Module 5, lesson 2 · 40 minutes

Start Right: what it is, what it changed, how sure we are

What has been reported about Marcel Altenburg’s Start Right model, which of its results are well supported, and which rest on a single promotional source.

You’ll be able to

describe Start Right as it has been reported: who built it, what goes in and what comes out;
state each reported result with its source and how far you can trust it;
explain the method as a space–time flow model, and say exactly where reporting stops and our reconstruction begins.

“Altenburg projected that the largest finishing wave would consist of 1,366 runners. There were 1,367.” The World Marathon Majors magazine called it “99.93 percent accurate”.

It is the number that made Start Right famous, and very likely part of why you are taking this course. It is also an excellent place to practise the most useful habit an expert can have: asking, for every claim, who said it, how they could know, and what it would take to be wrong. This lesson does that for everything reported about Start Right.

Marcel Altenburg

Marcel Altenburg is a Senior Lecturer in Crowd Science at Manchester Metropolitan University (MMU). He came to crowd science through operations rather than engineering. He served 12 years as a captain in the Bundeswehr and holds an MA in Human Science from Helmut Schmidt University. In 2009 he interned at the Berlin Marathon and then supported its control room for about three years before taking the MSc in Crowd Safety and Risk Analysis at MMU. He began working with New York Road Runners in 2016, and he runs ultramarathons.

Start Right’s public record, as reported. Dates are the reports’ own; the length of the Berlin period is approximate.

That record puts him where race operations meet crowd science: years in a control room, then the tools to model what a control room sees.

What Start Right is, as described

MMU describes it as “a unique mathematical algorithm to help race organisers visualise the flow of runners through the entire course”, which “makes 8.6million calculations to estimate the volume of people at any particular geographical point”. That is the university’s own description of its own work.

The fullest description of how it is used comes from two articles:

Inputs and outputs. The World Marathon Majors magazine says the model takes start times, course layouts and runners’ pacing data and predicts crowd flow, which race organisers use to place aid stations, toilets and medical staff. It says the model is used at “hundreds of events each year”. The magazine is a promotional publication of the Majors themselves.
How it is used in New York. Tom Vanderbilt’s 2022 article in Outside reports that the model uses data from previous races and “a hope that people more or less run at the pace they have said”. It outputs runner counts per five-minute finish window, works to a design target of “three square meters for each runner”, and recommended that New York go from 12 start corrals to 15.

Races named as users include London, Berlin, New York and Sydney. The magazine says Altenburg has worked on the planning of all seven Majors and in the control room for five; Running USA names London, Berlin, Chicago, New York and more than 100 other races, and reports his work on Grandma’s Marathon in 2021.

No peer-reviewed or technical description of Start Right’s method has been found. Everything above describes what goes in and what comes out. How it gets from one to the other has not been made public, and nothing on this site should be read as a description of the actual algorithm.

The results, and how sure we are

Table 1 lists every reported result with its source and a plain statement of how much weight it can bear. Read the third column as carefully as the second.

Start Right’s reported resultsFrom the most to the least securely sourced.
Race	What was reported	How sure
Boston 2026	“We have simulated the Boston Marathon more than 100 times to run it once for real.” Six waves instead of four; bus unloading, aid stations and finish chutes also changed.	Verified. Associated Press, 20 April 2026.
New York	Busiest five-minute finish window: 1,366 runners projected, 1,367 actual; “99.93 percent accurate”.	Verified as reported in two outlets. The race year is not stated. One window in one race.
New York	The model recommended going from 12 corrals to 15.	Reported by Outside, 2022.
London	Race director Hugh Brasher: the event “grew by more than 27 percent since 2019” while density “at certain key points” fell “by up to 28%” compared with 2019. The same article elsewhere credits “a 30 per cent increase in runners” and density cut “by 40 per cent” at Cutty Sark, with no timeframe.	contested The sources disagree. Use the director’s figures: more than 27% growth, density down by up to 28% at certain points.
Berlin 2019	“Helped reduce end-of-race congestion by nine per cent” and “allowed 4,500 additional runners”.	Single promotional source. No baseline is given for either number.
Lying-in-state queue, 2022	A prediction “within 50 metres and 14 seconds”. MMU’s own news says only “crowd control for the Queen’s funeral”.	Low. Reported only by the World Marathon Majors magazine, which does not say what was predicted.

Two things stand out. The best-supported result, Boston, is about a decision (six waves, many simulations) rather than a measured outcome: we know what changed, not how much it helped. And the outcome figures, London and Berlin, come from a magazine published by the Majors themselves, whose job is to celebrate the Majors. That does not make them wrong. It means they have not been checked by anyone without a stake in them.

Reading the famous number

Go back to 1,366 against 1,367. What does it show?

It is one number. One window, in one race, in an unstated year. A model can match a single number for reasons that have nothing to do with its quality, including luck. Move a handful of runners’ finish times by a few seconds across a window boundary and the count changes, so matching a count of more than a thousand to within one is partly fortune, however good the model.
It is the easiest output to get right. As lesson 3 shows, the busiest five minutes at the finish depends mostly on how many runners there are and how their predicted finish times are spread, and much less on the details of the course. It tests the field data more than the course model.
“99.93 percent accurate” is just 1 − 1/1,367. It is the error on that one count, not a measure of how accurate the model is in general.

None of this says Start Right is poor. The decisions it informs, such as more corrals in New York and more waves in Boston, are exactly what a good flow model should change. It says the famous number is not the evidence. What would be: predictions published before many races, at many points along the course, with the errors reported afterwards. This is the same standard Module 4 asks of any crowd simulation.

A colleague says Start Right is “validated” because it predicted 1,366 and got 1,367. What two or three questions would you ask before agreeing?

Good questions include: How many predictions were made, and how did the others turn out? Was the prediction recorded before the race? Which year, and was the model tuned on data from that race? Does the busiest finish window test the course model at all, or mainly the entry list? What were the errors at points along the course, where the density risk actually is?

What kind of model it must be

Knowing only the inputs and outputs, you can still say a good deal about what a model must do. It takes start times, a course layout and runners’ paces, and produces counts at any point on the course and runner counts per five-minute window at the finish. The simplest model that can do that is the space–time calculation of Module 4, lesson 4:

give each runner a start time from their wave and a pace from their predicted finish;
place each runner on the course at each moment: distance = (time − start) ÷ pace;
count runners in each stretch of road at each moment and divide by its area;
count arrivals at the finish in five-minute windows.

A model of this kind would work this way. Whether Start Right does is not public. It may, for example, correct declared paces using data from previous races, as Outside suggests it uses such data; it may model runners slowing in crowds, spectators, or the walk to the corrals. We do not know. The planner in the next lesson is built on the four steps above and labelled as ours throughout.

One more hint from the reporting. Outside’s phrase “a hope that people more or less run at the pace they have said” tells you where the biggest uncertainty sits: not in the arithmetic, which is simple, but in the input. Predicted finish times are self-reported. A model of this kind is only as good as the entry list.