The Man Who Wins Nothing
Every rating system in cycling counts results. A lead-out man's result belongs to somebody else.
On stage 4 of the 2023 Tour de France, into Nogaro, Jasper Philipsen won the bunch sprint. Mathieu van der Poel delivered him, throwing elbows to do it, and afterwards the commissaires moved van der Poel from sixteenth to twenty-second, the last place in the front group, fined him 500 Swiss francs, and docked him thirteen points in the green jersey competition. Philipsen kept the stage.
That is the lead-out man's entire situation in one afternoon. The work is visible enough that the officials can punish it, and invisible enough that no public ranking in the sport records that it happened. PCS, CQ, the UCI and our own Elo all count results, and a lead-out man's result is, by design, somebody else's.
There are good reasons for that. Results are what's recorded, they're unambiguous, and a system that tried to credit intentions would be guessing. Cycling does have tracking data of a kind: every bike at the Tour has carried a GPS transponder since 2015. But the feed is, as DC Rainmaker found when he looked into it, "relatively simplistic: Just position and speed", and nobody has built the public research archive that let baseball catch up with catcher framing.
The academic work on credit is thin too. Two economics papers measure teammate effects in aggregate (Candelon and Dupuy, 2015; Matthes and Piazolo, 2024), one of them in breakaways rather than sprints. A preprint from this year redistributes a season's UCI points within teams, with its latest version tuning its weights to agree with the rider ratings in Pro Cycling Manager, the video game (Calo-Blanco, 2026). The closest work is VeloRost (Rize, Saldanha and Moskovitch, 2026), which gives every rider separate ratings as a leader and as a helper, and finds that knowing a roster's helpers predicts results better. It credits helpers through their team's UCI points rather than through bunch sprints, though, and none of these papers ranks named lead-out men or puts a value on a seat in the train.
There is one study that gets close, and it is encouraging. Menaspà, Abbiss and Martin (2013) video-analyzed a single world-class sprinter across four years of Grand Tours and found that in sprints he won, he was closer to the front and had more teammates around him at sixty, thirty and fifteen seconds out, while the duration of the sprint itself was no different between wins and losses. What separated winning from losing happened before the sprint started, and some of it was the number of his own riders in the picture. That is a sample of one rider, 31 sprints on video, and it counts teammates rather than identifying them. The only other study like it, by van Erp, Marcel Kittel and Lamberts (2021), looked at 21 Tour sprints by a world-class sprinter and found that position 30 seconds out separated his wins from his losses, but the number of teammates around him did not. So the two agree that where you are matters, and disagree on whether your own riders are what put you there, which is a question for a much bigger sample.
What we have instead is every finishing order since 1996, and a record of who was on which team. That turns out to be enough, if you are careful about what you ask it.
The question I asked is narrow on purpose. Take every men's bunch finish where a team's best sprinter, defined by his rating before the stage and never by how he finished, arrives in the group. Ask whether he beat expectation. Then hand the credit to the teammates who were with him. Do that 64,319 times across 1996 to 2026, handing out 149,910 credits, and see whose name accumulates.
The first version of this was a disaster, and instructively so. Its top lead-out men included Primož Roglič.
The bug is precise and it is worth understanding, because it is the whole reason the metric needs a second idea rather than just more data. To pick which teammates count as "the train", I ranked them by sprint rating, and our sprint head does not measure what its name suggests. It measures within-group ordering: how you finish relative to the riders you actually arrive with, in whatever group that is.
So the rating is earned against wildly different opposition depending on who you are. A climber's arriving group is a mountain selection: a dozen other climbers, none of them fast. Beat them to the line often enough and you accumulate a healthy sprint rating, having never once been in a bunch kick. A lead-out man earns his in flat finishes against actual sprinters, which is a far harder room. The two numbers are not on the same scale, and comparing them inside a team ranks the climber above the specialist. Roglič and Steven Kruijswijk were being counted into their own team's sprint train.
What's needed is a way to say this rider is not working for someone else's sprint today, and the obvious candidate is how close he is on general classification. In raw seconds it is the worst of the options I tried. A time gap means something different in week one than in week three, and a fixed threshold in seconds is meaningless across races: the median rider sits 338 seconds behind the GC leader in a stage race of three days or fewer, and 2,245 seconds behind in a Grand Tour. Raw seconds separate the real lead-out men from the GC riders at an AUC of 0.711, which is worse than simply excluding whoever is best overall. (AUC is the chance that the key puts a randomly chosen lead-out man on the right side of a randomly chosen GC rider: 0.5 is a coin toss, 1.0 is perfect.) Turned into a percentile of the field, the gap does respectably, at 0.868. But it only exists where there is a GC to be behind: not in a one-day race, and not on the first day of anything.
