AI FTP Detection overestimating

I have been using AI FTP since release. Some friends and I do monthly challenges, which are essentially some type of solo ride on a pre determined segment. In august it was a hill climb that was just under an hour, so essentially an hour at FTP. My AI FTP was 303 and I rode for an hour at 295. I wasn’t feeling great that day after a very poor sleep, so I’m pretty comfortable that AI FTP is nearly bang on for me at least.


The majority of the feedback we have received is in line with @mcneese.chad’s experience. Of course, as we gather more data, this is something we will continuously refine.

AI FTP Detection isn’t meant to predict the results of any individual effort, because all-out efforts can be affected by nutrition, fatigue, and other subjective factors. It’s thus not meant to approximate or estimate what your ramp test result (or any other FTP test result) would be. What it IS designed to do is give you a training benchmark (FTP) that represents the full scope of your abilities and gives you the most successful and productive training possible.

Ramp tests formed just a small part of the millions of workouts and tens of thousands of individual FTP changes (including all different types of FTP tests and manual updates) the model was trained on. Crucially, it didn’t just look at athlete’s FTP changes, but also the subsequent results of their workouts at their new FTP.

It can be true that your FTP is 303 and your relative strengths vary across zones. This is the beauty of Progression Levels! Your Progression Levels will adjust according to those strengths and weaknesses to ensure your training is suited to your abilities.

We agree with Chad that it’s an incredible way of tracking fitness for most athletes. But if you feel that Ramp Tests or other capacitive tests work better for your training, the choice is yours!


I have been performing pretty poorly recently, but my AI-detected FTP went up another 5 watts. It was already about 10 watts too high, really.

I almost wonder if AI-estimated FTP is sort of like car horsepower; only accurate under certain conditions.

E.g., its what your FTP would be at sea level at 20 degrees Celsius and 50% humidity.

At 1,000 meters at 27 degrees Celsius and 70% humidity, doing the same numbers would be basically impossible.

I think AI FTP detection overestimates my FTP a bit, so I train with my FTP set slightly lower. However, I think FTP detection is an AMAZING objective tool for understanding my relative fitness level. They’ve trained it to estimate ramp test results; I’m sure they could train it to estimate 1 hour power. I don’t care. I find it very accurate on a relative basis and wish I could use it more, and I don’t have to test. Most of my complaints with TR are that I’m not getting enough of TR. I guess there will always be those who are all over the TR forum on multiple threads to preach against TR methodologies in favor of others, but IMHO, they are doing this cycling training thing better than anybody.


I have been thinking about this a-lot recently as I stopped following the plans (I don’t race and have no plans to) and just wanted to do more base/build style blocks and progress my time to exhaustion for sweet spot and threshold.

I think @WindWarrior makes a good point. He only changed his FTP a few times over a few year period and then tested it with a threshold type effort that he could sustain. Not a proxy. Modeled FTP, ramp tests, 20 minute efforts are all proxies. The longer threshold efforts (2x20 or anything longer than 15 minutes) are at the higher PL levels, and if there is a ramp test/button at the end of every 3 week block and it goes up 2 watts - it will bump your PLs down. So you won’t ever progress to the longer threshold efforts while following Adaptive Training- and will keep giving the short O/U efforts (think 9-12 minutes surrounded by Z2) and you can never really get a sense if you can hold this “threshold” or “FTP” number for an “appropriate” amount of time (30-70 minutes). It takes time to progress to 2x20 unless you start at 3x12.5ish (I forget the workout name).

Now if you don’t care about being able to hold your threshold for a long time (And I think saying only people who do TTs should do this is insanely short sighted - as it builds excellent muscular endurance and more importantly mental endurance) then following adaptive training and constantly taking small jumps in FTP that lower your thresholds and keep the threshold and sweet spot workouts with shorter intervals will work and you will get stronger. I think in this use case your “training number”/AiFTP is fine. But at some point that number will not go up (based on many factors) and the only thing you will be able to do is advance to workout with higher PLs and if that number is not set correctly (within a small confidence interval) you will learn very and “fail” quckly as OP pointed out.

