Although the Steve McGregor's NGP algorithm is proprietary, my understanding is that it is quite similar to normalized power.

As such, I wouldn't expect identical results from two different runs, even over the same course, unless the variations in speed were also identical.

Okay, maybe that's it. Yesterday was "less steady", even though it was identical at the averages level.

Do we know that's the algorithm they are using? Late last year there was a complaint that the TP NGP calculation was conservative for positive/negative grade. Ie, it would give back and take away the same amount of pace for traveling up/down the same hill, such as hill repeats...essentially treating a hill repeat as level ground.

I wouldn't think McGreggor's algorithm would do that.

I've been looking at the release notes, and the original trouble report and there's no indication THAT has been rectified.

The only people who apparently know all of the details of the algorithm are Steve and the folks at TrainingPeaks. (There is also no guarantee that they implemented it exactly as he designed it.)

My understanding, however, is that it follows the same general smoothing/weighting principles as normalized power, but with a few more running-specific little wrinkles.

ETA: FWIW, I have found that rolling NGP and Stryd-reported running power tend to parallel each rather well, to the point that you could argue that the former is really just as good as the latter, at least for post hoc analyses.

That McGregor fellow is a smart cookie...