Post workout responses from completed rides should absolutely impact predicted FTP. It’s real data. Any assumed future response is just guessed by the AI and implemented in a way that introduces propagating bias.
I like the hurricane path analogy as it relates to understanding predictions, but it seems like the complaint that user’s are making is that their hurricane path ends up striking Miami more times than not, and very few make it up to Jacksonville.
This is a terrific point, so I wanted to emphasize it. I think we’re hung up on the wrong pardigm. FTP now for TR is almost an artifact of the past. The system is trying to get you the most productive workouts over time. As Nate said, FTP is now just like a Rosetta Stone for using the old workout library.
However TR spins it FTP is still an excellent marker of aerobic fitness, and that just happens to be extremely important to this sport 
It is still one of the better ways of tracking your fitness gains and seeing where you stand. Obviously it’s not the only thing one should pay attention to but completely ignoring it or trying to say it is meaningless is a huge mistake.
I am happy to say that I largely agree with you, and I gladly use TrainerRoad precisely because I find the AI does a pretty great job of giving me what seem to be excellent workouts.
My original post was entirely in the spirit of constructive criticism because I want TrainerRoad to succeed.
However, I find that the new app interface with AI FTP Prediction sends a message that strongly contradicts the spirit of your post. Previously, ones’ FTP was recalculated once a month, and so paying much attention to it the rest of the time was pretty pointless. The app sent an implicit message that constant fixation on FTP was not important. Now, the AI FTP Prediction is recalculated often, it is displayed centrally in a giant font size, and the fact it is always changing attracts the user’s attention. The implicit message is to look at the AI FTP Prediction all the time.
If they really don’t want people to fixate on the AI FTP Prediction, then maybe the default should be to hide it (or at least minimize it) in the interface?
I don’t know. I would assume they include the survey responses in the model because it leads to better predictions than not including the survey responses. Surely they investigated this in the development of their model?
I would also assume that - in general, all other things being equal - the impact of picking “Very Hard” when the model has predicted 55% Hard / 45% Very Hard would be less than when it has predicted 95% Hard / 5% Very Hard. But with out more info about the model, this is all just idle speculation.
Yes, it will make a bigger difference, but it will statistically happen 9x less often.
A huge number of users are reporting very optimistic predictions that ultimately come down to earth over the course of the block. What I have described above is the reason why.
I expect that the TR team will ultimately acknowledge this and fix it, and we won’t see any longer term predictions until that happens (the bias propagation continues to expand the farther out you predict).
Good analogy.
The recommendation here is nice as well. I think what would end up happening, the forum would change from “My predicted FTP is wrong” to “i can’t believe i’m not part of the 50% that was within x watts”
Apologies if you didn’t find my post to be constructive. I certainly meant it in a spirit of constructive criticism. I tried to support my criticism very specifically and clearly, and then provided a suggestion for how to improve things.
I’m a fan of TrainerRoad, I’ve found the AI FTP to be quite reasonable and in line with the feedback I was getting previously, and I want them to succeed. However, I do think that the roll out of the AI FTP has been less than ideal in terms of how it has been communicated to users.
I think the proof is in “Are you getting Faster or better when you cycle?” Regardless of whatever number is generated. For example I’ve done the same race 3 years in a row and have trained with Ramp FTP (1st year), AI FTP (2nd Year), and Adaptive Training with a Masters Plan (3rd year). My FTP calculated before each race was high the 1st year, lower the 2nd year, and lower the 3rd year. However in the actual event I cut over 2 hours on my finish time from year 1 to year 3 (with a lower FTP). At the end of the event having a high FTP isn’t what got me across the line faster, it was all the work building the engine. Hoping this year to cut even more time from the race, until I plateau at some point.
That is an interesting conceit, but the premise ignores:
rider weight
rider body composition
rider fatigue
rider nutrition
injuries
familiarity with the course
equipment
racing strategy ie going all out vs conserving fatigue and output
Race performance is so multivariate that using results as “proof” of concept vis a vis FTP is like saying a team won the superbowl because they had a balanced breakfast that morning.
The hope is that the FTP is accurate, and therefore predictive. In your case a declining FTP with inclining performance immediately tells me there are other variables at play. It also calls into question the usefulness of a FTP measurement in the first place if in fact you had declining FTP and increasing performance; that is not helpful to predict year 4, nor should would it be reasonable to expect continuously increasing performance for reduced FTP unless again other variables are at play.
Hey Steve.
Hard workouts should make you feel tested but not gassed or hard to recover from. So compare your perception to execution, and that you’re good to go again in a few days.
Mentioned on a few of the many other threads - I’d encourage performance to be your measure of success. Perhaps a benchmark like a 20TT or 40TT.
Thanks for responding, how do you think my point in the inherent non-deterministic characteristics of AI affect your analogy of hurricane uncertainty?