The key that works is climbing ability: not GC position on the day, just how good a climber the rider is in general. It separates at 0.920 and beats the null: filtering on overall quality instead of climbing gets 0.816, so this isn't a disguised way of deleting good riders. A rider's GC rating does almost exactly as well, at 0.917, which is no surprise: for riders like these the two ratings measure nearly the same thing. I use climbing because it separates slightly better and says plainly what the rule is. Climbers are not lead-out men. The rule that ships is a conjunction: a candidate is dropped if he is both top-30 in the field on climbing and genuinely climber-shaped in absolute terms, unless he is a top-20 sprinter. Then the survivors are re-ranked and the top three become the train.
Filtering rather than down-weighting matters here. Dropping a climber from the pool promotes the real lead-out man into it; discounting him leaves him occupying a slot that the specialist never enters.
With that in place, the fit recovers the pantheon.
| # | Rider | λ (Elo) | Stages |
|---|---|---|---|
| 1 | Mørkøv Michael | 60.9 | 158 |
| 2 | Lombardi Giovanni | 39.9 | 73 |
| 3 | Corioni Claudio | 38.7 | 66 |
| 4 | Zanini Stefano | 38.6 | 76 |
| 5 | Fagnini Gian Matteo | 38.1 | 108 |
| 6 | Simon François | 37.3 | 58 |
| 7 | Renshaw Mark | 37.2 | 124 |
| 8 | Veelers Tom | 36.9 | 113 |
| 9 | Klier Andreas | 34.9 | 152 |
| 10 | Henderson Gregory | 33.9 | 113 |
| 11 | Richeze Maximiliano | 32.9 | 124 |
| 12 | Jansen Amund Grøndahl | 32.4 | 57 |
The site keeps this board up to date, refit monthly, as the Career λ column of the all-time lead-out board.
Michael Mørkøv is clear of the field by 21 Elo. Below him, the board is two generations of the same job. One is the men who delivered the fast men of the late nineties and early 2000s: Giovanni Lombardi and Gian Matteo Fagnini, who between them led out Mario Cipollini; Andreas Klier, Stefano Zanini and François Simon, whose best seasons the model spent in the trains of Erik Zabel, Paolo Bettini and Stuart O'Grady respectively. The other is the modern lead-out men: Mark Renshaw, Tom Veelers, Greg Henderson and Maximiliano Richeze.
The people who worked with them would give you the same names. Mark Cavendish, of Mørkøv, in 2021: "It's a known fact Michael is the best leadout man in the world." Brian Holm, who directed Cavendish's teams at HTC and Quick-Step, of Renshaw in 2016: "He is the master of the leadout, and no one is better than he is." The fit that produced this board was never told who any of these riders are, but I should be exact about the one place names went in. I chose the exclusion rule by checking which key best separated a list of well-known lead-out men from three GC riders, and all five of the men I just named are on that list. Its thresholds were then tuned on three riders from that list: Mørkøv, Veelers and Bert Van Lerberghe. So the filter was built to keep men like these eligible. The credit they then collected came from one place only: their sprinters beating expectation when they were in the train.
That is the part you already knew, which is what a new statistic is supposed to deliver before you trust its surprises. Claudio Corioni at third is the first name most fans wouldn't list, and he checks out. He rode for Lampre in 2006 and 2007, and the model credits him mostly to Daniele Bennati, who won five Grand Tour stages in 2007. Before that year's Tour, Bennati told La Gazzetta dello Sport that Corioni was "able to work well in the finale". After the biggest of those wins, on the Champs-Élysées, he said: "Claudio Corioni did the last two laps all-out; it was amazing." Both men moved to Liquigas for 2008, and when Bennati won a stage of that year's Vuelta, Corioni was one of the four teammates he thanked by name.
So far, so good. This is where a career average starts to look like the wrong shape.
I fitted λ on everything through 2019 and used it to predict 2020 through 2026. It helps, but only just: a small fraction of what the same links can do when they are asked the question a different way.
