Dr. Savelove or: How I Learned to Stop Worrying and Love Relievers

To recap where we last left off in our trip down the reliever rabbit hole: long-term trends on overall reliever dollar value do not seem to be changing that much over time, based both on the number of relievers registering valuable seasons and the proportion of “one-hit wonders” in the modern era.

| Group | Threshold | One-Hit Wonders | Total Qualifiers | % One-Hit Wonder |
|---|---|---|---|---|
| Closed Career (1991-2015) | Elite | 22 | 51 | 43.1% |
| Active | Elite | 11 | 27 | 40.7% |
| Closed Career (1991-2015) | Above-Mean | 70 | 127 | 55.1% |
| Active | Above-Mean | 24 | 54 | 44.4% |
But, as I noted in my follow-up last week, elite reliever seasons are becoming less reliant on saves over time:

Not only that, but my colleague Jeff Zimmerman also highlighted the extent to which fantasy managers are struggling to capture the saves which do occur. All this points to a quagmire – what are managers to do? One option is to punt relief outright, but that isn’t necessarily the best call, as our Benign RotoGraphs Overlord pointed out recently.
Without saves playing as clear a role in demarking good relievers, my goal in this analysis was to see what statistical tools I could leverage to make choosing the right relievers easier. I think, by and large, I have succeeded.
Who is successful over time?
Before diving too deep, I wanted to look at just how frequently relievers are repeating elite seasons. It turns out that the answer is not all that much:

Unsurprisingly, the vast majority of relief seasons are non-elite (read: >$13.90, or one standard deviation above the mean in my sample). As a refresher, the pool here is relievers who have ever had a positive-value fantasy season within our closed career (1991-2025) sample.
| Threshold | n | % of All Relievers Ever Reaching | % Advancing (Conditional) |
|---|---|---|---|
| 1 | 98 | 18% | NA |
| 2 | 48 | 9% | 49% |
| 3 | 22 | 4% | 46% |
| 4 | 10 | 2% | 45% |
| 5 | 7 | 1% | 70% |
| 6 | 4 | 1% | 57% |
| 7 | 4 | 1% | 100% |
| 8 | 2 | 0% | 50% |
| 9 | 1 | 0% | 50% |
| 10 | 1 | 0% | 100% |
| 11 | 1 | 0% | 100% |
| 12 | 1 | 0% | 100% |
| 13 | 1 | 0% | 100% |
| 14 | 1 | 0% | 100% |
As we can see above, 18% of that group eventually turned in an elite season and fewer still register more than one. The proportion that eventually registers (n+1) elite seasons drops off by roughly half for each n (excepting some sample-size noise) until you reach the final boss, Mariano Rivera. Here is the list of guys from that sample who have registered at least three such seasons:
| Name | n |
|---|---|
| Mariano Rivera | 14 |
| Trevor Hoffman | 8 |
| Billy Wagner | 7 |
| Joe Nathan | 7 |
| Jonathan Papelbon | 5 |
| Keith Foulke | 5 |
| Robb Nen | 5 |
| Armando Benitez | 4 |
| Francisco Rodríguez | 4 |
| Roberto Hernandez | 4 |
| Brad Lidge | 3 |
| Emmanuel Clase | 3 |
| Eric Gagne | 3 |
| Greg Holland | 3 |
| Heath Bell | 3 |
| J.J. Putz | 3 |
| Jeff Shaw | 3 |
| Liam Hendriks | 3 |
| Roberto Osuna | 3 |
| Rod Beck | 3 |
| Wade Davis | 3 |
| Zack Britton | 3 |
And here is the list of active relievers who have likewise registered multiple elite years:
| Name | n | Most Recent |
|---|---|---|
| Craig Kimbrel | 7 | 2023 |
| Josh Hader | 7 | 2026 |
| Aroldis Chapman | 5 | 2025 |
| Kenley Jansen | 5 | 2021 |
| Cade Smith | 3 | 2026 |
| David Bednar | 3 | 2026 |
| Edwin Díaz | 3 | 2025 |
| Raisel Iglesias | 3 | 2024 |
| Devin Williams | 2 | 2023 |
| Jhoan Duran | 2 | 2026 |
| Jordan Romano | 2 | 2022 |
| Kirby Yates | 2 | 2024 |
| Mason Miller | 2 | 2026 |
| Robert Suarez | 2 | 2025 |
| Ryan Helsley | 2 | 2024 |
| Tanner Scott | 2 | 2024 |
A couple things to take away from this: while the Hall of Fame doesn’t consider fantasy performance, we may be watching an unexpected Cooperstown-bound career unfold with Josh Hader. Also: Cade Smith has more elite seasons than Mason Miller? Turns out, yes. Miller’s fantasy turn in 2025 was not elite by this measure mainly because Robert Suarez cut off his post-trade path to saves. Add in the fact that his ERA was simply very good and not microscopic, and his season was worth about $10. Another takeaway is that this again points to the idea that big-picture trends in reliever value are broadly similar. Rivera excluded, the number and distribution of long-time elite active relievers seems roughly consistent with the historical sample.
Research Questions and Methodology
The information above isn’t actually that useful to fantasy managers, though. We want to know which players will break out, which will stay good, and which we need to avoid. Inspired by my experiments with the CART model in my earlier treatise on relievers, I wanted to try to leverage machine learning or statistical techniques to answer these questions. I primarily used two: elastic net logistic regressions and random forest decision trees.
