RotoGraphs Dynasty Rankings

All models are wrong, but some are useful.

It is with great pleasure that I can finally release what I’ve been working on for the past month: RotoGraphs Dynasty Rankings Model. We’ve traveled down a winding road with stops to handle prospects, prospects again, and aging curves. Now we’re at the end of the line.

Before getting to the juice though, I want to walk through the rest of the squeeze. I’ve pasted a table containing the Top 500 ranked players at the bottom of the article, so if you want to skip ahead, you can, but in case you want a peek under the hood at how the model actually works, keep reading. As has been the case, these rankings are based on a standard 5×5 roto, 12-team league with 23-man rosters, though the methodology is flexible and could be adapted to other settings and league types in future runs.

Equation

Here is the top-level formula for how the rankings work:

Here, we are calculating a discounted dollar sum over 10 years (inclusive of the current year, so this year plus nine future seasons). The ranking is then based on the sorted dollar sum value. To go into more detail, let’s start with a given player’s value in the baseline season.

Blended Jump-Off

For players with projected values (i.e. veterans, or prospects who are expected to accrue playing time), the baseline is a dollar value calculated from a mixture of pre-season projections, year-to-date performance, and rest-of-season projections, and various conditional penalties for injuries. The blend is as follows:

Here, YTD% is simply the percentage of games played for most teams divided by 162. Using a numerical example for where we are in the current season (YTD% = 75%), the weights (rounded) would be:

• 5% for pre-season projections
• 36% for YTD performance
• 59% for ROS projections

There are two key choices being made here: to use year-to-date values in the blend, and to subsequently down weight them. I’ll discuss each in sequence.

I don’t think it’s useful to ignore in-season performance. It is, after all, one of the only signals we have for whether or not a player has meaningfully improved! A noisy signal, sure, but one nonetheless. Ultimately this is where my personal philosophy as a modeler came into play, hearkening to Andrew Friedman’s famous view on free agency:

If you’re always rational about every free agent, you will finish third on every free agent.

If you base your player acquisition decisions solely on projections, you will rarely land the true breakouts. Ultimately, this weighting methodology is an injection of risk into the rankings – nudging overperformers (relative to projections) higher and underperformers lower.

That being said, I’ve chosen to down weight in-season performance relative to rest-of-season projections because if I don’t, year-to-date performance will overwhelm the jump-off point in a manner that isn’t consistent with their predictive capacity. The halving of the value is arbitrary and was calibrated based on experimentation; it is one of the areas of the model which could be subject to future optimization.

Consider the dual examples of Otto Lopez and Kyle Tucker. Lopez has been worth $31 year-to-date, is projected (via ATC) to be worth $14 for the remainder of the season, and was projected to be worth -$1 at the start of the season. Tucker, meanwhile: $5, $30, and $37, respectively. Their blended jump off values are $19 for Lopez and $22 for Tucker.

Look into your soul, dear reader, and tell me that you think Kyle Tucker is worth valuing at a baseline of $30 like rest-of-season projections would have you believe. I don’t, and I wanted that reflected in my model. On the flip side, I think you could defensibly argue that Lopez is bound for further regression and probably should be valued closer to his rest-of-season projected value. Fair enough! But valuing a $14 player at $19 is ultimately a pretty measured risk to take for someone who might also genuinely be a $25 player. For an underperformer like Tucker, the extra weight to year-to-date performance is, in my opinion, risk averse. A $20 valuation is still great – top 50 or so for this year – but to my eyes there’s serious bust risk going forward that simply isn’t reflected in his rest-of-season projection.

Injuries

Now, the blending methodology sounds okay on paper, but in practice is significantly effected by injuries. This could be:

• A player who is currently hurt, has been hurt for a little while, but is expected to be fine. This affects their year-to-date value in a way that is not representative of their go-forward value.
• A player who is only recently hurt, and is expected to be out for a while, up to and including the remainder of this season. This does not impact their year-to-date value (for now, though it will eventually), but does affect their rest-of-season projection in a manner that is similarly unrepresentative.
• Pitcher injuries *shrug emoji*

To this end, I devised a series of conditions to assign penalties based on the nature of the injury. Using Roster Resource data, for players who are expected to miss ~20% of the season – either on a year-to-date or total basis – I add a flag for that player to omit year-to-date performance and further adjust based on whether their rest-of-season projection exceeds at least half of their preseason projected value. As a note, I added the wavy tilde next to 20% because I’ve found that this has occasionally required some minor manual flexing based on some quirks in how injury information is disseminated from teams and subsequently recorded in the data. But the overall idea here is for a player who has spent around 20% of the season on the injured list to be subject to some additional attention in the model.