Stated plainly, that is obvious. Renshaw's λ is built mostly in his years with Cavendish, and in 2020 he is not a lead-out man; he is not a bike racer. Lead-out is at least as much an attribute of a partnership as of a person, and partnerships end. "The more they race together the better. That's a fact," as Holm puts it. Mørkøv led out Alexander Kristoff at Katusha, then Elia Viviani, Sam Bennett, Cavendish at the 2021 Tour, Fabio Jakobsen, and Cavendish again at Astana, and the rating that describes him with Viviani in 2019 has very little to say about 2024.
The fix is to stop averaging a career and start tracking a state. Give every rider a running total that gains when his sprinter beats expectation and decays between seasons, the way our other modifier heads do (though, as it turned out, much faster). Then a rating means what this rider has been doing lately.
That version transfers fine. It also, in the building, threw away almost everything I put into it.
I began with an elaborate construction: un-shrinking each result to the level it implied, and centering by the leader's own rating band to correct for the fact that a weak sprinter has far more room to gain than an elite one. That last effect is real and large: on stages where the designated sprinter wins, the mean rating gain runs +126.4 in the lowest leader band against +13.6 in the highest, a factor of nine. It felt like the correction the metric obviously needed.
It made things worse. Band centering was the worst of five constructions I tested, and it was actively destroying signal rather than cleaning it.
Then the soft cap. I capped each result's contribution so a freak day couldn't dominate a rider's total, and tightening the cap kept improving the fit, monotonically, from ±400 to ±150 to ±50 to ±8. That has only one interpretation: as the cap approaches zero, the function approaches the sign of the result. The magnitude was noise, and only the direction carried information.
Push it all the way to the limit and the metric becomes a bare count of good days, which predicts best on the random split. The published version stops just short of that, keeping the cap tight enough to throw away almost all of the magnitude while retaining a little, and adds one thing the pure count lacks: a weighting by how serious the race was. The choice that matters most is the asymmetry. A symmetric cap at ±8 recovers only two-thirds of what the bare count manages, while +8 / −3 recovers nearly all of it, because it refuses to blame the lead-out for the bad days.
That race weighting is where I expected the article's number and the model's best number to part company. It is baseball's minor-league translation problem in cycling kit: a lead-out man who reliably makes his sprinter overperform in 2.2 races is genuinely predictive, and will go on being predictive, and still isn't a better lead-out man than someone doing it at the Tour. Translating between levels usually costs accuracy at the source level. Here it doesn't. The weighted version holds about 8% more of the margin the rating has over simply knowing how fast your teammates are, and on the harder test (fit through 2019, predict the seasons since) it is the best version of all. The board and a prediction model can use the same number.
The decay, on two different validation designs, says a lead-out rating should have a very short memory. Both designs peak at 0.75 a year, and both are flat from 0.45 to 0.85, so the rate is not identified anywhere in that range. The published rating uses 0.45, which halves a value in a little over a year, still a far shorter memory than any other modifier in the system has.
Where the sport does argue is over which seat in the train matters. A train is not one job, and the people who run them are the first to say so. Holm, in 2016: "a good leadout is not just the final 500 meters to go. It's chasing the breakaways, riding in the crosswinds with 50km to go, keeping the sprinter out of trouble all day". Renshaw, describing the HTC train from inside it: "If it is a sprint stage, we try to control it all day with one or two riders. We make sure breakaways don't get too far ahead." Then: "3 kms to go is usually a pretty decisive time. That's when all our guys are doing their last turns." At the far end, Greg Henderson, André Greipel's last man, in 2014: "I drop him off with 200 meters to go." Note the units: the early work is measured in minutes of gap to a breakaway (at Cipollini's Saeco, the break-control man Giuseppe Calcaterra's rule was that with 50 kilometers to go the gap could not be more than five minutes), and the last turn in meters.
Mørkøv, who has ridden both ends of it, draws the physiological line. In 2019 he was riding in front of Fabio Sabatini, Elia Viviani's last man at Deceuninck-Quick-Step: "To be the last lead-out man I would say you need to be [a] sprinter because you need to be able to reach a really high speed. But to be in my position, in front of Fabio, you don't need to be super fast. What you need more of is to have a high-power output for 30 seconds to a minute, do a strong pull more than a sprint." Marcel Sieberg, who rode that seat for Greipel, put numbers on it: "I am usually at 900 watts to 1,000 watts for 20 to 30 seconds. The final sprint can be double that."
So the seats are different jobs, and the argument is over which of them is scarce.