You might remember logistic regressions from one of my earlier articles about k-rate and eventual fantasy viability. These types of models predict the impact of an independent variable (or variables) on the likelihood of a categorical outcome. In my model designs, which I will explain in better detail later, this is a binary yes/no outcome. The “elastic net” piece involves a bunch of statistical gobbledygook that does a combination of two things: shrinking the impact of variables that don’t end up important in the model, and accounting for situations when independent variables are highly correlated with each other. The former’s importance is self-evident; the latter is a more nuanced issue related to the calibration of standard errors. If you are looking for a more detailed read on the subject, see here, but the short version is that standard errors are what allows us to ascertain statistical significance. If we can’t trust those, then we can’t trust the model.
This framework is especially well suited for a project like this because baseball performance data suffer from precisely these two issues. There are a lot of variables which could feasibly independently contribute to a player’s overall success and it’s quite likely that many are correlated with each other. Random forest models are perhaps more well known amongst the general public; they are essentially souped-up decision tree frameworks like the CART model I leveraged previously. The RF algorithm creates many iterations of decision trees at random.
I used these methods on a bunch of data I collected to try answer the questions I’ve been circling around for weeks now:
- Conditional on having been good in a given season:
- How likely is it that they will never be good again?
- How likely is it that they will be good again in the following season?
- Conditional on never having been good:
- How likely is it that they will break out to be good in the following season?
One thing to note is that when I refer to “good” to describe a reliever, I mean something specific: 0.5 standard deviations above the sample mean ($10.70). I originally tried to model “elite” based on one standard deviation above the mean ($13.90 as depicted above), but the sample size of pitchers who had such a season was too small to meaningfully draw conclusions. The elite threshold did, however, hold for the third question noted above, so I used both the “good” and “elite” levels for the models.
The questions are classifiable based on a given set of conditions, so I made a series of binary 0/1 variables in my data to indicate them. Those constituted my dependent variables. I set each of the models to draw from a pool of various stats as potential independent variables:
- Age, good seasons to date, current good season streak, dollar value and its one-year lag, innings pitched, saves, K%, BB%, HR/9 and their one-year lags, percent change in the rate stats, and status as a reliever in the prior season
The above pool of independent variables is admittedly arbitrary; a different study might select different choose different ones or include more lags. But for the purpose of this analysis, I think it worked out fine.
Modeling approach and data
Before getting into the results, I need to briefly explain how to determine if models like these are actually useful. There are two main categories of steps: cross validation (CV) and back-testing. Each involves withholding a certain portion of the data from the model (test data) and seeing how a fit from the remaining data (training data) performs. The CV splits pitchers in the sample into equally sized, random buckets, and withholds one of them to constitute the test data. It then repeats this in various iterations. There is no temporal dimension – this means that it might be using data from 2025, 1992, and 2004 to predict data from 2008. You can think of this as a way to show that the model is good at making predictions about pitchers it doesn’t know about.
Now, we still have another question to answer. The CV above tests the model to see how it does on pitchers it’s blind to, but back-testing allows us to see how the model performs on time periods it doesn’t know. To do this, you create a cutoff point in the sample and withhold years after it to create test data and fit a model on the data from before the cutoff. It’s important that the training data are blind – that is, they don’t contain any contamination from future events and only “know” information up to the point at which the data are cut off. Marrying this sequential aspect to the pitcher dimension above helps us ascertain the usefulness of our models. There are various metrics which can be used and I will explain them as they arise.
My data are largely the same as other analyses I’ve published; I’ve used historical data pulled from FanGraphs to calculate auction values. My first two questions above rely on the sample of pitchers who have ever registered a good season, and my third question uses a much broader pool. In contrast to my previous articles which leverage a pool based on all pitchers who have registered a positive fantasy season, my sample includes all reliever seasons which accrued more than -$10 in value. My reasoning here is that if I want to predict – among relievers who might currently be bad – who will be good, I need to cast a wider net and include some players who may never have ended up good. If I use the same sample I had before, I’m already selecting for some level of talent.
Here are the baseline “hit” rates for each model. To be successful, my models should be able to come close to guessing these rates. To measure my first question (fall), I looked at each good season registered by a relief pitcher in my closed career sample (1991-2025) and created a flag for whether or not it was that pitcher’s last good year. The closed career sample was necessary to avoid truncation bias on the later end due to active pitchers’ careers not being over yet and also on the earlier end because some of the stats I mentioned earlier (like number of consecutive good seasons) might be misrepresented for someone if the model only knows about half of their career.
| Model | Sample Size (Unique) | Outcome Rate |
|---|---|---|
| Fall | 322 (160) | 46.6% |
| Repeat | 405 (191) | 39.0% |
| Good Jump | 3475 (1495) | 3.9% |
| Elite Jump | 3768 (1518) | 2.3% |
For my next question (repeat), I looked at each season after a good relief season and flagged if it was likewise good (and a relief season). The samples here are slightly different because the repeat models don’t require closed careers and therefore are able to consume active reliever seasons. This results in a slightly different arrangement of the good player pool.
The jump models naturally have a much bigger player sample since they are in fact conditional on not already being good or elite. If we’re dredging this pool, we need to dig pretty deep. The elite model’s sample is bigger because I allowed it to consider pitchers who had also previously been good, but not elite.
