Perfect. I’m going to steal that and add it as my email signature
2 Likes
Totally agree. My argument is that PLs are better understood as the level of workout one can complete within a training plan. As such, PLs should not consider hero rides in order to maintain consistent training. A coach understands this context. A computer needs the right input to provide the right output. If PLs did not ignore hero rides, it would be much more difficult to choose the correct endurance workout.
1 Like
Also, I meant ‘productive’ in the sense of what a productive workout is defined as in TR. E.g., if you are at PL x, a productive workout is x + y where y > 0. Whatever the PL is for 24 hours in old pueblo, a ‘productive’ workout under the TR definition would be more difficult. Lastly, I hope it was clear that I was joking 
1 Like
I listened to the podcast segment. Nate said basically 2 things explicitly:
- Don’t want PL’s to get abnormally affected
- Don’t want recommendations to be abnormally affected.
This is reasonable to me. My speculations were based on the language of other poster and his interpretation of the video.
All good and nothing to see here.
4 Likes
clearly both approaches are wrong. I mean why should I limit my ability to complain? :-p
(Joking, would rather have more info then be in the dark)
Am I the only one who thinks there are limits to modeling with just power data which limits how accurate the predictions are? That lack of accuracy is why things aren’t going as well as expected. (i.e. training machine learning on data that doesn’t have enough info in it)
1 Like
They are including at least one more data point, our survey responses about how the workout or ride felt. So they are probably trying to work on associations between the power data (which includes not only what we did, but a target in the case of a planned workout) along with that subjective “feel” information.
They have mentioned that heart rate data is also a possible consideration (although not in place at the moment AFAIK). Additional comments on things like HRV show potential growth for the future. Plenty of room to grow as the current state above is not their end goal, but it’s what we have right now.
Broadly speaking, it’s better than no interpretation & adaptation that was the case prior to AT, but TR and us all hope for better down the line.
2 Likes
Not in place yet to use that data and not something they recommend. (Suggesting you get hr data) As in the data set of people who ride with HR data could potentially be bigger if they were told it was useful and as they need a large data set for good training… Especially with HRV where people could easily enable turning on HRV capture on their Garmins so they could have the data to look at where as now very few people probably do that.
HRV stuff was from last year:
(Just in case anyone is wondering, I’m not claiming alpha 1 gives LT thresholds but can be used as a way to tell how hard to are pushing relative to other moments during a ride so closer to a level of effort that isn’t accurate or useful enough on its own but can help fill in more details about a ride)
I’m not sure what you meant by this, but they do indeed recommend you wear an hrm and have said that it will likely be used in the future. Apologies if I misunderstood your comment.
1 Like
I havent listened to the podcast the last few weeks but noticed on my calendar they changed a workout from Endurance to Tempo
This workout used to be Endurance PL 4.8.
Yet this workout with a higher IF stays as Endurance
I know they are always tweaking but havent seen this mentioned
My responses to how a workout felt really in my view they dont add much. Correlation to workout for some I am sure helps.
My heart rate data is a lot more relevant when you compare various workouits over time. I prefer data vs subjective but I do understand my N=1 may be different then others.
1 Like
Heart rate and power have always been relevant, for everyone.
3 Likes
They do use other information, at the very least the survey responses do matter. However, what other factors are included are a closely guarded secret. But they did say on several occasions that including more data in some instances made the predictions worse. E. g. heart rate data is super hard to parse across athletes as that is influenced by a ton of factors.
I’d like HRV to be useful, but I think we should be quite skeptical. At worst it complicates things for the athlete. Currently, I base my decisions off of (1) power, (2) RPE and (3) heart rate. When 2 out of 3 are in the red, that’s a sure sign I should take it easy. When 2 out of 3 are green, it usually is a sign I can continue. HRV could be an additional data point, but the question is how much additional info we gain …
If I had to guess, I think sleep data would likely be more relevant.
I think HR is one of the variables that influences…something right now. As I recall, TSS estimation from HR data was rolled out with AT.
1 Like
Maybe clearer now. This is what happens when I post at the end of a day off work with a headache…
This is why adding alpha 1 from hrv data can be useful as a more sensitive way of telling how hard you are pushing during the activity. If you only use power data there is no good way to know that. Sure there is the survey response but it’s not like you’re pushing at the same level the whole ride
2 Likes
That’s all they say they use hr for
So if you have a power meter no need to use hr. Except if you have extra data that can be used for analysis it seems like they should
Rpe isn’t something your bike computer knows anything about it what tr knows about. One answer at the end of an activity isn’t the same as continuously measuring during. The idea on alpha 1 is that it’s more sensitive from just heart rate
2 Likes
That’s right, but you know about RPE during a workout and inform TR of it with the post-workout survey.
My worry about HRV is that it doesn’t give more information than I already know. E. g. one night I sleep too little and not very well, and my HRV confirms that I am not well-rested. Does HRV give me additional information then?
PS I may sound like I am poo-pooing HRV. Quite the contrary, I wish it’d work. But the scientist in me tells me to remain skeptical
What we do know is that TR has tried to train models with more and with less input data. They said they found that including more data does not necessarily lead to better predictions, it is that you need to identify the right variables.
I’m more worried about data TR doesn’t (or doesn’t seem to) capture, e. g. L/R balance, sleep data, weight, etc.
2 Likes
I think you’re over simplifying hrv. As in your thinking the hrv sleep trackers measure which is more a specific calculation of hrv. What I’m referring to is hrv more generically (the raw beat by beat timing) which calculations can be done on. The main calculation that seems useful for a workout interpretation is alpha 1.
1 Like
Maybe. What I am thinking of are HRV scores you get from e. g. a Garmin smartwatch. These are averages and the result of processing raw HRV data. It is the data exposed to athletes that I am thinking of. I would be even more skeptical if TR had to process raw HRV data. TR cannot do it all. At least not well.
I think a lower hanging fruit is TR importing sleep dat, weight, etc. from Apple Health, Garmin, etc. I reckon sleep data to be much more significant than HRV.
not necessarily true, re: explainable NN. In image datasets, it’s been shown that certain layers pick up edges or changes in gradient of the photo.
My guess, is the underlying model is stack of boosted trees. XGBoost, LightGBM, HistGradBoosting, all have been shown to outperform NN on tabular data.
Any updates or rumours that can be shared?