Cavendish is unambiguous, and he answers it as a roster problem. On replacing George Hincapie: "You can't replace George. We're going to get two people to replace that one job. But we've got six riders who can be those two people." Renshaw, Cycling Weekly added in its own words, "will retain his position as last man." On why Greipel, a great sprinter, could not do the job: "Watch his sprints… he loses wheels. I can sit on Mark Renshaw's wheel — he can surf the side of the peloton for two kilometres and then still go fast. A lead-out man is a sprinter, but he's not a sprinter."
The two men at the end of Viviani's 2019 train put the difficulty earlier. Sabatini, the last man: "The most difficult part of a lead-out is finding a good position coming into the final kilometre, because normally by the time you reach the final kilometre the sprint is already done." And Mørkøv, the man in front of him, describing his own turn: "The hardest part is the timing because it's all about trying to arrive in the best possible position in the last kilometre. When you arrive in the position you want everything to go automatically."
Read together, that is one handover described from both ends. The last man says the sprint is decided before his turn begins. The man before him agrees, and says that deciding it is his job.
I had been imposing an answer on this: weights of 1.0, 0.6 and 0.3 by train position, picked by intuition and never checked. So I threw them away, accumulated every train member at equal weight so the data couldn't inherit my guess, and fitted a separate coefficient for each position.
The fastest helper in the group is worth about 1.6 times each of the next two, and the fourth- to sixth-fastest are worth something real but small: about a fifth of the fastest, and 3.4 standard errors above zero. My hand-set weights had the order right, but gave the second and third too little and the rest nothing. The published rating still uses them (1.0 / 0.6 / 0.3, and nothing for four to six), so every board in this article undersells riders who are usually the second- or third-fastest man in their train.
A warning about the labels, because they sound like seats and they aren't. "Last man" here means the helper with the highest sprint rating, not the one who physically rode last. Usually that is the same rider, since Mørkøv's own rule is that the last man has to be a sprinter. The exception that matters is Mørkøv himself. In Viviani's 2019 train, by both men's account, Sabatini rode last and Mørkøv in front of him. Across Viviani's 25 stages that season the model puts Sabatini in the last-man slot on none of them, and Mørkøv there nine times, more than any other rider.
So the model leans toward Cavendish's side of a live disagreement: the value it finds is largest for the fastest helper, who is usually the one who delivers. I'd rather say so than pretend the sport is unanimous, and the sharpest dissent comes from Cavendish's own directeur sportif. Holm, in 2014: "Renshaw is never really better than the guy ahead of him." Adam Hansen, on Lotto's sprints: "Last year, we probably had too many guys for the last kilometre and not enough guys for before." Mørkøv and Sabatini, who say the hard part comes before the last turn, aren't contradicted by any of this. The model cannot see where the hard part is. It sees who was in the group, and how fast they were.
There is a reading that lets both sides be right, and it's the one I'd bet on. Much of what Holm and Renshaw describe, "chasing the breakaways" and making sure "breakaways don't get too far ahead", decides whether there is a sprint at all, not who wins it. My model only ever looks at stages that finished in a bunch. A stage where the break stayed away never enters the sample, and neither does the value of the riders who would have stopped it. I have conditioned on that work having already succeeded. Whatever the fit finds for the slower helpers is only the part of a team's work that survives into the final kilometers, and nothing guarantees that the riders who did the early work are the ones it ranks fourth to sixth. The full value of that work may well be much larger; this design structurally cannot see it.
The gap between the fastest helper and the next two is on firmer ground, since both are conditioned identically.
It also eases a worry I carried for most of the project: that the metric couldn't tell the front of a train from the back. It can, roughly. On its own, a feature built from riders four-to-six predicts about a quarter as well as one built from the top three, and when the two are fitted together the top three keep most of it: the back of the train holds on to about a quarter of the front's value per point, and no more. What it cannot do is tell apart two fast riders who share the front. The de Jongh case at the end is one.
The payoff is that the fitted coefficient converts directly into Elo. The train's rating enters the sprinter's own, so a lead-out man's score can be expressed as how much better than his rating his sprinter performs when this rider delivers him, measured against a last man with no record at all.
A hundred and sixty-eight Elo is what the full model puts on the best lead-out man in the archive. On a stage where a helper with Lombardi's 2002 record, or Mørkøv's 2019 one, is the fastest man in the train, it expects the sprinter to ride as if he were rated about 153 and 162 points higher. It is a per-stage effect, not a change to the sprinter's rating, which already contains the results those lead-outs helped produce.
