Debut K-Rate and Fantasy Viability: A Study

Our esteemed RotoGraphs editor recently passed along a question to the team that came to him from a reader:

“If a player runs a high k-rate upon debut for a significant amount of time, how likely is it that they’ll ever be a viable fantasy hitter?”

While I may have gone a bit overboard in doing so, the analysis below endeavors to find an answer to this question. I have not broken new analytical ground regarding the overall notion (TL;DR: strikeouts bad), but the extent to which k-rate impacts the likelihood to be even minimally fantasy viable is both stark and robust. If you are looking for a quick heuristic about whether or not you can move on from a prospect after they spend some time in the majors, you might have found it.

Data:

To conduct this analysis, I used Retrosheet data for play-by-play outcomes dating back to 1996 to calculate a player’s k-rate over a given period of his first plate appearances. I also exported dollar values from the FanGraphs Auction Calculator over the same time range. I conducted my analysis in R, and used my Mac’s command terminal to scrape the Retrosheet data and parse the event codes ^[1]. Once I had my data, I began setting about the contours of the analysis.

Framework:

To answer this question, I built a series of logistic regression models. I won’t bore you with the gory details, but the basic question a logistic regression helps answer (after some math) is the impact on the likelihood of an event occurring given certain inputs. In our case, the event (dependent variable, or regressand) is whether a player was ever fantasy viable. The input (independent variable, or regressor) is k-rate upon debut. I designed the models in three basic categories: one using a continuous regressor, one using categorical buckets for k-rates in increments of 5% as the regressor, and one using a binary yes/no indicator for whether or not the player’s initial k-rate exceeded 35%. Using each of these frameworks, I analyzed three different plate appearance samples: 30, 60, and 120 (60 being the point at which K-rates reasonably stabilize, 30 and 120 being half and double that). I will explain each of these models in greater detail along with the intuition of the results.

Assumptions:

There are two major branches of assumptions involved here: sample period, and the definition of what constitutes “fantasy viable.” For the latter, I chose players based on a default 5×5 roto with 12 teams, a $260 budget, and 23-player rosters. Using these settings, I considered players who had at least one season in their careers which met four separate value thresholds: $1, $5, $10, and $20. In other words, I identified players who all had a season worth at least $1. Then, $5. And so on. That the lowest end of this scale is not a high bar to clear is by design – in our case, using a blunt tool helps illustrate the point cleanly. Yes, the results of an analysis like this could naturally vary somewhat based on the scoring settings you choose. Players who strike out more will likely fare worse in points leagues whose settings penalize strikeouts. Likewise, the impact may be muted in leagues which reward traits that tend to be correlated with strikeouts such as high walk rates and more power. But overall, given that 5×5 12-team roto is like the vanilla ice cream of fantasy baseball, I think the results are, much like vanilla ice cream, broadly generalizable.

For my sample period, I chose data from 1996-2025 ^[2] because it would give me an even 30 years of outcome data. Now, you may look at that sample period and think “Ah – but the k-rates have been rising since then! Surely that will impact your results.” And you’d be right.

To this, I have an answer: controlling for each year. Year controls introduce each year as a regressor in the model and theoretically strip the impact of that year from the results on the regressors we care about. Now, in order for these controls to work, I needed to make additional choices. I had to choose a cutoff date for debutants because in order for my year controls to not account for the impact of truncation (i.e., the notion that a player who debuted in 2025 has simply had fewer bites at the apple), I needed to go far enough back that players had a reasonable chance to accrue a viable season. Had I not, the year control would be, in effect, penalizing players who debuted in later years.

To determine a reasonable cutoff, I looked at how long it typically took a player to have a viable fantasy season:

For all players who had a season viable at the $1 level, roughly 95% of them did so by the end of the sixth season after their debut. Based on this, I limited my sample to players who had debuted no later than 2019, meaning all players had reached at least six years post-debut by 2025, the latest full season available in my sample. I like this for two reasons: one is that six seasons post debut represents roughly two standard deviations above the mean in this sample of data. The second is that six years post debut lines up neatly with current service time rules. In the analysis realm, you don’t always get a marriage of rock-solid quantitative and qualitative rationales, so you have to take advantage of the opportunities when they arise.

With this in mind, I’ll next move on to the models and how to frame their results. For each, I’m going to present results from the 60 plate appearance sample. The results across each sample are broadly similar, and I’ll discuss a few noteworthy differences as they occur ^[3].

Continuous K-Rates:

The graph below models the probability of being fantasy viable for every point along the k-rate spectrum between 5% and 50%. This helps verify our basic suspicion: as k-rates increase, a player becomes less likely to ever register a viable season.  

Notably, given the curved slopes, the effect isn’t constant across the spectrum of k-rates. In the graph below, the y-axis depicts the incremental impact on the probability of being viable based on where a player is on the k-rate spectrum. Effects are more pronounced earlier on the x-axis – the incremental impact of striking out 40% of the time relative to 39% is less damaging than, say, the impact of striking out 30% vs. 29%.

One potential flaw in the continuous model, though, is that any extent to which the relationship is weaker on one side of the distribution than the other is obscured. In other words, there is a chance that k-rates don’t matter statistically until they reach a certain threshold, at which point they become very important. This will create an “averaging out” effect, where a weaker impact earlier in the spectrum is overstated because of a strong impact at the end, and vice versa. 

K-Rate Buckets:

To solve this problem, I examined how likely a player is, given their place in a certain k-rate bucket, to ever be fantasy viable. My buckets start at zero and increase in increments of five percent up to the maximum in the sample. This allows me to both climb the ladder of k-rates like the continuous analysis above and also isolate whether or not probabilities at specific k-rates are statistically significant. A brief math note here: because of how these regressions work, I had to choose a “baseline” category. This matters in how the equation for converting the results to probabilities is set up but is otherwise not influencing the actual probabilities themselves. I chose the 5-10% bucket because it had relatively higher incidences of viability in the sample and can therefore be construed in this sense to be our “best-performing” bucket.