Cross validation results
I judged my cross validation results based on a handful of metrics: AUC (area under receiver operating characteristic curve), Brier Score, Brier Skill, and Top-20% Capture. The way AUC works is that first, the model assigns the probability of a given outcome based on the specification. In the Fall model, for example, the outcome would be whether or not the observed good season is in fact a pitcher’s last. Then, we pair pitcher seasons which do achieve the target outcome with those that don’t for every observation. The AUC represents the percentage of times that the model has correctly assigned the “positive” result to be higher probability. A perfect score would be 1.0, and in the table below, the metrics for the base rates above are 0.50, which is the result garnered from simply guessing based on the known proportions. One thing to note is that AUC doesn’t measure how accurate the probabilities were, but instead measures how well the model is ranking the observations. What constitutes a good AUC depends on the field you’re in. Obviously, you want a score higher than the base rate, and the more critical the reliance on a correct outcome, the higher you’d want your AUC to be in order to be considered acceptable. For example, fraud detection and anti-money laundering models at financial institutions require extremely high AUCs because the consequences of being wrong mean that the bank or a customer is being defrauded, or that a legitimate customer is being wrongly accused. Likewise, epidemiological models measuring probability of disease instance require a high bar to clear; you wouldn’t want to tell someone that they have high likelihood of a chronic disease (or the opposite!) and then be wrong. For our purposes – choosing relief pitchers – we can have a lower tolerance given the stakes are much lower. Across the board, the AUC values for my models indicate some utility relative to the baseline level. That’s good!
| Model | Algorithm | AUC | Brier | Brier Skill | Top-20% Capture |
|---|---|---|---|---|---|
| Fall | Elastic Net | 0.777 | 0.1930 | 0.235 | 35.7% |
| Fall | Random Forest | 0.723 | 0.2123 | 0.159 | 34.9% |
| Fall | Training Base Rate | 0.500 | 0.2523 | 0.000 | NA |
| Repeat | Elastic Net | 0.697 | 0.2142 | 0.106 | 30.4% |
| Repeat | Random Forest | 0.682 | 0.2150 | 0.103 | 29.1% |
| Repeat | Training Base Rate | 0.500 | 0.2396 | 0.000 | NA |
| Half-Elite Jump | Elastic Net | 0.760 | 0.0347 | 0.073 | 57.7% |
| Half-Elite Jump | Random Forest | 0.735 | 0.0348 | 0.068 | 53.9% |
| Half-Elite Jump | Training Base Rate | 0.500 | 0.0374 | 0.000 | NA |
| Elite Jump | Elastic Net | 0.838 | 0.0210 | 0.080 | 73.9% |
| Elite Jump | Random Forest | 0.774 | 0.0217 | 0.050 | 64.5% |
| Elite Jump | Training Base Rate | 0.500 | 0.0228 | 0.000 | NA |
Moving on to the others, Brier and Brier Skill are obviously related. Brier score is the mean squared error between the predicted probabilities and the actual outcomes. Lower is better, but the scores themselves are relative. You’ll notice that my jump models have much lower scores than the fall or repeat models, and that’s because the a baseline prediction of fail (i.e. not making the jump) was already extremely likely. That is, a person could simply predict that pitchers will never make the jump to being good and be right greater than 96% of the time. So again what you would want here is a score better than what you’d achieve from guessing based on the observed probability. Brier Skill measures how much better as a percentage than a baseline rate (in this case, the guessing the base rates above). Using the fall model as an example, the Brier Skill score of 0.235 is like saying the model is 23.5% better than the baseline (measured by squared error, not necessarily based on how many more predictions are correct). Here too, our models all represent some improvement over the baseline.
Top-20% Capture serves as a complement to AUC. It orders the predicted observations based on probability, selects the top 20%, and measures what proportion of the “hits” are observed within that group. Higher is better and we are looking for a score in excess of 20%, which is what we’d get from random chance alone. In all, this constellation of metrics is a broadly positive sign for the utility of the models. They aren’t perfect, but given the relatively low sample size of good relief pitcher seasons, they all can help us in evaluating relievers going into 2027. I’ll get there if you’ll bear with me.
Back-test results
I’ll skip the statistics explainer because you just suffered through it and the metrics I used to compare the back-test results are the same. For each test, I cut off the sample data at 2015, and tested all but the fall model on each subsequent season. For the fall model, I tested it on seasons between 2016 and 2020, because I needed to leave a large enough buffer of years afterward to have sufficient confidence to call that player’s last good season in the data the last of their career. As an example, looking at the table from earlier, Raisel Iglesias’s last elite season was 2024. While it might not be likely, it also wouldn’t be a total shock if he managed to reach that bar in 2027. So the model would unfairly identify his 2024 as a “fall” when it might not be. There are only two players in the entire sample who have gap between good seasons (John Franco in 1988 and 1997; Todd Worrell in 1988 and 1995) larger than six so that is the buffer I set.