So if you keep your actual threshold and work on improving aerobic endurance and time to exhaustion you should not be changing your FTP at the end of every block - only when you test it with longer threshold style tests or if RPE is way off (think 5/10 for a threshold workout rather than 7-8/10). Ego got in the way for me - I loved the AiFTP number. But it wasn’t accurate when I started extending TTE and used RPE to verify. I dropped it 5% and have worked up to 6.8 PL Threshold workouts (grain of salt they are custom, but it was a 2x25 at 100% degrading to 99% FTP with 7 minute rest). The artificially high number also made my high Z2 rides actually low tempo and my sweet spot rides - low threshold (which is unwanted fatigue for a base block).

These are obviously my opinions. I love TR and still use it (mostly using custom workouts) but I think no matter the platform and no matter at what way people derive an FTP it must be verified to have accurate training zones.


There are more ways than one to validate an FTP. Personally, if I am unsure, I validate with a few choice workouts and a hard outdoor ride along a known course. Insisting on a TTE test as the way to validate FTP is just dogma IMHO.

I prefer to trust my legs: I know what lactate threshold feels like in certain situations, and that is much, much more useful for judging whether an FTP — no matter how you obtained that number — is accurate or in the ballpark. I know the power I can hold on a given climb, and what my other data (heart rate, %age of FTP) typically are. Although to be fair, I only do the “full suite” if I had a long training break or the change in FTP is unusual.

I think @mcneese.chad’s method of seeing how he can cope with workouts is just fine if the change in FTP is small. Conversely, a TTE brings a lot of additional complications into the discussion, because it only gives a lower bound on your FTP. So if you pace your TTE test conservatively, the number it gives you is likely lower than your actual lactate threshold 2.

Why can’t it be both?

That is a non sequitur: to lowest order, any form of consistent, structured training will make you faster. Stuff like whether you do sweet spot or polarized, or FTP assessment protocols are much less important.

If I remember correctly (and please correct me if I am wrong), you had a big issue with FTP tests, because they didn’t agree with you physiologically and psychologically. So if I were a coach, it’d make sense to not test someone with that disposition less than other athletes. But I’d be very careful about making that prescriptive, i. e. to claim that this is a better method for the majority of people. Just saying.

Yes, but the more elaborate the testing, the more sources of error you potentially introduce. Especially if the testing method requires you to pick a target power in advance, that requires experience on your part and/or guidance from an expert. And it is can be hard to suss out smaller changes in FTP with TTE tests.

Please don’t take that as me saying that you are doing it wrong or TTE is useless, not at all. I’m just saying that you shouldn’t claim others aren’t validating their numbers if they don’t follow your preferred testing protocol. FWIW, personally, I can very often predict my gains within 2–3 W. (Losses are harder to predict.)


It’s as accurate as the data you put into it. If you always train at 8k feet at 80 degrees inside with no fans on you, and then go down to sea level at 50 degrees and ride outside…it’s going to drastically underestimate your FTP. Or the opposite if you train at sea level and then go up.

In my opinion it estimates me a little bit high, maybe 5-10 watts, but I don’t really care. Given a cool day, a little taper and starting at the lowest altitude around me I could probably do what it says I can for 45-60m. The important part is the workouts it spits out are appropriately hard, and if I want to make them harder I mark a couple as ‘moderate’ and if I feel I’m pushing the boundaries of overtraining or the workouts are beyond my capabilities then I mark them as ‘all out’ or ‘very hard’ and it levels it off or drops me down a rung so they’re suitably hard again. If my goal is to extend time in zone, then I leave my FTP lower than AIFTP recommends so the workout levels go higher and the intervals get longer. If my goal is shorter power and pushing those limits, then I make sure my FTP is set so the workout levels are in the 2’s-4’s.

From experience I have a good idea of what I can actually hold for an hour, two hours, three hours…etc in a variety of conditions/altitudes for pacing ITT type events or long climbs. For group races, it doesn’t matter that much. Alex Wild had a great quote awhile back that I’ve tried to take to heart. Paraphrasing… On race day the number does not matter. You can only do what you can do. What matters is whether you’re training appropriately/consistently.

Except that with TTE tests you don’t actually know precisely what your threshold is. If I can sustain 360 W for 40 minutes and 345 W for 70 minutes, what is my FTP?

Plus, in my experience the difference between my good and bad days can easily account for a few percentage points in FTP.

FTP by definition is a proxy, by definition it is a field test for your power at lactate threshold (2). Even in his earliest works that I am aware of Coggan makes that completely explicit, and he states that FTP ≠ hour power. The latter was just one early test protocol he suggested, because his test subjects were experienced, highly trained cyclists.

