And for those asking "Why not 1.x.x yet?": The API is stable but I'm waiting until the feature set feels truly production-complete. Getting there!
(don't let the version number stop you from using it though)

Link to relevant forum post (only accessible for those who have access to the developer program): developerportal.garmin.com/node/2303

I'm trying to understand the blog: Does the scanning back approach primarily ensure monotonicity and reduce computational complexity, with max effort detection coming from the 85% filtering? Or does the scanning back mechanism contribute to max effort detection in a way that I might have overlooked?

Intervals.icu uses pretty much the same method to find max efforts. Doesn't do any pre-smoothing of the power data. I never tried smoothing.

So there you have it. This was our proposed solution to interrogate real-world (messy) MMP data, identify maximal values in that data, and use (only) those to derive CP.

Final question: Do you apply rounding to the 3s smoothed data? (rounding determines how big of a drop will be considered a shoulder)

Thanks for the extensive answer! Your implementation is simpler than I thought: Because the paper mentioned "deflection points" I actually thought you were using the 2nd derivative with some cutoff.

Joking aside: Of course in some settings this is possible, but there are enough settings (like online training applications, my main clients) where this is not (always) possible/preferred, requiring an unsupervised approach.

I tried, but the .csv files refuse to talk to me...

I do like the shoulder method from Spragg et al. but it needs a cutoff that is very duration specific, making implementation (especially in a wider duration domain) difficult/arbitrary.

Thanks for answering.
Fitting on work-duration indeed makes sense. I still tend to plot power-duration because it's more intuitive (to me).
Do you know of any published work on hull/envelope fitting for power duration models?

@rchung.bsky.social any thoughts?

Some observations:
- Least-squares clearly underfits both W' and CP
- Asym-loss-A overfits Pmax
- Asym-loss-B looks most balanced (my subjective pick), but CP appears overestimated (note: from this data alone, not according to 2-param asymmetric-loss and my own estimations)

Question: Which approach do you think gives the best fit for the 3-param CP model? (Does it matter?)