| Model | Algorithm | AUC | Brier | Brier Skill | Top-20% Capture |
|---|---|---|---|---|---|
| Fall | Elastic Net | 0.754 | 0.2066 | 0.183 | 37.1% |
| Fall | Random Forest | 0.733 | 0.2153 | 0.148 | 40.0% |
| Fall | Training Base Rate | 0.500 | 0.2528 | 0.000 | NA |
| Repeat | Elastic Net | 0.693 | 0.2176 | 0.067 | 26.7%* |
| Repeat | Random Forest | 0.682 | 0.2190 | 0.061 | 33.3% |
| Repeat | Training Base Rate | 0.500 | 0.2332 | 0.000 | NA |
| Good Jump | Elastic Net | 0.760 | 0.0313 | 0.039 | 56.6% |
| Good Jump | Random Forest | 0.716 | 0.0357 | -0.097 | 50.9% |
| Good Jump | Training Base Rate | 0.500 | 0.0326 | 0.000 | NA |
| Elite Jump | Elastic Net | 0.846 | 0.0194 | 0.088 | 75.7% |
| Elite Jump | Random Forest | 0.768 | 0.0217 | -0.019 | 64.9% |
| Elite Jump | Training Base Rate | 0.500 | 0.0213 | 0.000 | NA |
Anyway, for each metric, we have broadly similar results as the cross validation table. This is a terrific sign. Some cracks do start to show, though: the random forest jump models are not producing brier scores better than the baseline rate, and the repeat elastic net model’s top-20% capture rate has a lower confidence interval below 20%, meaning we can’t statistically distinguish it from that baseline value. Still, though, this is encouraging.
With this in mind, let’s dig a little deeper into the actual tiers. I’ve trimmed the random forest results from the presentation and focused primarily on the elastic net model, which I’ll do from this point because it shaped up to be the better method overall.
| Model | Tier | Sample Size | Mean Predicted | Outcome Rate |
|---|---|---|---|---|
| Fall | 1 of 4 (Highest) | 19 | 63.4% | 78.9% (15) |
| Fall | 2 of 4 | 19 | 44.2% | 47.4% (9) |
| Fall | 3 of 4 | 19 | 31.2% | 36.8% (7) |
| Fall | 4 of 4 (Lowest) | 18 | 11.1% | 22.2% (4) |
| Repeat | 1 of 4 (Highest) | 42 | 68.8% | 50.0% (21) |
| Repeat | 2 of 4 | 42 | 49.0% | 50.0% (21) |
| Repeat | 3 of 4 | 41 | 37.4% | 34.1% (14) |
| Repeat | 4 of 4 (Lowest) | 41 | 25.4% | 9.8% (4) |
| Good Jump | 1 of 5 (Highest) | 316 | 13.1% | 9.5% (30) |
| Good Jump | 2 of 5 | 316 | 5.4% | 3.2% (10) |
| Good Jump | 3 of 5 | 315 | 3.4% | 1.9% (6) |
| Good Jump | 4 of 5 | 315 | 2.3% | 1.3% (4) |
| Good Jump | 5 of 5 (Lowest) | 315 | 1.2% | 1.0% (3) |
| Elite Jump | 1 of 5 (Highest) | 340 | 7.6% | 8.2% (28) |
| Elite Jump | 2 of 5 | 340 | 2.1% | 1.5% (5) |
| Elite Jump | 3 of 5 | 340 | 1.2% | 0.3% (1) |
| Elite Jump | 4 of 5 | 339 | 0.7% | 0.6% (2) |
| Elite Jump | 5 of 5 (Lowest) | 339 | 0.3% | 0.3% (1) |
These results – especially for the elite jump model – are pretty good. The fall model shows clear differentiation in the tiering but underestimates the amount of falls in the top and bottom tiers. The repeat model does not have satisfactory differentiation at the top, meaning we can’t reliably identify the top two buckets of probabilities from each other. We knew this based on the capture rate noted above. But there is a clear differentiation in the bottom bucket, which means we can say that we have some idea of who won’t repeat. The half-elite jump model is differentiated in a manner consistent with predictions but overestimates across the board. Here, we can probably reasonably trust an ordinal ranking of players based on this model but may want to shift down our expectation on the actual probabilities. The elite jump model is demonstrably the best. The vast majority of hits are clustered in the highest-likelihood bucket. This means that we can draw a shortlist from that bucket and have some amount of confidence that a given season’s breakouts will be drawn from that group. The model produced a roughly 3x multiplier on the likelihood of a hit relative to a random draw, going from a 1 in 40 shot to a 1 in 13 and was basically correct. This is a terrific outcome.
Show me what you got
I went into this exercise trying to figure out an answer for “if not saves, then what?” Well, we can figure that out based on the model results. Here is a list of the variables which, if scrambled in the data, have the most deleterious impact on AUC. Similar stats (K%, K% change, associated lags, etc.) are grouped where applicable.
