Looking Beyond Averages : Pacing Insights from Big Data

What’s the best way to pace a race effort? “It depends”.

Theory postulates that attempting a course with the highest steady state effort one can sustain without dipping too much into the extreme regions of the physiological intensity spectrum is the “optimal” way to pace.

But with that end in mind, all pacing profiles should divert to a mindless run at constant speed. The issue is that mathematically derived optimal pacing is still “open-loop”. Theory and practice differ and real world athletes tend to differ in their pacing approaches because humans are “closed loop” organisms.

Wouldn’t it be nice to inspect some real world data to see how people pace?

In this regard, two studies will be of interest to the reader. They are unique in that both took a “big data” approach to analyzing publicly available data from cyclists and runners. Below, I’ll quickly describe them and try cut to the chase.

In Empirical Analysis of Pacing in Road Cycling, Saupe et.al studied the pacing profiles of best performances from 12,200 competitive riders on Adelaide’s famous Norton hill climb. They used techniques to equate speed from the data to a power output and to classify data records according to the self-selected pacing profile displayed in the data.

Not only did they discover all types of pacing patterns that have been previously described in literature (positive, negative, even and parabolic profiles), but also some interesting patterns when riders were classified by performance level.

They found that slow riders typically achieved best results using a positive pacing strategy while this was less frequently seen used in better performing riders, who preferred sticking to an even profile. The authors concluded that better performing athletes have faster performance times but also greater “capacity to minimize slowing when external load is high”. I found that interesting.

The second study was from runners so we shift gears a bit. In Effects of Pacing Properties on Performance in Long-Distance Running, Leeuw et.al analyzed over 120,000 race results (10K, 21K and full marathon) of males and females from races organized by the BAA.

In the 10K distance, the 3 common pacing profiles were negative, even and positive pacing, whereas in the half marathon, it was a bit of negative pacing but mostly all positive pacing. In the full marathon however, all profiles discovered were of positive split.

So it appears that as the distance got longer, the “variation” in the types of pacing profiles dramatically decreased leaving little to no differentiation at the marathon level.

It gets even more interesting. The research found that in the 10K and half marathons, optimal subgroups of runners tended to keep their pace within a tight window. The same feature was found in the marathon, and the fast finishers displayed small fluctuations in the accelerations throughout the race, signifying that there was a bigger restriction on pace changes as the distance got longer.

The exact magnitudes of the pace window and the minimum % pace decrease for males and females according to the race format are in the study so I leave that for you to dig up.

To further reinforce the idea that slower performing athletes have room to optimize, a recent study looked into the same 12,000 cyclists from the Norton hill dataset to see how much a computed cycling strategy would compare to the self-selected pacing profile.

In On the Marginal Gains of Computed Optimal Pacing Strategies, computed pacing strategies from Best Bike Split and Powerbike were found to offer slight benefits in a majority of riders in terms of a lower average power for the same finish time.

However, it seemed that the best performing riders had the least to gain from the computed strategy and it looked like they were already self-optimized.

Summing up, what does this tell us?

Average pace or power is way too limited as a metric to describe athletic effort. I think we ought to bring pacing profiles into the conversation more and more.

In these highlighted studies, the best performers displayed the skill and training to “minimize the slowing down”. For the best cyclists on the uphill climb, the widely observed profile was an even profile. For the runners, there was no one pacing profile that fit all the distances but the best performers had some inherent notions about what window to keep that pace within.

And finally, computational tools look promising to offer an optimized pacing strategy for a given finish time. But currently, there is no way to add this into bike computers that cyclists use everyday to add an effective “feedback loop” component.

Fortunately, the data from the Norton hill study show that these are marginal gains at best. If I can generalize, people tend to self-optimize anyway with time, training level and skill.

Feel free to reach out to me if you have comment, corrections or thoughts on twitter.