Good explanation, but still the system needs some tweets I think. Example. I’m on an FTP plan. In this I have all the sugested TR workouts and strenghts sessions. My ftp given after my last TR check date was 263. In the weeks following it dropped to 259 (following the TR plan). Nothing to worries about. But this week I’ve had to change 2 days because of other onligations (wedding day forgotten 
) changed the strenght from saturday to thursday and my sweetspot from Friday to Friday to saturday. So just a small change. But this changed my ftp prediction from 259 to 251. I don’t think this is correct, and think will ignore it, but is is enoying and also a bit demotivating (as I feel it).
Started then my ftp plan at 255. End up wit a lower ftp is as I think not the intention of this plan when you only swob one workout.
I’m sure this will be fixed in the future, it is to share with you and if you have suggestions, please let me know, always open for it..
You’re ignoring fatigue. The entire point of the survey responses is to gather feedback data on fatigue. TSS/ATL/CTL are one way to estimate fatigue, but it is only an approximation and RPE (survey responses) is always supposed to override what the numbers say, per Coggan and the way the underlying Bannister model works.
If the model does not utilize survey responses then it would end up either under or over training the athlete. So since survey responses are necessary to get the right amount of training stimulus, then they are also important to predicting future FTP.
Now, perhaps the model over-reacts to an unexpected survey response to a single workout. I would think that it should take multiple factors being off before the model reacts significantly. However, since it’s a black box we just don’t know why it did what it did. Maybe it wasn’t just the survey response but also something in the HR data. To the user it’s going to seem to only be the survey response, since we don’t really know how the model “judged” the HR data.
Survey responses for completed rides are important. I never said otherwise. It’s the automatic estimation of the most likely response (data that the athlete hasn’t provided) that’s causing error.
I’m not ignoring fatigue; in fact it’s the opposite. In the example above, the model is ignoring the possibility that the athlete might be fatigued enough to mark it very hard, which results in an overestimate in FTP prediction.
That isn’t my impression, if you mark a predicted HARD workout as VERY HARD… my assumption/experience-on-TR (and my knowledge as a ‘Coach’ ) is that it lowers your predicted FTP.
As in…. “oh this rider can’t handle this work load or this hard of a workout… better dial it back.” Then reassess as weeks go by. You argue just now that the Survey is important… but were arguing they are not needed and shouldn’t be used… confusing reading, maybe just how it comes across.
The reasons why the post ride survey is helpful and needed in the process is that AI can’t predict how hard you are working (or your work/life/training balance, nutrition, sleep, recovery), you could be totally burying yourself in workouts (and recovering like a pro) after a few cycles you are hitting workouts just fine but its on the hairy edge. That 4x10’ sweetspot could have buried you…. essentially a threshold workout, but made it through. Without a feedback control loop it will continue until failure when faced w/4x12’ (not good for TR or the athlete), the intent is to modulate the control loop with the input before failure has a chance. With the goal of a steady progression and consistency over months.
Preventing.
No sense waiting for the bowling ball to hit the bumpers and then skipping from rail to rail on the way to the pins…. guide it more to the middle once trouble is detected as it heads towards the bumper.
I don’t think this is true. If it were true, future workouts would never changed based on the RPE assigned to them (they do change).
I’m going to disagree with this one too. I would expect a very different FTP if I completed a 2x20 at 300W and rated it moderate vs if I rated it all out.
I’m not sure what you mean by “the inherent non-deterministic characteristics of AI”. All of the AI/ML systems that I am aware of are entirely deterministic. The programmers just choose to include a different PRNG seed as part of the input in order to produce different outputs. Given the same input, they will produce the identical output. This is even true for MCMC methods! (Quantum computing might be a different story, but I don’t know of anyone using it right now for production AI systems.)
I would argue that the inherent non-deterministic characteristics of the humans (and the hurricanes) are far more of a problem! That was really what I was getting at with my post.
Providing only a point estimate of the most likely value of a human performance measure out into the future without any indication of the uncertainty around that estimate is not a great way to communicate. I think it is fair to say that there is a strong consensus about this in the scientific community.
I may be an outlier here. In my experience the predicted FTP has only varied a watt or two over the course of the 28 day window when I complete the prescribed workouts without pauses or reductions in intensity. And when I’ve had more oomph at the end of an intervals workout I’ve turned up the last set by a few percent and seen the predicted FTP climb a watt or two. My “cone of uncertainty” has been fairly narrow.
I, too, wish I had a deeper understanding of the mechanics behind the prediction algorithm because I’m a nerd that way. But I can’t argue with the results I’ve been getting. Whether or not the AI FTP is “real” or correlated with other FTP determination methods, I’m holding power levels for extended intervals that I could not have held 3 months ago so that’s a win as far as I’m concerned. The truest answer will come from my “A” race at the end of June.