| Model | Stat Family | AUC Drop |
|---|---|---|
| Fall | Good History | 0.091 |
| Fall | Age | 0.053 |
| Fall | Dollar Value | 0.051 |
| Fall | Saves | 0.043 |
| Fall | Innings | 0.023 |
| Fall | K% | 0.018 |
| Fall | BB% | 0.011 |
| Fall | Prior RP Season | 0.001 |
| Fall | HR/9 | -0.003 |
| Repeat | K% | 0.089 |
| Repeat | Age | 0.075 |
| Repeat | Good History | 0.035 |
| Repeat | Dollar Value | 0.011 |
| Repeat | Prior RP Season | 0.010 |
| Repeat | BB% | 0.009 |
| Repeat | Saves | 0.002 |
| Repeat | Innings | -0.002 |
| Repeat | HR/9 | -0.003 |
| Good Jump | K% | 0.070 |
| Good Jump | Innings | 0.032 |
| Good Jump | Age | 0.031 |
| Good Jump | Saves | 0.026 |
| Good Jump | Dollar Value | 0.008 |
| Good Jump | HR/9 | -0.001 |
| Good Jump | Prior RP Season | -0.001 |
| Good Jump | BB% | -0.002 |
| Elite Jump | K% | 0.091 |
| Elite Jump | Saves | 0.047 |
| Elite Jump | Age | 0.040 |
| Elite Jump | Innings | 0.036 |
| Elite Jump | HR/9 | 0.004 |
| Elite Jump | Dollar Value | 0.003 |
| Elite Jump | Good History | -0.000 |
| Elite Jump | Prior RP Season | -0.002 |
| Elite Jump | BB% | -0.005 |
For the fall model, it turns out the be the culprits we’d most expect. Whether or not the player has a demonstrated history of being good already, age, performance in the current season, and saves, and innings are the top variable. These are intuitive – all else equal, a younger player is less likely to fall off the cliff. Likewise, a player who just barely clears the bar, and does so for the first time, might have just had his career year. The other models is where things start to get interesting – K% and its related derivatives matter the most in each case, and is much more important in the model than saves.
When making that statement, it’s important to note that some of the other variables – history, dollar value – include information derived from saves, so the comparison isn’t 100% clean. That being said, those stats also contain information from strikeouts too. In any case, the data clearly point to a few takeaways:
- Fall model: If you want to guess the relievers who will stay good, choose the ones that have a good history, role, and, all else equal, are younger.
- Repeat model: If you want to guess the relievers who won’t repeat, bet on skills (or lack thereof).
- Jump models: If you want to find the next breakout, bet on skills first, and look for a path to a role.
This, of course, is what many fantasy managers are already trying to do. So in this sense, there isn’t necessarily a magic bullet. And unlike my analysis from a couple months ago, because regression coefficients aren’t super interpretable to a lay audience, I can’t give you a pithy rule or threshold to sort by for flagging players. But I can give you the lists outright. Here is what the fall model thinks about whether or not 2026’s good relievers will ever have another good season:
| Rank | Name | Age | 2026 $ Value | 2026 IP | 2026 SV | 2026 K% | 2026 BB% | Good Seasons | Good Streak | Fall Probability |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Mason Miller | 27 | $27.69 | 70.0 | 38 | 47.4% | 9.6% | 3 | 3 | 6.5% |
| 2 | Cade Smith | 27 | $26.28 | 74.0 | 41 | 36.3% | 7.8% | 3 | 3 | 11.6% |
| 3 | Josh Hader | 32 | $15.48 | 44.2 | 26 | 32.9% | 9.0% | 8 | 4 | 21.0% |
| 4 | Jhoan Duran | 28 | $14.80 | 59.1 | 33 | 35.6% | 4.2% | 2 | 2 | 28.2% |
| 5 | Jacob Latz | 30 | $23.14 | 72.0 | 30 | 30.6% | 8.5% | 1 | 1 | 34.2% |
| 6 | Didier Fuentes | 21 | $11.08 | 73.1 | 3 | 31.0% | 7.1% | 1 | 1 | 34.5% |
| 7 | David Bednar | 31 | $18.54 | 69.1 | 35 | 27.1% | 7.1% | 3 | 2 | 38.7% |
| 8 | Louis Varland | 28 | $24.26 | 78.0 | 35 | 31.6% | 7.5% | 1 | 1 | 40.7% |
| 9 | Aroldis Chapman | 38 | $12.62 | 51.2 | 35 | 33.3% | 10.8% | 9 | 2 | 42.9% |
| 10 | Grant Taylor | 24 | $17.02 | 84.0 | 10 | 31.5% | 8.9% | 1 | 1 | 51.1% |
| 11 | Mason Montgomery | 26 | $16.33 | 68.2 | 18 | 37.8% | 9.4% | 1 | 1 | 55.6% |
| 12 | Trevor Megill | 32 | $13.03 | 61.2 | 30 | 28.5% | 6.2% | 2 | 2 | 56.6% |
| 13 | Bryan Baker | 31 | $21.32 | 62.1 | 41 | 25.8% | 9.0% | 1 | 1 | 72.5% |
| 14 | Paul Blackburn | 32 | $12.42 | 85.0 | 1 | 20.5% | 9.8% | 1 | 1 | 93.1% |
Remember, the exact probabilities aren’t super trustworthy, but the ordering was well differentiated, meaning we can have some amount of confidence that Mason Miller and Cade Smith are the least likely players to never register another good season, and that Paul Blackburn and Bryan Baker are the most likely. The top end of the list isn’t shocking, and you’d be hard-pressed to find a fantasy player who expects a Blackburn repeat, but Baker’s place on this list is more conspicuous. He’s older, doesn’t have a long history of success, and his skills – while good – are clearly a tier below the rest of the group. Here, the model is giving us a clear fade signal.
