Is Drafting Relievers Worth It?

“I think you know my opinion on RP.”

• Trusted leaguemate and fellow fantasy writer Sam Lutz

Intro

For several weeks now, I’ve written articles poking around with how ADP and projections interact with each other. And against my wishes, relief pitchers have popped up in some way each time:

After finding last week that relievers are a main driver of the advantage projection-based ranks hold over ADP, I got to thinking: do we even need them? Or, more precisely – do we really need to spend good picks on them?

This is a tricky question to tackle, and its answer involves detangling a number of sub-questions, specifically:

• Is the dollar value of high-drafted relievers declining over time?

This question is straightforward. If the answer is yes, it helps bolster the case that we shouldn’t be spending material draft capital on relievers. If the answer is no, it tells us that the issue is more nuanced. Either answer requires further digging, with a question like:

• How well does ADP actually predict successful reliever seasons compared to other positions?

Here we will get a sense of whether the market is actually good at identifying high-performing relievers. If the answer is “not very well,” then it’s a hint that our draft resources might be better spent elsewhere, which can likewise be pinned down by asking:

• How do relievers perform throughout the draft relative to other positions?

This touches on the notion of opportunity cost. If reliever performance in the draft declines quickly, it means we might still want to pull the trigger with earlier picks. If it doesn’t, then it provides more support to eschewing early relievers.

Once we know the answers to the three questions above, we can build a recommended path of action for what to do about it.

Data construction and analysis framework

Building on my earlier analyses, I used historical season and Steamer projections data from FanGraphs and ADP data from FantasyPros dating back to 2015 (the earliest year of ADP data available). I adapted the code I used last week (which considers 12-team, 23-man rosters with a standard 5×5 roto setting) and used mostly the same methodology to build dollar values, with a couple of differences.

First, I removed the ADP pool cutoff when sizing my player pools. This was necessary because undrafted players matter here. If our goal is eventually to identify late relievers whom we want to throw darts at, we want to draw from all the players available for a given year.

Second, I changed the position adjustment pool sorting mechanism for projected dollar values to match what I did for the actual dollar values. After a period of reflection on the methodology, my mind has changed to think that ranking based on the initial pass of z-scores is more robust. It mirrors the way I calculated the actual dollar value for each season, more closely matches what the Auction Calculator does, and, ultimately, we are creating a ranking based on what the projections say, so I think that the projections should be what determines the order for how position adjustments get calculated.

Lastly, I changed how the position adjustment was calculated for UT-only players as well as pitchers. In my previous code, UT-only players received their own adjustment. While working on this, though, I realized that in the Auction Calculator they are assigned the first base adjustment so I mirrored that. For pitchers, I created a separate position adjustment for RPs and SPs. Notably, I did not calibrate the z-scores for each adjustment within each pool – this would drastically overrate relievers because they wouldn’t be getting penalized for their lower innings totals. Meanwhile, starting pitchers with a few nominal saves would be getting a huge boost to their values. To navigate this, I simply took the 96 best (12 teams times eight pitching slots) pitchers, calibrated the z-scores for the categories, and then used the worst SP and the worst RP from this pool to generate the position adjustment.

Now, you might ask – “Wait, if you’re going to look at ADP versus year-end performance, why do you need to bother creating your own auction values instead of just using Auction Calculator data?” And I would say fair enough! However, there is some other analysis coming up later which requires projected dollar values and because Auction Calculator does not create values for historical projections, I wanted my actual dollar value representation to be as apples-to-apples with the projection version as possible.

Once I had the auction values constructed, I nonetheless wanted to sanity check my results against the Auction Calculator export. To this end, I was generally quite happy with the result:

Overall, my results were very highly correlated across each year of the sample so I took this to be a reasonable affirmation that my process to recreate the dollar values was sound. I then attached the ADP data for each year and set about answering my lines of inquiry outlined above.

