AI FTP Detection overestimating

So, let me start by saying that I am very much on board with AI FTP detection and I think that using AI to track progress without needing demanding tests is great and TrainerRoad is doing great work.

With that said however, I think that the ramp test overestimates FTP (for me specifically, but I also think that it holds true for most riders out there). The ramp test puts such a burden on the anaerobic system that the aerobic system (which FTP should really be an indicator of) doesn’t get properly tested.

AI FTP detection uses TR’s treasure trove of athlete’s data to do its magic and come up with really good estimates of an athlete’s FTP, but most of these are FTP as determined by a ramp test, correct?

My most recent AI FTP detection estimate was 303w (which I think is a little high), but as I am starting a base block of training, I wanted my FTP to be more reflective of my actual MLSS power as working below this level is really what I think I should be doing during the base phase and increasing my TTE rather than just seeking improvements in my FTP.

I tried to perform Kolie Moore’s Baseline FTP test, which is:
good warm up (20-ish minutes for me) at zone-2 power
10 minutes 92-95% Target FTP
15 minutes 100% target
gradual ramp up from there until exhaustion

The average power for everything following the warmup is the FTP and the time that you are able to do the test is your TTE. To be a valid test, you need to be able to complete the 10 and 15 minute blocks.

I picked my target to be 294w (the same FTP as AI FTP detection picked for me at the beginning of my last block), thinking that this was probably doable, but not overly ambitious… I failed 5 minutes into the 15 minute block…

Remember, AIFTP detection estimated that my FTP was 303w right before I started this test. Are any others of the same mind as me? Do we think that AI FTP detection is over estimating FTPs? Or am I an outlier here?


Yes you’re not alone. Mine has also been about 5-10 watts above what I want. I am taking a different approach and making sure my threshold level is at 5 or above. This way I can keep working on my TTE and build on that. I’ve thought about writing to support but since I’m not following their system to the “T” I figured I wouldn’t . Last time I started with what they gave me for an AIftp my levels for threshold and Vo2max were so low I personally didn’t like it. I want to do longer intervals. Nothing against TR and still love what they’re doing for me, I’m just taking a different approach this year.


So far for me, I think it has been spot on. But I’m also one of those for whom the ramp test is fairly spot on and perhaps a bit on the low side. I have Meall (5x11m hard starts) this Saturday, in my 1st week after aiFTP, will let you know how it felt.


AI FTP seemed to be overestimating for me but that is mostly to do with the varied/ inconsistent data I’m giving it :joy:

1 Like

GIGO :stuck_out_tongue:


That’s what I’m thinking, as good as the model is, if the algorithm is trained on FTP test data that gives overestimated FTPs, that’s what it’ll give you.

1 Like

Meh, that is not exactly my point. Personally, I am getting far better results with AIFTPD than any testing I did since joining TR back in Nov 2015.

My point is that the actual data we give TR (power via imported workouts AND the ratings of those rides) is HUGE. I’ve seen examples of people using multiple devices with known disparities, applying one riding position for most riding (regular hoods) while using that AIFTPD value for stuff like extreme TT positions and other use cases where the data is not exactly representative of the rider as a whole or specifically.

It’s actually tricky to a degree, with all the potential impacts that exist but are ignored in TR. Power meter data for one, bike in use for another, and position on said bike(s) for a third. Include other variables like weather, inside vs outside and any other known influences and the picture can sure get messy.

To be honest, I am somewhat shocked at how well AIFTPD seems to be working on a larger scale. We do see edge cases and other instances where it is shown to be “off”, but those seem relatively rare from my viewing of topics and discussion here. Only TR really knowns, but if the general axiom is true here (we hear more about problems than successes), I expect that AIFTPD is working well for a large majority of users.


When was your last field test (indoors or outdoors), and how long were you able to hold your AI FTP?

1 Like

Last Ramp Test was in January 2022, before adopting AIFTPD values. I have not done a strict FTP-like effort in that typical 40-70 minute range. I missed out on two TT’s I normally do as one of my potential tests.

I am basing my evaluation on progress in workouts through plans since I adopted AIFTPD when released earlier this year. I am also considering my general performance in a range of events (super short BMX up to 1.5-3 hour races, and 5 hour centuries) this summer. My TR account is public and free to review for anyone wanting to dive in to the training and event results with all my performance notes.