If it does, I simply omit their in-season performance from the blend altogether and calculate their baseline assigning a 5% weight to pre-season projections and 95% weight to rest-of-season projections. Here, the contamination from the injury is playing too much role to derive useful value in their year-to-date performance.

If the rest-of-season value does not exceed 50% of their pre-season projection, I revert back to their pre-season projection with a 25% haircut as a penalty. For hitters, this is the end of the story. Consider Aaron Judge, with a -$22 rest-of-season projection, -$6 year-to-date value, and $60 pre-season projection. My blend methodology would rate him at -$12 which is patently incorrect. The haircutting methodology instead rates him with a $45 jump-off, which felt much more reasonable.

Another example is Nick Kurtz, with a $33 pre-season value, $19 year-to-date, and no rest-of-season projection given his thumb injury. Without the conditional methodology, it would actually be impossible to create a jump off point for him because he wouldn’t be in the rest-of-season data. This method sets his baseline at $24, which is perhaps low for someone who is obviously a very talented young hitter. But, the most predictive indicator for future injuries is previous injuries – ask my colleague Jeff Zimmerman about it. Even with the ding, he’s still ranked 14th in the model. And to the extent that he bounces back to his pre-injury production, the model will adjust quickly. But for now, all else equal, I would generally prefer to gently fade a guy who is set to miss roughly 40% of the season to injury.

Pitchers, meanwhile, are their own separate beast. I used the same general logic as with hitters, but for pitchers whose seasons end due to a serious arm injury, I haircut an additional 50%. For pitchers who have the “revert” flag but aren’t missing the entire year, I haircut an additional 25% no matter how their rest-of-season projection compares to their pre-season projection. Note though that these additional haircuts are on the already-haircut values. So the additional 50% haircut for serious injuries results in a total 62.5% haircut, not 75%.

To provide an example, the projections have given Tarik Skubal a slight ding – $49 rest-of-season versus $60 projected pre-season – but I have his baseline value at $37. He’s obviously still very good and fantasy players should act accordingly, but I think anyone modeling dynasty rankings should be extremely conservative with regard to pitcher injuries. I’m all for stuffing guys who are awesome until they prove otherwise – the best pitchers can be tremendously more valuable than even the second- or third-best arms. Jacob Misiorowski has been 50% more valuable ($61 versus $42) than Cam Schlittler, the number-two ranked pitcher. But once cracks in the armor start to show, I think it’s prudent to be swift in cutting them down in the ranks.

Aging Curve and Prospects

The rest of this explainer should be fairly downhill. Using the baseline value calculated above, and the aging curves that I detailed in last week’s article, I model out for a full 10-year horizon based on the player’s current age.

For example, a player who is 23 will receive bumps up to their peak age based on the curve, followed by decreases after they reach their peak. I floored negative-value seasons at -$10 because I didn’t think it added useful information to have a currently good but aging player suffer as extreme of a penalty as the aging curve would imply in edge seasons. It wouldn’t be consistent with how players are rostered in fantasy; if they’re sufficiently old and sufficiently bad, they’re cut. This is something that I might calibrate more precisely or change in future versions of the model, but ultimately it produced results that were overall reasonable and consistent with my intuition, so I stuck with it.