Question 1: Is the dollar value of high drafted relievers declining over time?

This, comparatively, is easy to answer. I separated players into 75 pick buckets up to 300 (with a 300-plus bucket containing all players with ADP higher than that, inclusive players who were undrafted, meaning there was no ADP data for that player in that year). I then looked at how relievers in the top two buckets (1-75, 76-150) have performed over time:

There is a lot of year-to-year noise, including years where the lower ADP bucket performs similarly or better than the higher bucket, but overall, the trends here are flat. To my eye, it seems that no, the value of top relievers is not materially declining over time.

One caveat I will flag as something that could be looked at in future analysis is that this doesn’t say anything about the specific composition of relievers making up these pools. It could well be that the earlier years have a more stable year-to-year base of relievers constituting the higher ADPs and that later years, while still roughly as successful on net, are churning more. This would be useful analysis but was a bit outside the scope of this specific project. So, moving on.

Question 2: How well does reliever ADP predict successful reliever seasons relative to other seasons?

This is where things get tricky. In my first pass, I simply correlated (using Spearman correlations) ADP and year-end value for each of these positions:

While it seems that ADP is demonstrably less predictive for pitchers as a whole, it is not meaningfully worse for relievers relative to starters. This splits the difference in terms of an answer to the question.

One flaw that the above analysis has though is that it’s not factoring in the performance of undrafted players in the data. This effect was harder to capture because relying on an ordinal rank like the above correlations isn’t possible for undrafted players. To get around this, I created a new variable based on my 75-pick buckets of ADP, and included undrafted players as a final bucket. In this new variable, the top 75 players receive a “1” for their value, while the next 75 receive “2”, and so on. This isn’t perfect, but it allows us to attach an ordinal (if blunt) rank to undrafted players for correlation analysis, and the results were encouraging:

Here, each position group is at least a bit worse off, which makes sense. Collapsing ADP into less granular buckets will introduce some noise into the process. Fortunately though we have enough of a sample to still derive useful results. RPs are the most effected compared to the prior correlation, and there is some distance now between the correlation for RPs and SPs.

To validate this, I sought another means of measuring the predictive success of a reliever’s ADP. Instead of just comparing against year-end value, I measured how much a player’s ADP bucket in a given year was correlated with their ADP bucket in the year prior. To this extent, ADP in the subsequent year represents a reasonable proxy for “year-end performance” of the previous season. Here too, the results support the notion that RPs are just a bit less reliable:

I took this constellation of outcomes to indicate an overall affirmative answer to this question. It does seem that ADP is not especially reliable, relative to other position groups, for relievers.

Question 3: How do relievers perform throughout the draft relative to other positions?

Unlike Question 2, this is another comparatively straightforward question to address. Here again I leveraged 75-pick buckets and looked at how relievers fare relative to other positions:

Finally, we have some hot tea! Relievers are the worst performing bunch in the earliest picks – more fodder to help answer our overall question – but also have the flattest rate of decline throughout the rest of the draft. Hitters in particular suffer steeply once you leave the early rounds.

Sorry Mason Miller, but this is a strong nudge to the overall notion that there may be a more optimal way to spend early picks than on relief pitchers.

Decisions, decisions

Now we’ve come to a reasonable conclusion: spending an early pick on a reliever is riskier, has an overall lower return on investment, and doesn’t appear to be meaningfully more beneficial than waiting until later in the draft. What can we do about it though?

Outright punting on relievers would not be my recommendation. It might work for you, but I think there is still a needle to thread – relief pitchers in the 76-150 bucket are still pretty good and based on my analysis from last week, there are some gems which could be unearthed late.

To this end, I leveraged a decision tree model to identify some traits which can reliably predict higher-value relief pitchers out of those drafted in later rounds. You can refer to the link for a more detailed definition, but a layman’s explanation for how a decision tree works is that for a given target outcome, it iterates over a series of potential explanatory variables to find which traits are the most meaningfully associated with that desired outcome.