It is not all the lead-out man's doing, and this is the most important caveat in the piece. Philipsen, who has had the best of it, puts it in a sprinter's terms: "Mathieu makes it easier for me to win, yes, but you have to be a good sprinter to finish the preparatory work he does." A helper's score is built from his sprinter's good days, so for a settled partnership it also tracks the sprinter's recent form, which an Elo rating absorbs only gradually. Give the model the sprinter's own recent record as well, and the lead-out coefficient falls to 53% of its size: about +82 and +87 Elo for Lombardi and Mørkøv at their peaks. That control overcorrects, since some of the sprinter's good run was his train's work in the first place, so the truth sits between the two numbers, and I would rather quote the one that cannot be accused of flattering the lead-out man. From here on, the Elo figures are the form-controlled ones.
A word on units, because the rest of the site uses different ones. These are points on the rating model's own sprint-Elo scale, the one it updates in. The ratings shown elsewhere on the site are display ratings on a 0–1000 scale, and near the top of the sprint field one raw Elo point is worth about 0.56 display points. So +89 here is about 50 points on the site's sprint rating, and +168 about 95.
Even the conservative number is large. This season's five best sprinters sit within about forty Elo of each other, and the tenth-best is 148 behind the first. Add +89 to the fifth-best on a stage where a helper like that leads him out, and he becomes the best in the field; add the full +168 to the tenth-best and so does he.
The form control also puts the two ratings from earlier on one level. A peak is the top of a running total, and at the top of the board it sits at a bit more than twice a rider's career average, so the fair comparison with λ is the running total's average per appearance. Across the top fifty, that average is about +22 Elo on the form-controlled scale, and λ, at face value, is +27. λ is itself shrunk toward zero by its fitting; undo that and it lines up instead with the full-effect average, +51 against +42. The two measure from different zeros (λ from the average helper, the accumulator from a helper with no record at all), so the fair reading is that they overlap rather than agree to the point.
The transfer market has noticed, if not in these units. When Decathlon signed Olav Kooij for 2026, the same announcement brought in Cees Bol, Robbe Ghys, Daan Hoole and Tobias Lund Andresen. Ten weeks later the team hired Mark Renshaw, a man from the top ten of both career boards in this article, as a sports director for its sprint project. He said the train was why he came: "the presence of a dedicated sprint train was the key factor that convinced me to join the team." Soudal Quick-Step extended Tim Merlier and his lead-out man Van Lerberghe to 2028 in the same announcement.
The riders are careful to say it wasn't a package deal. "We very deliberately did not sell ourselves as a package," Van Lerberghe told Het Nieuwsblad (via Wielerkrant, translated from Dutch). Bol, on the Decathlon build: "In the end, everyone has to say 'yes' individually, because you have to really believe in it yourself." Sometimes a sprinter and his man do move together (Shane Archbold went with Sam Bennett, and Cavendish took Renshaw and Bernhard Eisel to Dimension Data for 2016), but the contracts are signed one rider at a time. What teams do is announce them together and hire a specialist to arrange them. That is what you would do if you believed the value was in the combination and in who sits where, and it is the belief this metric is trying to test.
This summer's Tour gave the skeptic a case too. Van Lerberghe abandoned on stage 6, and Merlier won stages 7, 8 and 12 without him. One race proves nothing, since a sprinter has more than one helper, but it is the first thing a doubter will raise.
It also gives us something to grade. Nine months in, here is what three of the riders Decathlon signed alongside Kooij are worth by this metric. Cees Bol is at +45.4 Elo, fifth on the current board and just short of his career high, with a hit rate of 0.77 across thirteen appearances this season (two of them for Lund Andresen rather than Kooij) against a field average of 0.65. Robbe Ghys is at +15.2, on four appearances, and Daan Hoole at +6.0, on three. The honest reading is that one of the three is emphatically working and the other two have not yet done enough of the job to register. All three are hitting at or above average when they do it, which on three and four appearances means almost nothing.
Hoole is the interesting one, and it is a caution about the metric rather than about him. He is a time-trial specialist (he won the time trial on stage 10 of last year's Giro). Two features of the design matter here. The published rating credits only the three fastest helpers in the group, so a time-triallist riding with faster teammates earns credit only on the days he is among the three fastest there. And because his real work is the long, steady effort earlier in the day (break control and the early lead-out), it is the work the curve above scores lowest and, if the selection argument is right, the work this design is worst placed to see. A team that has just paid for that kind of rider and a metric that barely sees him need not be in disagreement; the metric is simply not looking.