For prospects, using the rank buckets discussed in my previous articles (with a slight modification; I ultimately did separate number-one ranked hitters to create hitter buckets of Top 1, Top 2-5, Top 6-30, and 30+) and used the 40^th percentile peak value for hitters within each bucket, and the 20^th percentile for pitchers. These came out to be:

So, to use top-ranked Jesús Made as an example, the model is treating his expected peak season to be roughly that of Cliff Floyd’s best year. Meanwhile, fourth-ranked Walker Jenkins’ best season looks like Hank Blalock’s to the model. It then uses the aging curves to extrapolate downward to their current age and outward until they hit 10 years past their current age. For the ranks, I used the FanGraphs lists, as opposed to Baseball America (which the historical data are based on) to calibrate the model’s prospect preference to match our in-house philosophies.

I have set the model to be very pessimistic with pitchers:

To me, this is conceptually aligned with the conservativeness by which I treat them for injuries and generally reflects the evergreen logic of TINSTAAPP. For prospects who rely on list rankings to derive their dynasty value, these assumptions result in the best hitters ranking roughly in the 50-100 range, while the best pitchers are generally closer to 200. If you’re a pitching nut, before coming at me with a pitchfork, go back to the 2024 prospect list and look at the sparse number of pitchers who have returned meaningful fantasy value so far. Out of 37 pitchers ranked, it’s Paul Skenes, Mason Miller, and Jacob Misiorowski. And sure, these guys are awesome, but that’s not a very good hit rate.

For players ranked on both mid- and pre-season lists, the model uses the mid-season rank. For players who appeared on a preseason list but accrued sufficient playing time to lose prospect eligibility, the model uses the higher of their values based on their prospect status or the blended value methodology detailed above. This only lasts for the season that they appeared on the list; subsequent years will rely only on projections. I like these mechanics because it results in the model correctly stuffing a guy like Kevin McGonigle, who would be ranked much lower based on his prospect status. Meanwhile, it affords a bit of patience to guys who struggle at first but reacts fairly swiftly after that grace period ends. There are new top prospects literally every season – it doesn’t make sense to hang on to a player who isn’t producing when you can (and should) be digging to find the next guy who will to take their place on lists.

To circle back to aging curves briefly, last week I promised a comparison of the model rankings between a delta method-based curve and a GAM-based curve. The flat shape of the GAM aging curve, in my opinion, highly overrates older players:

Reasonable minds could disagree about where these players should settle in a final dynasty rank; maybe 39 is too high for Chris Sale. But to that point, if 39 maybe is too high, then 14 is not reasonable. So while the GAM model might perform marginally better in certain testing setups, I don’t think it is well suited for this exercise.

Discount Factors

After applying the aging curves, with a 10-year model of projected dollar value for both current veterans and prospects in hand, I applied a 25% discount factor to future years:

Here, the discount for 2026 would be 100%, because 1.25⁰ equals 1. For 2033, it is 21%. That is, the model only values a 2033 season at 21% of what it would value an identical season in 2026. This rate is aggressive, but I think is appropriate. Flags fly forever. With something as unpredictable as fantasy baseball, even in dynasty, you should chase championships, not prospect rankings.

In addition to the general 25% discount factor, I have added an additional 4x multiplier on catchers, and a 6x multiplier on relievers.

Again, I know these seem extreme but consider the volatility associated with each position. This is another area I hope to calibrate more precisely over time, but for now, I think the results are reasonable. To bring it all back:

Then, take the sum of each player’s discounted values and sort it to create the rankings:

Conclusion

An astute reader may notice the occasional oddity in the rankings, like Walker Jenkins being ranked one spot higher than Jesús Made, despite being a lower-ranked prospect. This is because given his age (21 vs. 19), the model sees him as closer to his peak and therefore likelier to begin accruing fantasy value on your rosters sooner. Since nearer seasons are discounted much less steeply than later ones, Made’s relatively longer ramp up to viable fantasy production results in more of his best years being discounted relative to Jenkins’. With something like this, while I might personally value Made more, I also think the model is built soundly and want to listen to its results to see what kind of information can be gleaned. I will spend time in subsequent articles diving deeper into the actual rankings, pointing out some things that I like, don’t like, or just think are interesting. It may will undoubtedly be wrong in various ways, but I’m optimistic that it will also be useful.

*Note: Already, a hitch: this model run was done prior to knowledge of Jacob deGrom’s removal due to triceps fatigue, and not enough information is known yet to have sufficient inputs for a re-run. So, interpret his ranking with caution.