My target outcome here was a relief pitcher season exceeding $10 in auction value. This is just slightly below the average value of a reliever chosen in the top 75 picks as noted before. As a round and easily interpretable number, I felt it served these purposes well. My first step was to figure out what the blind hit rate for each bucket would be to create a baseline which the model would need to exceed:

To help interpret this table, for a relief pitcher in the 76-150 bucket, we would need to exceed a coin-flip. There have been 123 RPs chosen in that range over time and 63 of them eventually returned a $10 season.

Full transparency here – this is my first foray into any sort of machine learning, so I was learning as I went along. The model I used was the CART (classification and regression tree) model, which is a part of the base package in R. This is among the simpler versions of these types of models that exist. That said, I think it still proved pretty useful for our purposes.

For this analysis, the explanatory variables I used were the following:

IP, SV, HLD, WAR, FIP, ERA, WHIP, K/9, BB/9, K/BB, HR/9, K%, BB%, K-BB%, AVG, BABIP, LOB%, age, projected dollar value, projected rank, ADP rank change, ADP rank

Any regular stat above is coming from Steamer projected values. The latter four are derived through my data builds. Projected dollar value was the product of the build I mentioned earlier, while the projected rank is a sorted rank based on those outputs. ADP rank change measured the change in ADP relative to the prior year, and ADP rank is simply a rank based on sorted ADP. There are surely more that could be used in a different model run, but I felt that this was overall a good start.

I set the model to focus on the 76-150 and 300+ pools. The idea here is that you might be able to replace an RP chosen in the top of the draft with one in the middle and a few targeted dart throws or post-draft waiver adds.

Drafters love this one weird trick!

First, for players in the 76-150 ADP range, the model landed on one useful split: projected rank. This is a nice complement to my analysis from articles in recent weeks, and one that arose completely independent of any authorial intent. Players with a projected rank better than 140 had a 68% hit rate – a meaningful improvement over the roughly coinflip odds noted above. And, crucially, higher even than the hit rate of RPs in the highest ADP bucket!

I acknowledge that it may seem obvious at first (better players are better), but as noted in prior weeks, ADP doesn’t always sort players in the way that projections would. Below are the players from 2024 and 2025 between ADPs of 76 and 150, delineated by the projection-based rank:

Here, I sorted by projected rank, and marked players who were “hits” in green and players near the threshold in yellow. You’ll notice that ADP is bouncing all over up and down the range.

2025 featured fewer relievers within that ADP bucket but those above the top 140 projected rank still performed relatively well. On the other hand, 2026 has been a mixed bag so far:

The overall sample is small, and among this group whose projected rank is better than 140, we only have one hit out of six based on YTD value. To this point though, I’d say that 1.) this framework still would have helped you avoid the three worst performing RPs from the bucket (Helsley, Pagán, and Estévez) and 2.) we haven’t quite finished the season and these numbers could look a lot different by the end of the year.

Among the deep pool, a 1.7% hit rate is not a meaningfully high bar to clear. The model found that players who projected for at least one hold (technically, greater than 0.88425) and whose WHIP is projected to be lower than 1.28 have a 6.8% hit rate. That’s pretty good! The cutoffs make sense, too – this is essentially flagging relievers with a defined late inning role (the holds threshold) and who project to not be terrible (WHIP).

Here are the players who met the dart throw criteria heading into 2026. I’ve included ADP rank and projected rank to show that there isn’t a clear signal to be derived from either, though that is already implicit given the decision tree didn’t pick it up.

That’s four hits (and one very, very near miss) in 46 chances for an 8.7% hit rate, just slightly better than our expectation.

As a further check, I looked at whether there was any combination in the top 75 sample to make those choices more efficient. It turned out that there wasn’t, probably partially due to the small sample size, and partially because a baseline 61% hit rate is going to be tough to meaningfully beat in any scenario.

