- 12 workouts, not 12 weeks. It’s listed in the release post and support doc.
Yip I wasn’t sure, thanks 
See my post above. Last week was rest week and Wednesday, Jan 26, 2022 was my last ramp. (I felt I was overshooting which is why I did the extra ramp then)
Chad has confirmed its 12 wo’s at least not 12 wks of workouts.
My training has been consistent since late November, but I haven’t done much of anything over threshold besides over/unders. I’m wondering if that is somehow influencing the prediction.
I wouldn’t think so - that’s the case with most users who are in base.
I ran the FTP estimate out of curiosity, and it gave me a 1.1% increase. I am in the last week of my second block of build, and somewhat fatigued at this point. I am curious as to what my ramp test in two weeks will output as a result, and how that window right after a rest week would be reflected in the FTP estimation. Given that we get stronger when we recover, how is the estimation process taking that into account.
I think the estimate it gives is what you would get from a ramp in a rested state based on training history and other factors.
Not what you would get right now in a fatigued state.
I’ve just checked your prediction on our end and am seeing a slight increase (+0.88%). Could you please try again and if you’re not seeing something similar, please reach out to Support.
Interesting. I’m seeing the same thing as you now. Thanks for taking a look!
Nice! We resolved a bug in the model today related to some workout data not being read properly. Could have been why you were seeing a bad prediction.
Depends on the PM. I believe Assioma pedals take chainring shape into account, and others may have some type of adjustment as well. Or, I could be thinking I read something that didn’t exist and you’re anticipation is correct:)
I’ve got assiomas but you can’t enter shape into it. Plus I was listening to a developer who was explaining thru have to assume circular chainrings for precise measurements. Maybe stages is fine because it measures how much the crank flexes, rather than the power input.
It’ll still be accurate with ellipticals but it just won’t be precise. I think anywhere between 0.5 and 5% has been seen but 2% is common. But before it gets too off topic! At least it’s accurate so it won’t bug the AI out with the switch too much 
I have some interesting observations on the FTP estimation model.
- Last night after a hard workout, it showed 353
- This morning before my workout, it showed 354
- This afternoon after an easy spin (Bald Knob -1) it showed 357.
They all seem to feel about right.
There definitely is some genius going on with this ML model (and not just because it keeps revising me up, lol). It seems to be accounting for freshness and the time it takes for prior training to turn into fitness. I can even see it redefining the taper, as it will know exactly what to do to produce the highest number possible for race day. This is some really cool stuff. It will ultimately be a necessary tool for coaches.
What you are probably seeing is “noise”. The predicted FTP is actually a range / confidence interval, and you are only being shown the midpoint. So it would be expected that the midpoint could move around day to day, with the possible variation being influenced by the size of the confidence interval.
Assume for sake of this example that the prediction only uses your last 12 workouts. So everyday that you run the prediction, one workout gets dropped, and a new one added - assuming you are doing a workout every day. So depending upon what happened in the dropped workout and the new one, this slight change in the inputs to the prediction model will lead to a change in the output - the predicted FTP. The possible magnitude of this change is governed by the confidence interval of the prediction. Another way to think of the confidence interval is a measure of the sensitivity of the prediction to a slight change to the inputs. A tighter confidence interval means the predicted FTP is less dependent upon the presence of anyone workout. A larger confidence interval means the predicted FTP is impacted more by the presence or absence of each workout
I suspect @AlphaDogCycling has this figured right.
The interesting thing to me is the non-uniqueness of the FTP input with PL and how that affects predictions. I suspect progression rate at a given FTP input - no matter what it is - has to be an important driving factor in the suggested change.
Stated another way, I think this is more reliable in suggesting reasonable changes in a given FTP input than about getting an absolute value. It’s a subtle difference for sure. Be interesting to see how it plays out.
Just done this FTP estimation - Pretty spot on i reckon. Good work TR.
My girlfriends first TR experience was a 40min long ramp test due to a low initial estimate, so not exactly an idea introduction to structured indoor training 
Being able to get a decent estimate off the bat and jump straight in to training would have been so much better, anything to reduce the barriers (perceived or otherwise) has to be a good thing
I think ramp is still best starting point followed by a few weeks of training if stuck indoors. You have to give the system a good starting guess. Least annoying is probably to just do unstructured rides outside in a group that are hard.
Have done ftp estimation or a ramp test in the past two weeks? If so it won’t show up