Strava is a wonderful social media site for endurance athletes to share their rides and runs, praise each others’ efforts, and try to poach the best times on specific routes around town.
I’ve been a fan of it for a long time, as are hundreds of thousands of other cyclists, triathletes and runners. One of the especially neat features of the site is that professional athletes often share their training and competitive logs, so you can look at exactly where they go, how fast they move and how much power they produce. It’s mind-boggling—and addictive for a power meter junkie like me. Even before I had a power meter, it was neat to see what Strava estimated my output to be during a ride. It’s a great combination of science experiment and race against yourself.
But I’ve wondered for a while just how accurate these power estimation algorithms can be. It looks like the guys at ERO Sports have figured out how to be extremely accurate with them. Creating your own personal wind tunnel with a power meter and a velodrome track is impressive stuff. But even that group is putting controls on the elements– homogenous surface, perfectly flat course, going in circles and no wind.
Strava is a much wilder proposition. Their algorithm has no idea what conditions you’re riding in, and attempts to crunch your power data just based on your speed, weight and the road gradient (which, since it’s determined by GPS, has its own degree of associated error).
Then it occurred to me that I could actually test it out.
I rode in an 80km time trial back in March. The course is on Fiesta Island in San Diego, which is approximately a 4.15-mile loop. The map from my Strava profile is below. The wind direction was generally northeast at 5mph, giving us a pretty good push from the start/finish line and a tougher go of it on the back stretch, just for reference.
I ran the comparison by uploading the file to Golden Cheetah cycling performance software and creating a copy of it on my desktop. Then, using the ride editor, I deleted all the power data cells and saved it as a .tcx file. Strava allows both formats to upload to their site. So I just uploaded both. Then I copied and pasted all the power readings for the segments designated on the course. There were anywhere between 12-15 segments per lap of the course. I was surprised that the number did not remain constant. I’m not sure why Strava only recorded performance for some segments on particular laps. Whatever the case, it did record the segments the same way for each “version” of the race. So the data remained comparable.
First, overall data. Strava estimates that my average power for the ride was 218 watts. Using the power meter, both Strava and Golden Cheetah say that the average was 225 watts. That’s an error of 3.1%. Not too shabby!
I then compared estimates versus measurements on a per segment basis. The longest segment on the course is 6.8 miles. The shortest is 0.2 miles. The race was 12 laps, so lots of repetition here. When it calculates output on segments, Strava’s average error increases to 7.3%.
So what happens when Strava estimates power at shorter distances?
The problem is that sometimes the algorithm overestimates and sometimes it underestimates. Over a long enough distance or time, those pluses and minuses average each other out. But what I looked at was the absolute error on each individual segment. In other words, I didn’t care if it was under or over by 5 watts. I only accounted for how much it was off. So a negative 5 watts on one segment and a positive 5 watts on another one weren’t able to cancel each other out. This was the observed error on 340 total segments, or approximately 28 segments repeated 12 times in a period of 2.2 hours.
Going a little further into the data, I found that the algorithm’s error was greater than 10% on 65 of 340 segments, or 19.4% of the time. It was greater than 20% on 26 of the 340 segments, or 7.65% of the time. The largest error I found was 42.7%. Graphically, here’s how they rank against each other. The graph of actual measured power is on top, and the estimated version is below.
It’s worth noting that this was a flat course with significant wind effects. I may run a similar analysis on a ride with bigger climbs and see how it works. My suspicion is that the Strava algorithm will be more accurate when aerodynamics are less of an influence.
I think this indicates some valuable things for people interested in upgrading to the “premium” version of Strava.
If you’re interested in training with power but don’t have a power meter yet, I know of few better introductory tools than Strava’s system. You won’t be able to effectively use power to help you train or race until you actually have a power meter on your bike, but Strava at least gives you a good indication of what a 200 watt effort feels like versus 300.
You can play with the data to an extent and get a “feel” for things before you decide to dive into a meter. Once you do that, it’s up to you whether you want to also spend the bucks on the premium subscription or work with the Golden Cheetah software for free. Personally, I’m a cheapskate. I also think that Golden Cheetah gives you a little more capability to change how you view and assess your data, but that also incurs the extra responsibility of knowing what you’re looking at and what you’re looking for. If you just want the basics and a nice, clean visual, go with Strava premium.
The ultimate finding here is somewhat of a nuance, since the estimator only has a relative value for a person without the ability to use a meter in real-time to begin with. You can’t really gauge your performance (or that of anyone else, for that matter) on a segment unless they actually had a gauge.
Thank goodness all those King of the Mountain titles are based on time!
If you’re interested in getting faster, you’ll be fascinated by FASTER: Demystifying the Science of Triathlon Speed. In Faster, astronautical engineer and triathlon journalist Jim Gourley explores the science of triathlon to see what truly makes you faster—and busts the myths and doublespeak that waste your money and slow down your racing. With this knowledge on your side, you can make simple changes that add up to free speed and faster racing.