In the table below, I have omitted displaying the results for the buckets on the very high end of the spectrum; the sample sizes are quite small and the results don’t provide meaningful information. I have also omitted the $20 performance threshold because few of the outcomes are statistically significant. They are, though, directionally similar to the other viability thresholds.

Each column demonstrates the likelihood for a player in the given k-rate bucket to ever be fantasy viable at the various thresholds. For example, a player in the 30-35% bucket is only 22.8% likely to be viable at the $5 level, according to my sample.

This table is information dense, and not every result is statistically significant, so I’ll highlight a couple patterns. First: at the $1 threshold, the results aren’t significant until the 15-20% bucket. At $5 and $10 viability, results aren’t significant until the 25-30% bucket. This means we can say that up to each of those levels, a player’s likelihood of viability is not materially different from the baseline k-rate bucket of 5-10%. Once we cross the 25% threshold though, we see clear decreases with a nadir at 35-40% for each measure of viability.

Interestingly, players in the 40-45% bucket fare slightly better than the bucket prior. This pattern is not present in either of the other samples I tested. While this may just be statistical noise, perhaps 60 plate appearances is a threshold at which players with certain power-heavy profiles struggle, but ultimately recover. Meanwhile 30 plate appearances is not enough of a sample to detect this dynamic, and players who are still striking out greater than 40% of the time after 120 plate appearances are a much less hopeful case. Notably, the 30-plate appearance sample exhibits similar behavior at the 30-35% bucket relative to the 25-30% bucket, with statistically significant results. This pattern is certainly worth future exploration and could be the subject of an article of its own.

Over/Under 35% K-Rate

In search of a pithy takeaway for this analysis, I wanted a point at which we can be relatively certain that a player’s chances at being viable are cooked. Informed by the more granular bucket groupings above, I separated players into groups above and below a 35% k-rate.

This model setup is where the results are most stark. Based on these results, it simply does not take very long to ascertain whether a player will ever be worth rostering. Players who strike out more than 35% of the time in their first 60 plate appearances are no more than 13% likely to be viable by any standard, and are roughly two-to-three times less likely to be viable than players striking out at rates below the cutoff.

Allow me take a brief detour that is too large for a footnote, and that comes with a math warning: an astute reader may notice that the $5 viability probabilities are actually higher than for $1 viability. Likewise, the lines in the continuous model for $1 and $5 viabilities eventually cross at k-rates in the 30s. These each imply the same thing: that it is somewhat likelier for a high k-rate player to be $5 valuable than it is for them to be $1 valuable. Logically, this isn’t possible; a $5 viable season must, by definition, be a $1 viable season. So a $1 viable season must be at least as likely as the $5 viable season. I don’t think this result is necessarily a problem; I checked the data and the number of $5 viable seasons in the sample is indeed lower than $1 viable for k-rates above 35%. Most likely, the year effects are interacting with the target variable slightly differently, and because these models are estimated independently from each other and do not enforce this nuance, we get a bit of weirdness in the results. As you’ll see in the Discussion, there is a lot of overlap in the confidence intervals for these viabilities. So, in short, I suspect what we’re seeing here is some amount of statistical noise, but it is worth noting. That there does not appear to be a statistically discernible difference in a player’s odds to be $1 viable vs. $5 viable after a high k-rate is an interesting finding on its own and worth further exploration.

Back to your regular programming: these odds are especially damning when considering the likelihood to become an elite fantasy player ($20 or more). Only 25 players in the sample struck out greater than 35% of the time after 60 plate appearances and went on to post a $20 or better season:

One could look at this list and be somewhat optimistic; there are a few Hall of Fame and several Hall of Very Good careers. But to this point, I caution about selection bias – we are looking at players who posted $20 fantasy seasons. Of course they’re good. These players are the outlier, not the norm.

Discussion

To validate the results, I wanted to look at the eventual performance of high-strikeout post-2020 debutants (n = 89) to see how frequently they achieved viability at various levels. Below is a table showing the expected results for an 89-player sample along with the 95% confidence intervals. We should expect the number of players who debuted with greater than 35% k-rates to achieve viability to fall between the intervals at each threshold.

This comparison isn’t perfect – because of the truncation problem mentioned before, we are likely undercounting the eventual number of viable seasons from this group of players. Likewise, any year-specific environmental effect (hello, Covid-19) is not accounted for. That said:

Pretty good! We have slightly fewer $5- and $10-viable players than the mean estimate, which is aligned with the intuition discussed above. We do have more $20-viable players than the mean estimate, but still within the upper bound. The small sample size of players at that level make precise interpretation difficult, but the results within these viability thresholds are overall encouraging and lend, in my opinion, a notion of reasonableness to the results.

Where we have some issue, though, is the $1 viability threshold. Our number of viable players actually exceeds the upper bound, which runs counter to the caveats discussed above. Why would this be occurring?

I’m not sure. I suspect it may be a combination of several interacting factors: $1 as a potentially weak indicator of value, Covid weirdness, and overall k-rate environment. Let’s consider the latter two first. Everyone knows the overall strikeout environment is increasing (or, at least, is elevated post-2020):

Explicitly, our model is missing this because it’s not incorporating data from these years. All else equal, as k-rates increase, the number of fantasy viable seasons at higher k-rates also increases.

Next, debutant data around the Covid-19 pandemic is, for lack of a better word, weird.