Renshaw's own career is the tell: two world titles in the team pursuit (he rode the qualifying rounds), a thin road palmarès, and, by Cycling Weekly's count, nineteen Tour de France stage wins racked up alongside Cavendish. One of his best road results came from a lead-out: second on the Champs-Élysées in 2009, behind Cavendish. Before the last corner, Garmin's Julian Dean tried to take Tyler Farrar up the inside, but Renshaw and Cavendish "had the best and fastest line", Cycling Weekly reported, and when Dean had to brake he blocked everyone else on the way out. The two came out of the corner clear, and by the line Renshaw had time, Cyclingnews wrote, "not only to look behind, but to claim second by a margin Columbia's rivals would love to win by." Cavendish, in 2011: "There's no one in the World that can do his job; he's the best in the World at what he does."
It is also a hint about where these riders come from. Two of the modern names at the top of these boards came up on the track: Mørkøv as a Madison and six-day rider, Renshaw as a team pursuiter. Mørkøv's own explanation is that he has ridden about 50 professional six-days, "and every night in a six-day race you will have 10 or 15 times that you sprint for the finish line," against a road rider's twenty or thirty sprints in a year. Two riders prove nothing, but if the skill is timing a sprint from inside a moving line, the velodrome is where you get the most practice.
Hollow points are seasons with no lead-outs: the rating is carried forward, not re-earned.
The decay does the work you'd want it to do here. Lombardi's line is the Cipollini era in miniature: a steady run leading out Erik Zabel at Telekom, then a spike in 2002, the year Cipollini won Milan–San Remo and the world championship, and gone within a few seasons of the job ending. Renshaw's rating falls away after 2015 as the work thinned out, though he was still riding for Cavendish, and the decay finishes it after he retires in 2019. Mørkøv's 2019 peak sits above a decade of steady accumulation. And Dries Van Gestel, who is not a household name, is currently rated higher than anyone in the sport because he has spent two seasons delivering Paul Magnier at Soudal Quick-Step, and Magnier keeps beating his rating.
| # | Rider | Elo now | In full | Career peak | Hit rate | Leading out |
|---|---|---|---|---|---|---|
| 1 | Van Gestel Dries | 68.4 | 127.8 | 75.7 | 0.72 | Magnier Paul |
| 2 | Stuyven Jasper | 55.9 | 104.5 | 58.7 | 0.65 | Merlier Tim |
| 3 | Consonni Simone | 49.4 | 92.4 | 52.7 | 0.64 | Milan Jonathan |
| 4 | van der Poel Mathieu | 46.2 | 86.5 | 50.1 | 0.78 | Philipsen Jasper |
| 5 | Bol Cees | 45.4 | 85 | 46.3 | 0.69 | Kooij Olav |
| 6 | Van Lerberghe Bert | 43.4 | 81.1 | 61.9 | 0.67 | Merlier Tim |
| 7 | Stewart Jake | 43.3 | 80.9 | 50.2 | 0.64 | Vernon Ethan |
| 8 | Theuns Edward | 40.3 | 75.4 | 40.8 | 0.75 | Milan Jonathan |
| 9 | Smith Dion | 38.6 | 72.1 | 48.1 | 0.7 | Vernon Ethan |
| 10 | Planckaert Edward | 38.1 | 71.3 | 48.6 | 0.69 | Groves Kaden |
This board is live on the site as the current lead-out ratings, rebuilt with every data update, beside the all-time career highs and a board for each season, starting with 2026 so far.
The current board has one outside check. Before this year's Tour, ProCyclingUK picked the best lead-out riders in the race, by eye: Jonas Rickaert, Van Lerberghe, van der Poel, Jasper Stuyven and Bol. Four of those five are in the model's current top six, and the fifth is its clearest miss: Rickaert is 37th on the current board, although he was on the list of well-known lead-out men I used to choose the exclusion rule.
The likeliest explanation is the weakness I flagged after Figure 4: the published weights undersell riders who are usually the second- or third-fastest man in their train, and Rickaert is one. He is a lead-out specialist rather than a sprinter. His own team introduces him as a lead-out man, and VeloRost picks him out as one of Alpecin's "dedicated leadout specialists" (Rize, Saldanha and Moskovitch, 2026). The site ranks his sprint rating this season in the 53rd percentile among WorldTour riders, against the 74th for his teammate Edward Planckaert, so on the 13 stages the two finished in the group together, the model ranked Planckaert ahead of him on 9. Across his career Rickaert takes the last-man slot on only 24% of the stages he finishes in the group with his sprinter, and on none of the six this season, but he is one of the three credited helpers on 83% of them.
