AI FTP Detection

Hey Everyone!

I was wondering if AI 4DP Detection is ever going to come to Systm, as I personally think it would be a fantastic feature to add.


But then we wouldn’t have all the joy that Full Frontal provides. :nauseated_face::face_vomiting:


Please not!
@David.McQuillen.KoS @Coach.Mac.C
Seriously, please don’t go this route.

@Beau_Goddard a huge assumption here, and likely incorrect but you are welcome to correct me if so, you got this from the TrainerRoad crowd, and no doubt you are on their forum, and the debacle with AIFTPD when they started it was diabolical as everyone was keen to have their ego boosted with a better FTP number, then the wave of uncomplete workouts started happening, and then the wave of distrust to the detection process, and the wave of users manually decreasing their numbers on their own, and the the wave of frequency of when AIFTPD is undertaken, all of this with an unnecessary noise of disgruntled users who didn’t get/don’t have the (correct, whether up or down) numbers detected for them.

Very easy, do the hard yards and do a FTP test.

I really hope WahooX does not employ an AI process for FTP. Even if it is given as an option.


I definitely get what you’re saying, and by no means would I want it to replace any fitness tests, but I think it has a place in training. I have been using for a while now, and their ftp detection algorithm is very accurate for me.

And the benefits of this are allowing me to gauge my fitness during the racing season, where I would not usually have time to do any fitness tests. Additionally, as someone who has just started significantly ramping up training volume, I am seeing rapid increases in fitness, and the 6-weekly fitness tests in my training plans result in large ftp jumps between each (which could be avoided using ftp detection, and optimise my training even further).

With that being said, there are definitely some other features that I would prefer the people at Wahoo to focus on before considering this idea.


So, why not stay with for your FTP assurance? Why bring another platform into the equation, which will cause some doubt over any number that is different to

Everything else you contend need not be stated as you could just rely on’s FTP offering. Using thier offering prohibits everything else standing in your way, and you only have one platform to guage your progress and not be drawn into a minefield of which platform is giving you the correct numbers.

Using a second platform which also offers a detection algorithm is going to cause you to lose time and take your focus away from optimising your training as they will not be spitting out the same numbers and you will be spending said time questioning the why.

The above is the crux for me, not aimed at you, but you do raise it and I have seen it across this platform and TR. Even where RGT proposes a higher FTP number than the latest 4DP test done.

A user prescribes to a particular platform, likes it for the FTP number they are given, and when interacting with other platforms, start to question either that specific platform or the others giving different numbers. This tactic doesn’t change regardless of what/how all the platforms supply the holy number.

Almost as if the user is searching many platforms to get the highest number. And then doubting other platforms which give a lower number.

Offering an AIFTPD is going to produce a lot more noise in this regard, as evidenced at TR. To me, its a bird nest where SYSTM need not lay its eggs.

I am of the stance to ascribe to a erstwhile training platform, and submerse myself in their science and believe they have it right.


Using me as an example…I have only done SYSTM, Zwift, and outside rides for past six months or so (and about half of them with only HR and no power) but have been syncing all rides to TrainerRoad.

Last week TR FTP detection suggested my FTP was 270. I didn’t really believe it (too high) so did a ramp test…which gave me an FTP of exactly 270. Quite impressed.

I have a FF scheduled for next Sunday…would gladly take a (good) AI estimate instead. The key, though, is that it will only work if it takes into account all my rides across all platforms and doesn’t solely rely on power. SYSTM is nowhere near that today…so FF it is.


I’m going to bang the same drum I always do and say that FF is a specific testing protocol. You cannot compare its results, even the 20min power measurement, to any metrics outside System. Not even HM really, though they are known to be close for some riders, and used as a guide, as they’re designed by the same team.

That’s not to say that some form of adaption couldn’t be incorporated (I’d bet good money that the Wahoo team have considered how and if this can be done, but changing the testing protocol shouldn’t be done lightly).


One big difference (at least off a reading of the TR AI FTP press release) is that still requires maximal efforts in its FTP estimations, where the TR method does not. The method will dynamically pick a few of your best efforts of arbitrary duration, but if they’re not truly maximal efforts it won’t be accurate. Despite the name, it seems like AI FTP is estimating more than just one metric but rather some representation of your whole power curve. I haven’t been a TR user since like 2015, so I have no idea how their workouts and targets are structured (i.e. what’s their analog to 4DP?).

Personally, my guess is it’s feasible to create an ML model that comes pretty darn close to predicting 4DP FF numbers for most people. And I’m guessing SYSTM team has done this already but dealing with the outliers of “close enough” and “most people” makes it dang hard to turn into a shippable product.

1 Like

As somebody who knows a little something about software, and a little about the complexity of the biology/physiology of fitness, I am extremely skeptical about any prescriptive program for fitness, or any estimate of fitness parameters outside of a test.

There are too many variables, too much individual variation, and not enough data.


@Heretic I am reminded of a comment you made previously which I think was - “all models are wrong, some are useful”. I would definitely agree with the complexity of variation and designing something for everyone. Some athletes already see that in their Full Frontal and Half Monty results.


I have made that comment in several places, and probably here as well. It is a well known saying in the modelling/simulation community.

The discrepancy between HM and FF is a good example. The discrepancy between indoors and outdoor performance for some people might be another. You can see it here in the way that some workouts are easier for some than others. The Omnium (which I have not done in a while) is one of my favorite workouts, for some, it is a horror show.