You are conflating testing with training. For training it may make no sense whatsoever to prescribe threshold intervals that are longer than 20 minutes each. Long FTP tests like a 1-hour TT simulation or a TTE test induce a lot of (mental and physical) fatigue, so this is likely not a suitable workout. (Of course, there are exceptions, e. g. if you follow a 40k TT plan and you are in the specialized phase.)


My argument is that if you can’t hold whatever number (however you derived it) for 20 minutes with an adequate warmup then the rest of your training zones are not accurate. Maybe for the style of riding they do it isn’t important they do a 20 minute effort - but to determine if the number is accurate so the sub threshold efforts and workouts are in the right zone is important.

You and other people are saying you can “feel” when your FTP is accurate. Great - you are basically going off of RPE. But TR isn’t using RPE to set training zones - it uses a number - which I think most people who use the platform can not hold for even a fraction of time to make their training zones accurate for longer efforts - regardless if they need/want to do them.

I am giving you my opinion, and it’s based off of my experience with ramp tests and AiFTP detection. I think they’re fine tools - but they need to be verified. And not if you want to do TTs or long hill climbs. Because then your Z2 rides are actually Z2 and not tempo and your sweet spot is actually sweet spot and not threshold.


Yes, but who is claiming otherwise? My point is that you don’t need to test it every time. And if I feel the need to validate my FTP test results thoroughly, I include a longer effort at threshold.

No, I’m talking about something much more specific: I can feel the tipping point when I cross LT2 with certain standard workout (like 4 x 10 minutes). RPE to me is much more vague and broad.

You are conflating power zones with TTEs at specific power levels. And you make claims about people using this platform that cannot be verified.

IMHO the biggest barrier to holding “longer” efforts (with longer I mean longer than springs or so) is purely mental. It is why long endurance rides on the trainer get hard: it isn’t the power that makes it difficult, it is the boredom. At higher power levels, it is the pain that you need to be able to deal with. If you test untrained people, they will frequently get nowhere near they physical limits. How should you then set power zones, which, by definition, are delineated by physiological markers?

Yes, and this is what @mcneese.chad and I have advocated for in this thread, too. It’s just that our methods are not your methods. I’m not even arguing that my methods are superior, just that they work for me, and that I think many people here royally overthink this. Things like “It isn’t FTP if I can’t hold the power for x minutes.” not only puts a lot of pressure on you as an athlete, it can lead to you setting your zones incorrectly. You correctly mentioned this before: if your FTP is wrong, your training zones are off. You focussed on what happens if your FTP is set too high. But also training with an FTP that is too low has negative consequences. Threshold workouts are then sweet spot workouts, and for some training philosophies (e. g. polarized), the difference between the two zones is essential. Over-unders could become under-unders. Ditto for certain VO2max workouts.

Do you need to roll out your full test suite for a 10 W bump in FTP? Probably not. Especially if the increase roughly corresponds to what you have come to expect in that period of your training given your training history.


This is not true. Once you get up there, a change in FTP will bump PL a bit down, but not enough to never give you a long FTP effort. Just had an aiFTP update and it changed my TH workouts to 5x11, 3x20 and 2x30.

See my profile for how aiFTP and AT will get you there: - TrainerRoad

I think the formula for AI FTP Detection is as follows:
0.95 * 20 min power
0.85 * 5 min power
0.8 * 3 min power
0.7 * 1 min power
Drop the lowest one or two values, and average the remaining values.

This is of course a best guess estimate. It’s a currently between the average of the top 3 and top 2.

Last TR AI FTP Eval came to 336w.
According to my last 42 day durations are as follows:
1min: 526w * .7 = 368.2w
3min: 390w * .8 = 312w
5min: 335w * .85 = 284.75
20min: 284w * .9 = 255.6

Traditional just use my 20 min power ftp: 255w
Average of all four: 305w
Average of the top 3: 321w
Average of top 2: 340.1

I’m still able to consistently complete productive and stretch workouts with my FTP set as is.

This is not how AIFTPD works (according to TR). It can use sub-maximal efforts to give its estimate, thats why its new/interesting - it does not require you to go all out over any distance to get a reasonably accurate estimate.

1 Like

I did a ~20min effort in August which saw me bump my FTP up by 12 watts. AIFTPD instead suggested lowering my FTP by about the same amount. I’m not following a TR plan, but as I understand it that shouldn’t make a difference.

It’s not a requirement to follow a plan for getting AIFTPD results. BUT… it does help if you are getting in data along with the 1-5 rating for workouts imported. That subjective result aids the objective review they are doing of whatever data you give TR via the imported files.