Sort by hit rate instead, among the 554 riders with at least 50 credited stages, and he is back among the best: his 0.783 is 13th (Figure 7). Hit rate is a blunter tool than the rating, since it ignores the level of the race, and at 50 stages one standard error is about 0.07, so the order within the table is loose. But it counts every credited stage the same, whatever the slot, which is where the rating shortchanges Rickaert.
| # | Rider | Hit rate | Stages | Main sprinter | Peak |
|---|---|---|---|---|---|
| 1 | Guerra Héctor | 0.824 | 51 | Cardoso Manuel Antonio Leal | 2008 |
| 2 | Lombardi Giovanni | 0.822 | 73 | Cipollini Mario | 2002 |
| 3 | Nordhaug Lars Petter | 0.82 | 50 | Boasson Hagen Edvald | 2012 |
| 4 | Trenti Guido | 0.819 | 72 | Boonen Tom | 2006 |
| 5 | Gentili Massimiliano | 0.815 | 65 | Fischer Murilo Antonio | 2005 |
| 6 | Boucher David | 0.809 | 63 | Bouhanni Nacer | 2012 |
| 7 | Martin Tony | 0.808 | 73 | Cavendish Mark | 2009 |
| 8 | Coledan Marco | 0.807 | 57 | Mareczko Jakub | 2018 |
| 9 | Bagües Aritz | 0.804 | 51 | Aberasturi Jon | 2021 |
| 10 | Bouet Maxime | 0.796 | 54 | Alaphilippe Julian | 2015 |
| 11 | Raileanu Cristian | 0.79 | 81 | Salby Alexander | 2025 |
| 12 | Lund Anders | 0.784 | 51 | O'Grady Stuart | 2009 |
| 13 | Rickaert Jonas | 0.783 | 69 | Philipsen Jasper | 2025 |
| 14 | Scheirlinckx Staf | 0.778 | 54 | Farrar Tyler | 2007 |
| 15 | Sieberg Marcel | 0.777 | 157 | Greipel André | 2009 |
| 16 | Müller Martin | 0.774 | 53 | Pollack Olaf | 1999 |
| 17 | Millar David | 0.771 | 70 | Farrar Tyler | 2010 |
| 18 | Sánchez Luis León | 0.769 | 91 | Rojas José Joaquín | 2008 |
| 19 | Auger Ludovic | 0.769 | 78 | Stoyanov Plamen | 2003 |
| 20 | Van Keirsbulck Guillaume | 0.768 | 56 | Dupont Timothy | 2018 |
The same rating runs on the women's races, on far less data: 16,921 credits since 1998, about a ninth of the men's 149,910. The top three on the current women's board all lead out Lorena Wiebes, and so does the sixth: Mischa Bredewold, Lotte Kopecky, Barbara Guarischi and Marta Lach (Figure 8). The Elo figures use an exchange rate fitted on the women's own races, 17.04 points of Elo per point of rating against the men's 6.49, so they are amounts of Elo rather than a bare ranking. The two boards are still measured against different fields, so each ranks riders within its own peloton.
| # | Rider | Elo now | Career peak | Hit rate | Stages | Leading out |
|---|---|---|---|---|---|---|
| 1 | Bredewold Mischa | 101.8 | 105.3 | 0.93 | 28 | Wiebes Lorena |
| 2 | Kopecky Lotte | 93.6 | 103.1 | 0.83 | 24 | Wiebes Lorena |
| 3 | Guarischi Barbara | 87.5 | 111.6 | 0.76 | 62 | Wiebes Lorena |
| 4 | Bäckstedt Zoe | 85.2 | 88.1 | 0.76 | 17 | Consonni Chiara |
| 5 | Ton Quinty | 79.7 | 83 | 0.68 | 19 | Baker Georgia |
| 6 | Lach Marta | 63.6 | 102.5 | 0.77 | 31 | Wiebes Lorena |
| 7 | Jaskulska Marta | 63.1 | 71.8 | 0.76 | 25 | Coles-Lyster Maggie |
| 8 | Brand Lucinda | 61.5 | 82.8 | 0.75 | 64 | Balsamo Elisa |
| 9 | Moors Fleur | 61 | 76.3 | 0.73 | 15 | Copponi Clara |
| 10 | Georgi Pfeiffer | 60 | 125 | 0.74 | 47 | Kool Charlotte |
The live version is the women's lead-out board.
