Not my link. I do not know anything about a Valencia Marathon study. Did that study share the number of observations for each device type, the mean for all device observations, and the standard deviations? If the author does not understand statistical analysis, there is a pretty strong likelihood that the presentation is not statistically relevant.

I do know that my Garmin watches (multiple) are an order of magnitude more precise than my personal Apple devices. This is based on hundreds of runs and comparing standard deviations between the two types. I believe that the design of fellrnr's analysis is flawed in that it looks at individual activities in a way that exaggerates absolute GPS position errors that wash out over an entire activity. And, they virtually vanish when looking at many activities over time.

I prefer my 935 over the Fenix.
Had the watch for about 2 years now but still looks new. Few friends have the Fenix 5 and the wear & tear are showing through the powder-coat & doesn't look quite as nice outside of training/racing.

Exactly correct on the note about statistical relevance. I’m guessing this is just an aggregation with no significant statistical analysis or control. Also, it’s impossible to control for the actual path traveled (tangents, etc.) which can add distance. Running tangents is typically the most accurate, especially with a lot of turns. I’m not familiar with this course.

I would go with the 935 and not a refurb.

Oops, sorry for the misquote.

Yes it includes the sample size for each, as are means. Since there are error bars, I presume confidence intervals were calculated using a standard deviation and not just made up.

Apple seems to be spotty for multisport. I'm more curious about legitimate Garmin competitors like Polar, since fellrnr says the following: "The 935 has mediocre GPS accuracy at best, and I would not want to do any serious training using the 935 GPS. The level of error is large enough to really screw up a marathon training plan."

Due to the flawed design of the fellrnr test protocol and analysis, I would not trust any conclusions. In most cases, the foot pods were calibrated by GPS. Let that sink in...

Or, here is another simple illustration (the reality is more complicated, but this is a start)... say GPS has a ~20' accuracy for a given location and a foot pod has about a 0.3% accuracy per distance traveled. If you compare GPS distance versus foot pod distance for 1/4 mile using just two data points, the foot pod's error would be about 4'. That is pretty freakin' awesome. On the other hand, GPS might be off by 40' (3%). But go the other direction... compare total distances over 10 miles instead (still with just two GPS data points for the sake of the illustration). The GPS error would still be around 40', but now 0.08% as a percentage of the distance measured. Foot pod would still be 0.3% and 160' off.

Given that high-level information, think about how you would design a protocol to exaggerate the nature of GSP measurement versus the nature of foot pod measurements? Fellrnr pretty much hit it out of the park, and he is unaware of what he did. If I were running short distances in caves, tunnels, trees, and treadmills, foot pod is a no-brainer. That is effectively what his testing demonstrates, and little more.

Good explanation, thanks. Makes sense for footpod vs GPS discussions, but what about the variation between GPS units on the same protocol?

I have a 935 and charge it when it looks like it's getting low. About every couple of days. But I'm sure it could go a week or more without a charge. It has never died on me during a workout. I do, however, fully charge it before going on long, multi-hour workouts or ski days. Overall extremely happy with the 935's battery life.

---

"The first virtue in a soldier is endurance of fatigue; courage is only the second virtue."
- Napoleon Bonaparte

The problem is that his protocol injects a significant amount of random noise in every device's data that masks the true accuracy issues that devices might have. I suspect that the absolute GPS position accuracy errors are even double-counted in the way he does his tests and analysis. Also, he does not apply a reasonable weighting to the things tough for GPS, which again introduces more random noise that invalidates any usefulness from the study. It's a shame, because he is collecting a crapton of data, but it is done in a way that masks real-life performance differences.

For example, his statement that the 935 is so poor he would not use it for marathon training fails the rational man test. I suspect that is driven by the random noise above. I have hundreds of miles running with a friend with a 935. His watch consistently matches mine within 0.2%. Most often, our watches are identical. That seems like pretty amazing precision to me, and those routes typically have tree cover, turn arounds, and semi-trail

If he had just done a single 5 mile course and counted that as a data point, he probably could have eliminated the random noise by an order of magnitude. Then, he has a very challenging route that is probably far worse than people encounter in real life. His intention is noble, but some watches may respond to turns poorly, while others may respond to obstructions poorly, and others may handle long straights poorly. Designing tests around those scenarios might also help people understand products that might match their real-world conditions better. As it is, the protocol is designed to make all GPS throw up without knowing why.