I do 4DP testing (HM and FF) because it helps calibrate the workouts. I do not assume it tells me too much about exactly how I will do an outside course.


@Heretic Sure - I can relate to that one along with AWWNH At the Velodrome.


Damn I remember that one.

“All ok, all ok, seems fine. building up, seems ok, AAAARRRGHGHHGHGHGHGHGHHHHHH”


Ha, I checked the workout history for the only time I did that session. Cadence on the kilo effort:


How I managed to get out of that hole I’ve no idea. I know how I got in to it though!

Similarly, power:




Among others, there are two intense but great workouts I like in SYSTM which are not only great workouts, but as a bonus, also quantify performance metrics. They are the HM and FF. I particularly enjoy doing them in the context of the 7 day Test Prep Plan that includes the HM on day 3 and FF on day 7. Do this one week of workouts and it will automatically detect your 4DP metrics!


I suggested the OP stick to as he said it’s very accurate for him. So offering OP to stick with something that he is comfortable and confident with.

I wouldn’t though. So maybe address this with the OP. :person_shrugging:t3:

I’d gladly use for ftp detection and change my SYSTM parameters manually, but that would not update my athlete profile (sprinter, pursuit, etc) in SYSTM. Would love it if you could modify your profile without having to go through a 4DP, which is accurate but horrible, or Half Monty, which might be accurate for some but not for me.

And on cue, as highlighted above, TR finally releases their AIFTPD to their masses out of beta.

And as mentioned above with discrepancies please don’t badly occurring from the machine learning through all their use cases, there are prescriptive actions to consider.

All rides are taken account, that are recorded with power meters/produce power data. Rides with no power recordings are deemed to be included but it’s better if power meters have been used. There lies a huge outlier. For me personally as not all my bikes have power meters.

The percentages they give are far from 100% accuracy. Flip the meat of the skeleton to something that is meaningful, a doctor who doesn’t have a 100% success rate to operate on you, you’d go for that? Or a trial in nutrition or other medication that has less than ideal markers being close to 100% days f desired effect, you would take such medication/nutrition? So why accept the numbers it spits out.

Aspects of stress, fatigue, nutrition, weather amongst others impacts recorded ride data, which is recorded against your efforts to give you an FTP.

You would think that with +150 000 000 use cases that the accuracy of machine learning could get closer to 100% accuracy than what it has achieved currently.

So why so much faith in AIFTPD if there are nominal factors that can detract from a true assessment. Because one doesn’t have to sweat and feel the pain for it.

If there were less detracting factors from getting an accurate detection I would be all for AIFTPD. SYSTM has way less use cases to decipher their own system and if a large dose by TR cannot produce near accuracy, this platofrm should stay clear from offering it.

AI FTP detection has its benefits and flaws. As do all testing models and methods.

Personally, I do a lot of Zone 2 riding and volume with some intervals thrown in, but rarely do I ever do 20 minute maximal efforts. So if you look at my profile there is a long slow downward drift of my estimated FTP between my previous test and my most recent test. Then my most recent test bumped my FTP up higher than my previous FTP. So, did my FTP really drift downward for 6 weeks between tests? Likely not, because I actually improved my FTP by doing that training.

Does this mean that has a flawed model? Maybe in terms of how I train. But the actual number it estimates based on my actual tests and maximal effort is pretty accurate. But the time in between tests it doesn’t.

So assuming TR uses a different model, with my volume and Zone 2 training, how are they going to estimate my FTP in between tests? HR? HR is wildly variable based on all kinds of external variables and stressors.

Even tho Wahoo SYSTM relies on the users performing tests at regular intervals, it does just as good of a job (if not better) than the AI estimating used on other sites. And if I smash a workout today and the AI jumps my FTP, does that mean my recovery ride tomorrow will be harder than it would have been if I hadn’t hit yesterday’s workout as hard (since it will use a percentage of my new FTP)? And does that really make sense?

Your FTP can vary day to day based on context and all kinds of external variables that a computer model can’t take into account. So, it will never be 100% accurate. But SYSTM’s 4DP model has set up a specific set of circumstances that can be accurately and reliably repeated.

Sir Neal and Wahoo have continually said that they are not using AI FTP estimation because in their studies they’ve found it to be less accurate than their current testing model. Especially because SYSTM workouts are based on more than just your FTP (i.e. 4DP) which makes it work better for training than training that is solely based on FTP percentages. And it would be much more difficult to have AI estimation of all 4 4DP metrics. So, that’s why SYSTM is likely not going to be using AI estimation any time soon. I’m sure Sir Neal or Sir David can provide a much better and more eloquent / knowledgeable / detailed explanation.

1 Like

This!!! As much as it pains me to say and to do, FF for the win.

1 Like

This is not the place to discuss Machine Learning which is really just basically a technique for sophisticated statistical analysis. Nonetheless, I would also prefer to stop using the term AI. There is nothing intelligent about a ML model.

Without knowing how they do parameter selecting, or the distribution of the cases, one cannot evaluate the model. Too much of the wrong data can result in over training of the model which decreases the accuracy of the model. As mentioned in other posts, data about fatigue, sleep, and nutrition is probably missing form the data points. There is no discussion on how they avoid bias in the model.