Originally Posted by RChung

Yup. I've noticed that too. Golden Cheetah uses Phil Skiba's xPower, so it's pretty clear Strava is using Skiba's xPower. XPower uses a fixed "time constant" of 25 seconds which means a fixed 1/25 = .04 = 4% decay rate rather than Andy's fixed 30-second "uniform kernel" moving average.

You almost surely know that Andy uses as defaults "time constants" of 42 and 7 for CTL and ATL. Many people mistakenly think that means CTL is based on a moving 42-day average of TSS. Just like Phil, Andy's time constant of 42 means a fixed decay rate of 1/42 = .0238 = 2.38% per day, and a time constant of 7 means a fixed decay rate for ATL of 1/7 = 0.143 = 14.3% per day. Fixed decay models are all exponential, so Andy called his CTL and ATL "exponentially weighted moving average" models but in other fields we just call them fixed or constant decay.

There's some (medium but not overwhelming) evidence in the literature that Phil's fixed 4% decay model for xPower matches HR a *tiny* *little* bit better than Andy's fixed 30-second moving average, but I don't think the difference between NP and xP is all that huge. GC and Strava went to xP not because of the difference but because TrainingPeaks originally tried to strong-arm them on NP while Phil said "no problem, you can use xP if you want." It was during the strong-arm that I found out that TrainingPeaks had trademarked the name they hadn't come up with.