I mention this now since it was specifically highlighted in another response recently, and not everyone knows the ratings are useful even for non-TR workouts.


This is partly what I was getting at yesterday. A more complete thought…

Like any estimate, of anything, it is possible for AI FTP to overestimate or underestimate. You don’t need to accept the estimate and can use your own from a field test or guess. If you are using Adaptive Training, it can manage adjusting workouts if AI FTP happens to be off. You can possibly improve accuracy by rating outside rides on the post-workout survey.


Nope. I neither do long threshold workouts or time trial events. I exclusively race in central park, prospect park and floyd bennett field, and am a newly minted cat 3. This put me in the A field with the CRCA, making me have to do top efforts just to stay on with the cat 1/2/3 guys. I clicked on a cat 2 in the field and he actually ramped on TR as 393w ftp.

1 Like

I will say, this is one of the best parts about adaptive training, it doesn’t give you AIFTP as gospel (must use it or else the adaptions don’t work). Trainer Road allows users (sometimes to the users demise) to customize their FTP based on their own feelings about the result of AIFTP with respect to other tests, RPE, workout quality, anything. And then TR will adapt the plan based on this updated FTP number and the users performance and feedback regarding the work done going forward to adjust this FTP up or down.

Even though it can over- or under- estimate FTP; which, by their own admission, is just the number they are assigning to your fitness as a way to prescribe workout levels, and isn’t really tied to other meanings of FTP (hour power, or power at maximal lactate steady-state). They do usually get pretty close, so hats off to you TrainerRoad.

1 Like

There is no simple, neat formula for AI FTP’s algorithm.

Yes, correct, but I think this is forgetting two things:

  1. Fundamentally, AI FTP was introduced so that people no longer have to test their FTP (with whatever method). We can argue whether it wouldn’t be better if athletes actually did FTP tests or not. But I think that forgets that athletes are individuals and not averages. Many athletes hate FTP testing. I personally like them a lot and am looking forward to them, but I know I might be in the minority.

  2. The question isn’t whether AI FTP can be off or not, but whether it performs better than other ways of determining your FTP. Also field tests can give you numbers that are higher or lower than your power at LT2. Generally speaking, the longer the test, the more likely it is that athletes under test. In my experience, AI FTP also seems to err on the conservative side, since underestimating your FTP is less problematic than overestimating it. Still, when your FTP is set too low, this also has undesired effects, e. g. over-unders could become under-unders, and you aren’t actually training to deal with lactate in your legs.

Good that you mention it, I think this is important and needs emphasis. TR has been designed to accommodate inaccurate FTP estimates, and you have the tools to address that. No matter how you determine your FTP, the proof is in the pudding: are you able to finish workouts consistently or not? You know when all workouts seem too hard, and you might know the difference between being slightly over vs. slightly under your lactate threshold. Trust your legs. You don’t need any complicated testing procedure, listen to your body.


For to be reflective you need hard efforts. You can vary what time period it uses.
In prior years I would ramp tests and typically being frequently riding hard tended to have a higher eftp then my TR FTP.

This year had my own eftp at more then 30watts below my TR AIFTP.

I just havent done any hard rides this year of any duration of intensity that would trigger icu to update.
TR has been seeing my 2-3 workouts a week of TR workouts and my outside endurance rides of 1-2 per week. My AIFTP had been slowly trending upwards over the last year.

IN reviewing my own power for 2022 I noticed most of my PRs this year were 5-10 watts below what I would normally do. I was riding a lot easier but was curious if I had really regressed. I had the same question of is AIFTP too high. Roughly 2 weeks ago I did a 30 minute ride at what I considered threshold pace. It wasnt overly hard but was nice and steady and consistent throughout. That ride resulted in the best power I have ever done in 7 years of riding for the 19 to 30 minute time period. It wasnt a huge increase but it was the highest…I am now 62 and since I hadnt been riding hard I was thinking my power was not improving.

So I have no idea if AIFTP is overestimating…but it feels like it is in range for me…I can do my workouts…and while I havent been pushing the intensity other then short intervals in the workouts I am able to ride outside easier with more power when I try. AIFTP for me is working to enable my workouts to be productive. I follow TR low volume plans but only do 2 of the prescribed workouts. I do two endurance rides as well. So it is dealing with a mix of data with my rides. I do believe it is working well.