Marathon Cohorts is one of our most popular tools. It answers a simple question: what kind of training are runners in my goal time cohort actually doing? By grouping marathon finishers by their finish time, you can see the distribution of 5K times, 10K times, weekly mileage, and more for each cohort. It's a powerful way to benchmark your training against others who already achieved your goal.
This kind of data-driven guidance used to require a coach or deep dives into training literature. With Cohorts, it's finally available as an interactive tool. But the original version had a problem: it was built upon noisy real-world data. If you played around with the tool enough, you would notice that the times wouldn't always move in the direction of your mental model.
The issue comes from sampling noise. Each cohort represents a different group of runners, and the percentile boundaries measured are subject to a bit of sampling randomness. Let's look at the 50th percentile (median) 5K times for a few different cohorts:
Notice how the median 5K time for the 4:00 cohort (23:36) is faster than the 3:55 cohort (24:03)? That doesn't match our mental model—slower marathon finishers shouldn't have faster training times. And in general they don't, but if you zoom in enough on the data, you would see these discrepancies.
With our V2 of Cohorts, we've fixed these issues. By looking at the smoothed version of the trends of each percentile, we are able to reliably give stats that both represent the runners in the cohort and that smooth out any sampling error. The chart below shows the difference. Dashed lines are the raw percentiles; solid lines are the smoothed version. Hover to compare specific values.
The update is now live, and you can check it out here. We've kept the original V1 available as a legacy item here as well.