Uhhhhh… so for today’s workout, I switched out Batu (TSS 71, IF .84) with Red Kaweah (TSS 72, IF .85) and my FTP prediction went from from +3 to -7. I only made the change because Red Kaweah would be easier to remember and accomplish outdoors. No way that this small of a change could have such a significant impact. And note that yesterday my FTP was predicted to go to 284 and today, with no changes being made, it dropped to 281.
I’m taking a closer look at this now, but keep in mind that we aren’t looking at just one workout when we update your prediction; it’s all of the workouts between now and your next detection.
If you change one workout, often many others will also change.
It looks like those two workouts are similar enough, but for whatever reason, you’re twice as likely to fail Red Kaweah -2 than Batu, which likely affects the workouts following it.
I get that and see that when I flip-flop to the the slightly harder ride, it results in small changes that decrease my weekly TSS by 2 this week and 4 next week. But ironically, next week is a taper week because of my first A race is next Saturday.
For whatever its worth, I have been quite happy with the AI changes so far and really don’t care what my absolute FTP “is” other than its utility in setting my zones. That being said, I am an ML modeler (environmental sciences) and to me, the model for this FTP period should have mostly converged by now, three weeks in. I would expect the outcomes should only sensitive to a string of multiple major outliers. In this case though, the differences between those rides are so far within measurement error alone, that it has to be a glitch.
But they aren’t. The model thinks you have a relatively high probability of failing it. How do the power profiles of the two rides compare?
Not sure that 2.6% can be considered a relatively high probability compared to 1.3% in this situation.
But there’s also a higher chance that it will be hard (11% higher) or very hard (3.5%) compared to the previously assigned workout which can change upcoming workouts. All of that put together causes changes in estimated FTP. At least that’s the way I understand it…but who knows
It’s twice as high as the initial to go with a twice as high Very hard and about a third higher chance of hard.
In modeling terms, the difficulty probabilities are intermediate predictions built on all of the empirical data I have given the model through my recorded rides. There are recurrent or recursive ways of feeding those back into a machine learning model to predict the final dependent variable, which is FTP. However, if the final dependent variable is highly sensitive to to small changes in its intermediate predictions, while there was no change in the empirical data they were based on, it indicates those intermediate predictions are weighted too highly.
And further, it makes no biological sense! No way, in actuality, that a TSS difference of 1 in one ride wipes out two months of structured training. This is the first principal of modeling… do your outcomes make sense?
This is just an artifact of the predictions using discrete steps. They said when this was released that it would be ideal if they could use monte carlo simulations for predictions, but right now was just taking the most likely single path.
That means if the predicted most likely path changes as a result of some value just barely tipping over a crossing point, it will sometimes make a bigger than expected difference in the predicted outcomes.
All models are wrong, some are useful.
100% and I would say if they need to use discrete steps at this points, it is discretized too coarsely for such a small change to elicit not only a pretty big absolute change in modeled outcome but a actual change in sign. The whole reason I posted was to help them dial in their this model.
But you are predicting in 9 days what the best wattage is for you to train it. It’s not looking at months worth of data. It’s noisy, but the model thinks you have an increased chance of telling it either your current watts are just right or too high. My guess and reality based on many other posts is failures are highly correlated. It seems that when someone posts that they failed a ride they usually fail multiple, so that would also make picking a ride that has a higher likelihood of failure more impactful. Your biology might not change but your mental fortitude might.
Should a weather model treat all data in a month equal?
That wasn’t showing your probability but mine, however both of those workouts are relatively low AL, so a failure I imagine would be pretty detrimental in AI ftp. You may have found an edge case.
And in general, it is sensitive to a workout that has a higher likelihood failure rate, I swapped out a sweet spot workout for one with twice the failure rate and got a 4 point drop 9 days out. It’s also one of 2 hard workouts before detection, as next week it’s a recovery week.
And BTW, nice work with the George Box quote. University of Wisconsin (my alma mater) giant.
For me the issue is around having confidence that the workouts being served are not changing too much based on one small deviation from the previously expected path.
What I mean is, if we hear that the reason the predicted FTP has changed so much is because it looks to the future workouts to make that prediction, it follows that the new planned workouts must differ materially from the previously planned workouts.
I don’t want my training plan to be that adaptive. I would much prefer small and gradual adaptations that nudge in one direction or another. Particularly when the deviation from plan was something so subjective as rating a workout very hard instead of hard, which could be due in the end to many factors.
This is my concern, too. FTP predictions seem to wildly swing based on relatively small changes in workout.
I have stopped looking at the predicted FTP because it’s no longer a useful metric, but as has been said above, it is undermining my faith in TR being able to help me improve.
I would ignore the FTP because I really don’t care what it is compared to others. Buuuuut, as I understand it, TR FTP is what sets your zones and therefore your workout difficulty. I don’t want one small change in one day’s workout to mean that I’m going to get a month of workouts that are 5% easier. And ironically, this sort of over responsiveness of the model is precisely what the TR podcast regularly speaks against! It is supposed to be that fitness and performance are built with consistent structured work over time with progressive overload. Not “oh my god you had one workout with a TSS 1 point higher than recommended so now you’ve lost fitness.”
I’ve been having this issue too in the last couple weeks. I’m wondering if they are updating the FTP prediction model? I have a current FTP of 324 and my initial prediction was ~340, then it went down to around 325 after a while, but after my 1hr z2 ride yesterday I got a bump up to ~340. Compliance has been pretty reasonable (missed a couple weekends due to travel).
@eddie , can you take a look and see if anything strange has been happening over the last few days? (feel free to share anything to you find from my calendar)
Quite
I’m coming back from 6 months off due to a back injury, so any training is better than no training. However, I have a friend who is a high level masters athlete (World Champion duathlete) and she had a brief illness over Christmas/New Years and TR AI significantly reduced all her workouts and predicted FTP. She’s finding it incredibly frustrating and on the verge of giving up on TR.
Agreed that it seems like something in the model was changed or updated. So far I have been really pretty happy with the AI, but in the last week or so its gotten very twitchy.