Back on the men's side, I have two ratings for the same riders, the career λ from earlier and the decayed accumulator, built on the same data with the same exclusion rule. Across all 530 riders who qualify for both, their rank correlation is 0.43: clearly related, and a long way from the same number.
Thirteen riders sit in both top twenty-fives, and they are exactly the names you would expect: Fagnini, Mørkøv, Richeze, Lombardi, Guido Trenti, Corioni, Renshaw, Jürgen Roelandts, Henderson, Veelers, Van Lerberghe, Sieberg and Sabatini. When two estimators built on different principles agree on a pantheon, that pantheon is probably real.
The disagreements run in one direction each, and both directions make sense. The riders the accumulator likes and λ does not (Van Gestel at 5 against 108, Danny van Poppel 6 against 116, Mike Teunissen 17 against 110) are riders with recency: currently active, or concentrated in a few heavy seasons. The accumulator is a decayed sum, so it rewards doing the job a lot and doing it lately. The riders λ likes and the accumulator does not (Grøndahl Jansen at 12 against 219, Henk Vogels 14 against 119, Tony Martin 18 against 115) have the opposite shape: a strong per-appearance effect on relatively few appearances. λ is an average, and averages do not care how often.
Neither is wrong. They are answering "who moved the needle most per stage, net of his team" and "who has been delivering sprinters, recently and often", and those are different questions about the same job.
One last thing, because it is the sort of result that makes me trust a metric more than a clean leaderboard does. The best rating in the 2026 peloton belongs to a rider almost nobody would name, and the fourth-best belongs to a former world champion who does this job, by his own account, out of loyalty: Philipsen, he said in 2024, "helped me achieve victory last year in Paris-Roubaix, and I don't forget things like that". Mathieu van der Poel's hit rate as a lead-out man is 0.78 across 36 stages, against a field average of 0.65, the highest of anyone near the top of the board. Leading out Jasper Philipsen it is 0.81, across 26. He has put his own number on it. In July 2023, on Sporza's Vive le Vélo, he said (via WielerFlits, translated from Dutch): "If I lead him out perfectly, he finishes it off eight times out of ten." His bar is higher than the model's (a perfect lead-out and a win), but it is the same kind of number, and it is close.
This season Philipsen has beaten expectation on all seven stages van der Poel has led out, and it is worth being clear what that means. One of the seven is stage 7 of the Tour, which Philipsen did not win. "It was a perfect leadout and the sprinter did not finish it off today," Alpecin's Philip Roodhooft said. Beating expectation is not winning. It means finishing higher than his rating said he would, and that day he did.
At Milan–San Remo in 2024, van der Poel, then in the rainbow jersey, chased down the attacks over the Poggio and led Philipsen out on the Via Roma. Philipsen won. Van der Poel finished tenth. "Having won it last year makes it easier because I think I still had the legs to sprint," he said afterwards.
Nothing in the results file records what that was worth, and this is my attempt at a number for it.
References
- Calo-Blanco (2026). Measuring individual contributions to team success in professional road cycling. arXiv preprint 2602.11831, version 2. arxiv.org/abs/2602.11831
- Candelon and Dupuy (2015). Hierarchical organization and performance inequality: evidence from professional cycling. International Economic Review 56(4), 1207–1236. LISER record
- Matthes and Piazolo (2024). Don't put all your legs in one basket: theory and evidence on coopetition in road cycling. European Economic Review 170. RePEc record
- Menaspà, Abbiss and Martin (2013). Performance analysis of a world-class sprinter during cycling grand tours. International Journal of Sports Physiology and Performance 8(3), 336–340. doi:10.1123/ijspp.8.3.336
- Rize, Saldanha and Moskovitch (2026). Bayesian estimation of leader and helper skills in professional road cycling. SN Computer Science 7(5), article 387. Open access. doi:10.1007/s42979-026-04925-6
- van Erp, Kittel and Lamberts (2021). Sprint tactics in the Tour de France: a case study of a world-class sprinter (part II). International Journal of Sports Physiology and Performance 16(9), 1371–1377. doi:10.1123/ijspp.2020-0701