Among the projected repeaters, and again – this model’s strength is going to be in the bottom of the list – we see something similar:
| Rank | Name | Age (2027) | 2026 $ Value | 2026 IP | 2026 SV | 2026 K% | 2025 K% | Good Streak | Repeat Probability |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Mason Miller | 28 | $27.69 | 70.0 | 38 | 47.4% | 44.4% | 3 | 80.8% |
| 2 | Cade Smith | 28 | $26.28 | 74.0 | 41 | 36.3% | 34.7% | 3 | 66.7% |
| 3 | Didier Fuentes | 22 | $11.08 | 73.1 | 3 | 31.0% | NA | 1 | 56.0% |
| 4 | Josh Hader | 33 | $15.48 | 44.2 | 26 | 32.9% | 36.9% | 4 | 56.0% |
| 5 | Jhoan Duran | 29 | $14.80 | 59.1 | 33 | 35.6% | 27.9% | 2 | 50.9% |
| 6 | Jacob Latz | 31 | $23.14 | 72.0 | 30 | 30.6% | NA | 1 | 45.3% |
| 7 | Louis Varland | 29 | $24.26 | 78.0 | 35 | 31.6% | 25.3% | 1 | 44.2% |
| 8 | Mason Montgomery | 27 | $16.33 | 68.2 | 18 | 37.8% | 30.1% | 1 | 44.1% |
| 9 | Grant Taylor | 25 | $17.02 | 84.0 | 10 | 31.5% | 34.4% | 1 | 42.2% |
| 10 | Aroldis Chapman | 39 | $12.62 | 51.2 | 35 | 33.3% | 37.3% | 2 | 37.3% |
| 11 | David Bednar | 32 | $18.54 | 69.1 | 35 | 27.1% | 34.3% | 2 | 34.7% |
| 12 | Bryan Baker | 32 | $21.32 | 62.1 | 41 | 25.8% | 30.3% | 1 | 25.3% |
| 13 | Trevor Megill | 33 | $13.03 | 61.2 | 30 | 28.5% | 31.2% | 2 | 24.4% |
| 14 | Paul Blackburn | 33 | $12.42 | 85.0 | 1 | 20.5% | 19.4% | 1 | 7.8% |
The jumpers are where things get really interesting, because this is the area with the strongest opportunity for profit. First, the top 20% of the “good” jump probabilities. The ordering and proportional difference of these probabilities is fairly decent, but are probably running a touch hot:
| Rank | Name | Age (2027) | 2026 $ Value | 2026 IP | 2026 SV | 2026 K% | 2025 K% | Good Jump Probability |
|---|---|---|---|---|---|---|---|---|
| 1 | Riley O’Brien | 32 | $3.64 | 64.0 | 38 | 23.2% | 22.6% | 17.3% |
| 2 | Aaron Ashby | 29 | $8.48 | 89.1 | 4 | 30.3% | 28.1% | 13.9% |
| 3 | Bradgley Rodriguez | 23 | $3.80 | 74.0 | 2 | 26.9% | 29.0% | 12.1% |
| 4 | Hogan Harris | 30 | -$2.26 | 70.1 | 17 | 28.3% | 23.5% | 10.7% |
| 5 | Erik Sabrowski | 29 | -$0.04 | 48.1 | 1 | 37.4% | 34.7% | 9.7% |
| 6 | Gregory Soto | 32 | $3.85 | 67.0 | 16 | 27.7% | 25.1% | 9.2% |
| 7 | Yoendrys Gómez | 27 | -$6.58 | 67.0 | 22 | 21.9% | 20.9% | 9.1% |
| 8 | Hagen Smith | 23 | -$2.26 | 28.0 | 0 | 37.3% | NA | 8.8% |
| 9 | Abner Uribe | 27 | $6.47 | 59.1 | 6 | 26.5% | 30.2% | 8.7% |
| 10 | Brent Headrick | 29 | $5.05 | 78.1 | 0 | 30.1% | 32.6% | 8.5% |
| 11 | Ryan Zeferjahn | 29 | -$3.90 | 69.2 | 7 | 30.9% | 28.4% | 8.5% |
| 12 | Bryan Abreu | 30 | -$2.37 | 62.2 | 9 | 29.9% | 35.5% | 8.4% |
| 13 | Edgardo Henriquez | 25 | $0.59 | 60.0 | 2 | 29.3% | 23.1% | 8.3% |
| 14 | Brandyn Garcia | 27 | $2.58 | 51.1 | 5 | 31.2% | 21.2% | 8.3% |
| 15 | Juan Morillo | 28 | $4.19 | 64.1 | 3 | 28.0% | 23.7% | 7.5% |
| 16 | AJ Blubaugh | 26 | -$4.83 | 96.0 | 1 | 23.4% | 28.2% | 7.1% |
| 17 | Yohan Ramírez | 32 | $3.87 | 85.0 | 2 | 28.2% | 29.0% | 6.9% |
| 18 | Jack Dreyer | 28 | $3.97 | 64.2 | 2 | 28.0% | NA | 6.7% |
| 19 | Erik Miller | 29 | -$3.94 | 51.1 | 5 | 32.6% | 17.2% | 6.5% |
| 20 | Garrett Whitlock | 31 | $8.73 | 52.1 | 4 | 29.8% | 31.1% | 6.3% |
| 21 | Tyler Wells | 32 | $8.33 | 72.0 | 6 | 26.0% | NA | 6.2% |
| 22 | Samy Natera Jr. | 27 | -$6.75 | 31.2 | 3 | 36.6% | NA | 6.0% |
| 23 | Elvis Alvarado | 28 | -$3.94 | 55.1 | 6 | 28.4% | 27.3% | 6.0% |
| 24 | Pete Fairbanks | 33 | -$6.36 | 41.1 | 19 | 31.1% | 24.2% | 5.9% |