I also wanted to look at the hit rates of other positions, and found results that are consistent with what we saw in the overall value charts provided earlier in the article:

You might notice that the bucket splits here are different. Because we are now talking about strategy in an actual draft, I wanted to more closely mimic a draft for the hypothetical league we’re analyzing.

The first key split is in the top 75 bucket. I found that the earliest ADP for any reliever in the sample was 37, so when comparing hit rates for relievers with the idea that you would replace one with a hitter or an SP, the fairer comparison would be against hit rates of a pool that begins at that point.

I also pooled the 151-276 to represent roughly the back half of drafts, and pooled 241 and beyond inclusive of undrafted players to represent the players available in roughly your last three rounds of the draft.

As you can see, the hit rates for hitters drops off much more steeply than for relievers. Notably, SP hit rates stay relatively high, and are even roughly 8% for the worst bucket, a finding that is certainly worth some future analysis.

What are the chances?

At this point, we have most of the puzzle put together. We’ve determined that relievers are in all likelihood worse investments, both in terms of realized value and the opportunity cost relative to other positions, and also identified a few very straightforward filters that can be introduced into projected data to meaningfully improve your chances at finding a good reliever later in drafts.

When thinking about the broader implications of draft strategy, I also wanted to look at the value of a typical pick in this range to put these probabilities into a more concrete example of the trade-off being made.

Again building off of some prior analysis, I built a pick value model using spline regressions. This time, because I had the full run of historical projection data, I did use the projected values to stay apples-to-apples in the comparisons we’re doing. Using the modified pick bucket structure from the previous section, here are the average dollar values associated with each bucket:

Based on the hit rates we identified earlier, you can spend half the draft capital on a mid-round reliever and still exceed the odds of finding a good one than you would have earlier in the draft.

To conduct a slightly more nuanced exercise, consider two additional scenarios:

• Option A: Use random draws to make one pick in the 37-75 range on a relief pitcher (61% hit rate), one pick in the ADP range of 76-150 on a hitter (28.4% hit rate), and two targeted RP waiver adds (6.8% hit rate)
• Option B: Use a random draw to pick a hitter in the 37-75 range (46.6% hit rate), one targeted relief pitcher in the 76-150 range (68% hit rate) and two targeted RP waiver adds (6.8% hit rate)

Each will cost roughly $35 in expected value based on those picks. Option A results in an expectation of 1.03 hits and a 24.6% probability of at least two hits. Option B results in an expectation of 1.28 hits 38.5% probability of at least two.

We should all be feeling like Kronk, because these are meaningful increases. The number of hits expected in Option B goes up by roughly 25%, while your odds at a double hit have larger than a 1.5x multiplier. To mix my meme metaphors, this was quite literally free real estate.

Conclusion

There are a few caveats to discuss when thinking about the above analysis; one is that the hit rate analysis treats all hits above $10 the same. In other words, a $25 player counts the same in this framework as a $10 player. This is obviously not the case in real life. Another is that due to the non-linearity of pick value across the drafts, flattening them when comparing hit rates and average values might produce distorted results. This, though, is an unavoidable trade off. If I go too granular, sample sizes would become too small to draw from. Meanwhile, going too large exacerbates the problem I just described. Overall, despite these issues, I believe that this nonetheless constitutes an interesting and exciting analysis.

If nothing else, it represents an incentive for future research to optimize the buckets a bit more precisely, or more accurately account for “big” hits. That isn’t all, though; this framework could easily be replicated for different league or scoring settings, could springboard a better quantification of the expected dollar value of a given draft choice, or provide the opportunity to apply a decision tree framework to more than just relievers as we think about draft strategy. I’m sure that there are likewise many others that I haven’t named, and I can’t help but conclude this article feeling like I’ve found a grove of (relatively) low hanging fruit. It should be fun to see what else we can tease out of these data over the coming weeks and months.