Per the basic TR ethos, I am looking at this largely from a “Is it working?” aspect. The AIFTPD values I have used all season have seemingly set good targets for my workouts and allowed me to complete TR AT plans without major modification and minimal “blow-ups” while still pushing me to get better feelings in and results at my events this season.

Compared to years past with testing methods from the 2x8m, 1x20m and even the Ramp, I am having less trials and tribulations surrounding what my FTP is, as well as the workouts that it drives. For me at least, that is a success even if I lack testing related to it.


ok, the simple answer is you haven’t validated the TR FTP estimates. The adaptive plans should be better for the vast majority, the previous plans were essentially cast in stone and largely ignored where the user was in their fitness journey.


but is that due to AI FTP being correct or because progression levels add extra buffer for error?

It’s more basic than that.

I have not tried to make a conclusive set of result for people to compare, just giving my $0.02 opinion. But in service of your question, a quick mental review:

  • I have been using AT (with Progression Levels & related Workout Levels) since it was introduced in March 2021. I had the later half to 2/3 of my 2021 season all the way to Feb 2022 using AT + Ramp testing. My results were decent, but I also had some questionable moments than I remember from this most recent season.

  • I have been on AIFTPD + AT since Feb 2022 to the present, which is a similar start about 1 year later than above. My personal feeling is that I had less stress around the lack of testing, and I feel my results in workouts based upon the AIFTPD were more consistent and reliable than the year with AT + Ramps.

  • I know I had more manual concerns and likely adjustments to my Ramp test results in 2021, in effort to account for the oddities I sometimes presented (quitting too early in many cases, digging just a bit too deep in a few).

  • As ever, N=1 and subject to so much other stuff that happens over 2 seasons, but for me… the results seem clear. I had as good and likely a better season of training with the combo than just AT alone.

  • Here’s a pic of the last 2 years, showing the relative lack of FTP change, but still progression this year vs the varied up/down and more frequent adjustment of FTP in the prior year. This year was more “set it and forget it” with acceptance of AIFTPD as well as suggested plan changes.

  • I did some minor tweaks to sub in SS workouts on Sat’s when I felt smoked and didn’t want the scheduled VO2 or Thresh, as well as adding my own strength and/or endurance rides on top of my Low Vol plan. But those are minor and expected tweaks AFIC. In a way, I seem to feel like I did more/better with the LV + AT + AIFTPD this year than I did with the prior year that was MV + AT that appeared like more TSS at the very least.

  • Just sharing my experience and opinion for reference.


curious how your AIFTP value will change this fall compared to last year with the substantial drop in volume. For me in previous seasons, my ramp test based FTP values drop less than long form sustained type efforts when my volume drops significantly like yours, which may not be picked up in ramp testing but may be picked up in AIFTP. I usually only see a minor drop in FTP likely due to sustained anaerobic contribution to a ramp test and MAP, whereas true threshold drops pretty fast for me if I’m not doing those types of efforts regularly and/or lots of endurance and tempo to keep CTL up.

I am curious to how that will go as well. I got a glimpse in late July as you can see there is a drop in AIFTPD. That might partly down from volume, but also related to some of my outside riding not “counting”. A mix of BMX racing that I don’t record but is super taxing.

That lowered AIFTPD felt a bit “easy” within a few Thresh and SS workouts, but I left FTP alone and only adjust Workout Intensity a few times and no more than 103%. Could be more of the same as I am done with event and “training” for a while. That may change if they get the Workout Levels V2 in Early Access in the near future.


shouldn’t those count towards AIFTP still? Unless they were all without power

1 Like

Yeah, sorry. I got that all messed up with mixed thoughts :stuck_out_tongue:

Trying to edit to clean it.

I feel you, my CX races end up with like 100w avg power and HR near and above my LTHR, theres no way to fully encompass the stress from rides like that. Probably even worse with BMX since the bike is so different as well as upper body demands

1 Like

If anything, TR AIFTPD seems a bit low for me.

All of my workouts have been outdoors for the last couple of months and I’m consistently doing higher numbers than prescribed when doing threshold & VO2 sessions. eFTP is also almost 7% higher than AIFTPD.


what type of threshold intervals? Like 2x20 or 2x30 at greater than 100% FTP?