| 25 | Luis Medina | 28 | -$4.04 | 73.0 | 0 | 29.3% | NA | 5.8% |
Finally, the elite jumpers:
| Rank | Name | Age (2027) | 2026 $ Value | 2026 IP | 2026 SV | 2026 K% | 2025 K% | Good Seasons | Elite Jump Probability |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Didier Fuentes | 22 | $11.08 | 73.1 | 3 | 31.0% | NA | 1 | 17.1% |
| 2 | Trevor Megill | 33 | $13.03 | 61.2 | 30 | 28.5% | 31.2% | 2 | 11.3% |
| 3 | Bradgley Rodriguez | 23 | $3.80 | 74.0 | 2 | 26.9% | 29.0% | 0 | 9.0% |
| 4 | Hagen Smith | 23 | -$2.26 | 28.0 | 0 | 37.3% | NA | 0 | 8.8% |
| 5 | Riley O’Brien | 32 | $3.64 | 64.0 | 38 | 23.2% | 22.6% | 0 | 7.5% |
| 6 | Aaron Ashby | 29 | $8.48 | 89.1 | 4 | 30.3% | 28.1% | 0 | 7.5% |
| 7 | Yoendrys Gómez | 27 | -$6.58 | 67.0 | 22 | 21.9% | 20.9% | 0 | 6.5% |
| 8 | Erik Sabrowski | 29 | -$0.04 | 48.1 | 1 | 37.4% | 34.7% | 0 | 6.5% |
| 9 | Ryan Zeferjahn | 29 | -$3.90 | 69.2 | 7 | 30.9% | 28.4% | 0 | 6.1% |
| 10 | Hogan Harris | 30 | -$2.26 | 70.1 | 17 | 28.3% | 23.5% | 0 | 6.0% |
| 11 | Edgardo Henriquez | 25 | $0.59 | 60.0 | 2 | 29.3% | 23.1% | 0 | 6.0% |
| 12 | Brent Headrick | 29 | $5.05 | 78.1 | 0 | 30.1% | 32.6% | 0 | 5.6% |
| 13 | AJ Blubaugh | 26 | -$4.83 | 96.0 | 1 | 23.4% | 28.2% | 0 | 5.5% |
| 14 | Abner Uribe | 27 | $6.47 | 59.1 | 6 | 26.5% | 30.2% | 0 | 4.7% |
| 15 | Luis Medina | 28 | -$4.04 | 73.0 | 0 | 29.3% | NA | 0 | 4.3% |
| 16 | Bryan Abreu | 30 | -$2.37 | 62.2 | 9 | 29.9% | 35.5% | 0 | 4.3% |
| 17 | Gregory Soto | 32 | $3.85 | 67.0 | 16 | 27.7% | 25.1% | 0 | 4.1% |
| 18 | Yohan Ramírez | 32 | $3.87 | 85.0 | 2 | 28.2% | 29.0% | 0 | 3.9% |
| 19 | Justin Martinez | 25 | -$7.44 | 13.1 | 0 | 32.7% | 32.8% | 0 | 3.8% |
| 20 | Brandyn Garcia | 27 | $2.58 | 51.1 | 5 | 31.2% | 21.2% | 0 | 3.8% |
| 21 | Samy Natera Jr. | 27 | -$6.75 | 31.2 | 3 | 36.6% | NA | 0 | 3.8% |
| 22 | Alex Vesia | 31 | $0.01 | 58.0 | 3 | 31.2% | 33.8% | 1 | 3.7% |
| 23 | Anthony Nunez | 25 | -$7.81 | 59.2 | 3 | 25.9% | NA | 0 | 3.6% |
| 24 | Jack Dreyer | 28 | $3.97 | 64.2 | 2 | 28.0% | NA | 0 | 3.6% |
| 25 | Mason Fluharty | 25 | -$2.02 | 62.2 | 1 | 24.7% | 26.0% | 0 | 3.6% |
| 26 | Juan Morillo | 28 | $4.19 | 64.1 | 3 | 28.0% | 23.7% | 0 | 3.6% |
| 27 | Jeff Hoffman | 34 | -$0.53 | 66.0 | 6 | 34.9% | 29.3% | 1 | 3.5% |
| 28 | Erik Miller | 29 | -$3.94 | 51.1 | 5 | 32.6% | 17.2% | 0 | 3.4% |
This is probably the most valuable nugget of the research. Here, you have a short-list of relievers who might make the jump to elite performance. If Didier Fuentes stays in his relief role, I’d be buying everywhere, especially if the Braves don’t retain Iglesias. It’s also interesting that Trevor Megill is near the top of this list but the bottom of the others, but not necessarily inconsistent. An 11% likelihood to break out is not mutually exclusive from a low chance to repeat his existing level, or a higher chance to never be that good again. Of course, the outcomes are – one cannot break out next year and also not be good – but here we have an example of a high-risk, high-reward range of possibilities.
I also like this list because it lets you choose your own adventure based on what you personally value. Do you prefer to lean on saves in your dart throws? Trevor Megill and Riley O’Brien might cost too much in drafts, so go for Hogan Harris, or Yoendrys Gomez. Prefer age as a separator? Look at Fuentes, Hagen Smith, or Bradgely Rodriguez. Raw stuff? Smith and Erik Sabrowski are racking up tons of Ks. There’s a little something for everyone here.
