Adaptive Training: Post Your Custom TR Rides Which have Mis-Scored Progression Levels

This is a call for everyone who has created a workout using Workout Creator (this is NOT for outside workouts) that has mis-scored progression levels to post the workout here, and why you think this is mis-scored.

Hypothesis (@IvyAudrain jump in if I’m off base): TR is using ML to score custom / Workout Creator workouts. So if these are mis-scored, then correctly scoring outside workouts won’t happen. So if we can give TR a sufficiently large pool of mis-scored workout creator workouts, hopefully this pool can be used to better train the AT scoring model, and help get us more quickly to having correctly scored outside / free ride workouts.

Examples of a mis-scored workout:

I submitted the above two to TR support, and the response I got back was:

The reason for the discrepancy is that workouts that are built by TrainerRoad don’t have the warm-up or cooldowns included when determining the levels. This means that we are only using the main working part when we calculate the level. However, these points are set manually so it’s not something that we’re also able to do for custom workouts. I’m not exactly sure if the entirety of a custom workout is included for calculation, but I can check on this and get back to you.

Since the warm-up that you add to your workout has a fair amount of length and gets up to 110% FTP, this can cause the levels to increase a fair amount from the TR version as there is now effectively an entirely new “extra” set of work in the beginning.

Not sure how it helps with outside workouts, but perhaps TR could train up an ML algo to identify warmup and cool-down sequences. It is good to know that the workout score excludes those, and certainly there can be pathological cases where the warmup is secretly a hard vo2 effort!

A simpler approach would be if workout creator let you annotate the WU and CD portions of your custom workout (eg a hidden text event with #WU to mark the end of the warmup and #CD to mark start of the cooldown; while on the topic and #erg or #slope command would be nice for sprints).

1 Like

My hypothesis is that TR is using the same AT functionality to rate custom TR workouts as it will be using to rate outdoor workouts. So if this functionality cannot correctly rate simpler custom TR workouts, then forget it accurately rating noisier outdoor workouts.

This was rated as a Sweet Spot 5, even though its almost entirely tempo


I would have rated this as a Tempo 5 or 6

That plus the rating for my customized version of Clarence King -2 leads me to believe that the way that custom workouts is now rated as a weird bug where the zone is selected off of the “highest” power value, and it isn’t doing a detailed zone analysis. So this can lead to completely wrong assessments - see below for an egregious example.

My concern is: if TR automated rating system cannot accurately rate custom workouts which should be relatively simple, how is it going to accurately rate more complex outdoor workouts?

My customized version (just adds warm up & cool down) of Clarence King -2 is rated as VO2 max 8.0 vs. standard version which is rated as Threshold 5.3

Judgement call on these… Some FasCat criss-cross intervals rated as VO2max instead of SS/tempo, that I might expect based on searching for “criss” or over-unders in TR catalog. Specifically the Criss-Cross, Tempo > VO2: 3 x 10 on that page.

Looking at completed criss-cross workouts in WKO, my relative VO2max had enough time above 85% / 90% to qualify as vo2 stimulus. Very similar time in high % vo2 as TR’s Baird variants that I’ve completed.

So no biggie on those being vo2max from my perspective.

Similarly:

Somebody posted about modifying Needham and I made that one in WorkoutCreator.

4x15-sec sprint intervals. The ones with 15-sec on/off are anaerobic, the ones with longer off (say 15 sec sprint, 90-sec recovery) are sprint. Those came from the FasCat weight lifting for cyclists off-season plan.

My TR library has 177 custom workouts, I just did a quick scan and those are the ones that stood out as “is it or isn’t it?” Might have missed a few. None of these are cause for concern.

That was probably my modified needham from ages ago

1 Like

Your example reinforces my underlying point: the current algorithm for scoring custom workouts doesn’t align to the way that TR scored the out of the box workouts. So, without any evidence, I think TR missed a big step: they should have turned their scoring algorithm loose on the TR catalog, and see how close it comes to the current score. Until the agreement is pretty good, then either the manual scoring is biased, or the scoring algorithm isn’t returning values that match the human expectation.

All, we appreciate the good intentions here, but you guys don’t need to go out of your way to screenshot and collect these. You’re always more than welcome to shoot specific examples over to support@trainerroad.com if you feel so inclined, but Custom Workout misclassification is a known issue and we have a good handle on what’s happening behind the scenes. To address one other specific concern:

The system that’s currently misclassifying some custom workouts is not the much more dynamic and intelligent system we have in development for unstructured outdoor rides, so you don’t need to worry about that. We’re on top of both use cases and neither are in their finished state.

4 Likes

Great, thanks for the update! Glad to hear its a known issue and hadn’t seen it on the list.

Sean,

Thanks for the additional detail. I started this thread because I did report the mis-scoring to TR support, and the response I got came across basically as “oh well, the scoring differs”. Not aligned to your response above: we know the custom workout scoring algorithm needs work, and we are working on it. If that had been the response, then I wouldn’t have started this thread as a way to get what I see as a problem addressed.

Cheers

1 Like

No problem, and my apologies for the underwhelming response you received from support. Our support team is absolutely world class but with so many things in flux, sometimes things inevitably get overlooked.

1 Like

Today I did custom vo2max wo.

It included a some short vo2max with short rest.
It indicated the rating was 8.4 and a not recomended label. (my old AT level was 3.2)

I still did it, and was not easy, but definitely not impossible as AT would suggest (my new AT level is 8.4)

So i have 2 hypothesis.
The AT calculations can’t handle short burst of power (making them outliers) or my vo2max adaptation was much higher than AT gave me credit for (unlikely).

custom workout levels are known to be incorrect, so you might try retesting your hypothesis with a similar standard vo2max wo.

vo2max workout levels are pretty high relative to others, so it’s not unusual for you to see higher numbers than in other systems. Bluebell, which is like literally the first VO2Max workout anyone does is a 4.5, so your 3.2 was probably either due to decay or that you hadn’t done any Vo2Max since you changed your FTP.

2 Likes
  • Correct.

Unless there is something actually broken with VO2, this post likely should be merged to the thread above.

2 Likes

I knew I read something similar before…
Yeah… merge away

1 Like

Almost the same Wo. The only difference is the vO2 max at the end

One vO2 max is 135% a d the other is 200%. Just 30s each

The calculation and wildly different tho