even with a projected rest-of-season winning percentage right around .500.

AL Central

Now this is what I’m talking about. The White Sox are a fascinating team, and I completely sympathize with the projection-based modeling. The White Sox might be above .500, but they’re not doing it in a way that looks sustainable. They’ve been outscored on the year. Many of their veteran players are having their best stretches as big leaguers; rookies Munetaka Murakami and Sam Antonacci are playing far better than even an optimistic projection would expect.

That’s why the projections are still down on the Pale Hose. The Bayesian model doesn’t know any of that, though. It just knows that, on average, teams that play much better than their preseason projection in the early going tend to do better than the projections think the rest of the way. There hasn’t been much daylight between the four teams chasing the Guardians so far this year, and the Bayesian version of our odds takes that seriously. If the division looks this muddled, and the projection-based odds are in such disagreement with the season-to-date odds, the Bayesian model splits the difference and forecasts a tremendously interesting divisional race. The only problem is that it might be for second place – the Guardians have a divisional lead, and both models think they’re better than the teams closest to them in the standings.

AL West

The story here is that the Mariners have gotten off to a slow start, not that anyone in the rest of the West has taken advantage so far. Our odds give them a huge chance of winning the division because we gave them a huge chance of winning the division in the preseason and not enough has transpired to change the math very much. The Bayesian model looks at the fact that they’re 23-26, not particularly close to their 88-win preseason projection, and leans on the season-to-date odds more heavily as a result. The season-to-date odds see a three-way race in the division. Blend that in, and you get the Mariners slightly out in front, but with the A’s and Rangers close behind.

This general shape – the preseason favorite comes out slowly, while the rest of the division is mostly business as usual – is where the Bayesian model picks up most of its relative value. The FanGraphs projection playoff odds model has a specific weakness: It’s sometimes too certain of playoff likelihood early in the year. The season-to-date playoff model often has the opposite problem. The Bayesian model sorts through those shortcomings in a reasonable way, and it’s easiest to notice when the division is bifurcated like this.

NL East

Oh man, this is a fun one. I’m truly not sure what to think about the Phillies, the team with the largest downgrade in playoff odds using this system. You can completely understand what the FanGraphs odds see. It’s a team full of stars who project like stars. We think the Phillies will have the fifth-best offense and the second-best defense/pitching combination over the remainder of the season. They’re not even below .500 now! Sounds like a surefire playoff team to me.

The thing is, the Phillies haven’t really played like that elite team this year. They’ve been outscored by 19 runs, and they’re scoring fewer runs than average while allowing more runs than average. I’m fascinated by this divergence. We don’t really have enough history to test situations exactly like this very often. The Phillies have played quite poorly in 2026, but they’ve also gotten lucky in the early going – they have a winning record. If they’d played to their run differential so far, the FanGraphs odds would be much lower on them. But the FanGraphs odds don’t consider the recent past, aside from how it feeds into player projections. They just start with today’s record and project forward.

I’m truly not sure whether the Bayesian method will handle this well. It has inherent limitations here, and it’s also built on regressions. Philadelphia falls into a potential hole in the model – there aren’t a lot of teams projected as well as the Phillies who scuffle early to begin with, and few of those outperform their run differential en route to an above-.500 record while doing so. This model is a little better than the FanGraphs model, but there’s a lot of irreducible uncertainty when it comes to projecting playoff teams.

The Mets, naturally, take a small hit here. The Marlins and Nats get a boost from their early play. None of that seems wild to me. But that enormous gap between how the different methods see the Phillies is really interesting – and to me, it says that the Phillies themselves are really interesting as a result.

NL Central

Oh look, FanGraphs odds were too low on the Brewers before the season, what a surprise! I’ve written about this before. Our odds seem particularly ill-suited to projecting the Brewers because the places where they find edges are generally not well-modeled. Baserunning, defense, the combinatorial effects of baserunning and low-strikeout batters; we don’t do a great job handling these factors, and we know it. The Bayesian model has that covered, though; the Brewers are outperforming the projections, so it simply looks at the projections less.

The Cubs aren’t the losers here, though. They’ve played almost exactly like we expected them to. The Cardinals have exceeded expectations. There’s really nowhere to subtract in the division. That’s why the Central is the division that adds the most cumulative playoff odds when moving from a projection model to a Bayesian one. These teams have been great so far, while their playoff competitors have scuffled. That explains the odds change very well.

NL West

Boring! But really, what can I say? Most of these teams are who we thought they were. The Giants, the only exception to that, have dug themselves a hole so deep that regardless of what odds system you use, they’re unlikely to make the postseason. When priors match observations, Bayesian inference looks at the result and mostly leaves the priors alone. Seems good to me.

How This All Works

I tested out a number of different Bayesian approaches when designing this study. The one I used here is the simplest of those methods, though none of them are that simple. I broke the season up into months. For each month, I built a pool of candidate weightings to give each model. More specifically, I varied “how much to trust the prior (the FanGraphs odds)” and “how much a deviation from the prior should change the model’s opinion.” I then tested each of these candidate weightings on prior years to see which had done the best, and selected final weights for each month based on that.

This is a regression-based model, which means it’s tuned based on what has actually happened rather than a first-principles approach to how much each model’s input matters. That’s an inherent limitation, because past performance may not be indicative of future results. But on the other hand, regardless of how I’m picking the weights, I’m testing everything out of sample first. In other words, I’m building my model through 2017 and testing it on the 2018 season, rebuilding it that year and testing it on the 2019 season, and so on. This Bayesian model beats our two existing models; it reduces error by about 2.4% compared to our FanGraphs odds.

With those weights in hand, my Bayesian method calculates each team’s odds by looking at their form so far in this season, accounting for schedule, and comparing it to their preseason FanGraphs-projected winning percentage. The larger the disagreement, the less weight the model puts on the FanGraphs odds and the more it uses the ones based on season-to-date play. It’s as simple as that. You could design a more complex way of handling this, but I like that this one is relatively straightforward. Look at the two, re-weight your priors based on the information you’ve received, and move on.

I’m not saying that you should take these odds as gospel, to be clear. Playoff odds have inherent limitations. We’re still just guessing about the future. But if you’re looking at FanGraphs odds and wondering why the White Sox can’t get any respect even with the fourth-best record in the American League, well, our odds have historically had a problem with underrating teams like that. This is my best guess at a solution.