Honk if you hate saves
Before I leave you, dear reader, I want to return to saves. To prove that they weren’t all that useful in the jump models, I refit them but excluded saves from the pool of variables that they could use. Predictably, the results were a bit worse but broadly similar:
| Model | Algorithm | AUC | Brier | Brier Skill | Top-20% Capture |
|---|---|---|---|---|---|
| Good Jump (No Saves) | Elastic Net | 0.751 | 0.0355 | 0.051 | 56.7% |
| Good Jump (No Saves) | Random Forest | 0.715 | 0.0359 | 0.038 | 49.1% |
| Good Jump (No Saves) | Training Base Rate | 0.500 | 0.0374 | 0.000 | NA |
| Elite Jump (No Saves) | Elastic Net | 0.819 | 0.0215 | 0.060 | 69.0% |
| Elite Jump (No Saves) | Random Forest | 0.747 | 0.0223 | 0.025 | 55.3% |
| Elite Jump (No Saves) | Training Base Rate | 0.500 | 0.0228 | 0.000 | NA |
Same goes for the back-test:
| Model | Algorithm | AUC | Brier | Brier Skill | Top-20% Capture |
|---|---|---|---|---|---|
| Good Jump (No Saves) | Elastic Net | 0.737 | 0.0324 | 0.007 | 50.9% |
| Good Jump (No Saves) | Random Forest | 0.677 | 0.0379 | -0.163 | 41.5% |
| Good Jump (No Saves) | Training Base Rate | 0.500 | 0.0326 | 0.000 | NA |
| Elite Jump (No Saves) | Elastic Net | 0.825 | 0.0200 | 0.060 | 73.0% |
| Elite Jump (No Saves) | Random Forest | 0.726 | 0.0227 | -0.064 | 45.9% |
| Elite Jump (No Saves) | Training Base Rate | 0.500 | 0.0213 | 0.000 | NA |
And the back-test tiers:
| Model | Algorithm | AUC | Brier | Brier Skill | Top-20% Capture |
|---|---|---|---|---|---|
| Good Jump (No Saves) | Elastic Net | 0.737 | 0.0324 | 0.007 | 50.9% |
| Good Jump (No Saves) | Random Forest | 0.677 | 0.0379 | -0.163 | 41.5% |
| Good Jump (No Saves) | Training Base Rate | 0.500 | 0.0326 | 0.000 | NA |
| Elite Jump (No Saves) | Elastic Net | 0.825 | 0.0200 | 0.060 | 73.0% |
| Elite Jump (No Saves) | Random Forest | 0.726 | 0.0227 | -0.064 | 45.9% |
| Elite Jump (No Saves) | Training Base Rate | 0.500 | 0.0213 | 0.000 | NA |
Here are the players whose outlooks change the most based on model:
| Name | Age (2027) | Rank (Saves) | Rank (No Saves) | Rank Change Without Saves | Half-Elite Jump Probability (Saves) | Half-Elite Jump Probability (No Saves) |
|---|---|---|---|---|---|---|
| Yoendrys Gómez | 27 | 7 | 47 | -40 | 9.1% | 3.5% |
| Pete Fairbanks | 33 | 24 | 62 | -38 | 5.9% | 2.6% |
| Riley O’Brien | 32 | 1 | 35 | -34 | 17.3% | 5.0% |
| Steven Okert | 35 | 53 | 25 | +28 | 3.3% | 6.2% |
| Hogan Harris | 30 | 4 | 24 | -20 | 10.7% | 6.3% |
Unsurprisingly, most of this list relied on saves to boost their odds in the original model.
| Name | Age (2027) | Rank (Saves) | Rank (No Saves) | Rank Change Without Saves | Elite Jump Probability (Saves) | Elite Jump Probability (No Saves) |
|---|---|---|---|---|---|---|
| Riley O’Brien | 32 | 5 | 51 | -46 | 7.5% | 1.9% |
| Paul Blackburn | 33 | 61 | 16 | +45 | 1.4% | 4.8% |
| Yoendrys Gómez | 27 | 7 | 47 | -40 | 6.5% | 2.0% |
| Garrett Whitlock | 31 | 46 | 23 | +23 | 1.9% | 4.1% |
| Hogan Harris | 30 | 10 | 31 | -21 | 6.0% | 3.2% |
For the elite jumps, the story is the same.
Conclusion
I can’t help but end this feeling somewhat like Stu Pickles:

In any case, I hope this has been illustrative of a few things: predicting fantasy baseball is hard, and the edges you can glean are small. But with a sufficiently targeted scope, it can bear fruit. In general, I like this framework for fantasy analysis because it can be applied to just about anything if you ask the question right. I can revisit my k-rate analysis and look for ways to improve or add to it by including a broader history of variables to calibrate more precise player-specific likelihoods of success. I can look for busts or breakouts for other position groups. I can probably even tweak these models to be more effective. It’s an exciting frontier, and is something that I’ll be adding to the docket for what is becoming an increasingly busy off-season plan.
Jonathan is a contributor for RotoGraphs. He is a Tigers fan living in Philadelphia with his wife and dog and requests that you leave your best pizza topping combinations in the comments.


I didn’t read the article. Relievers bore me. But I loved the title. I thought you had to be over 65 to remember on e of the great under appreciated classics Like Dr Strangelove.
there’s no fighting in the WAR room!