Trying to get out of the sample size of 1 on workout failure / adjustments.
OK to share my calendar publicly for insights (To clarify this post has a specific focus on Failure so I’m not interested in starting a conversation about paddleboard TSS)
What I discovered recently is that if you are on a TR plan and looking ahead and you want to update a specific workout on the plan if you select anything HARDER it will flag it as “not recommended” and it will predict you will fail it (lowers AI FTP etc.) Based on conversations with TR support this is a feature not a bug.
So the question is; what drives failure, what is the acceptable % of a failure in a training plan, and why is this threshold not configurable?
From the Pod Jonathan mentions 2% as an average for failure across athletes using TR AI and I believe that this dimension is then implemented a hard cap on the AI Workouts. For Example:
Here is my last week and the workout today “Big Dog” is listed as Moderate
Now lets say I’m confused why TR would plan a week with 2 moderate days and 1 hard day in a block and choose to select a harder alternative because I feel good (Why it’s doing this and why I’m choosing an alternative is an AI horse that is being beaten to a pulp in other threads)
In general I find failure hits due to circumstances outside the scope of the data that TR is gathering eg. Heat, Sleep, Nutrition, Travel, too much paddleboarding etc.
Do other people run into the same 2% threshold? How do others feel about this being a feature?
Maybe I’m misreading this, and I haven’t looked at your conversation with support, but I do think it makes sense to have failure rates set for each zone.
For instance, with those sweet spot workouts you’re referencing, while you might not actually fail them if you gave them a shot, they likely wouldn’t bring the right stimulus and therefore wouldn’t be recommended.
Each workout zone has to have workouts that aren’t recommended, since the training in each zone has limits.
I haven’t looked into this yet, but I just wanted to see if that has any influence on your view on things.
The 2% threshold you’re talking about is obviously a different subject, though..
Today’s workout for me has a 2.9% failure, and was the AI selected workout. I don’t get a warning that it’s too likely to fail. I’m not sure what all is taken into account to determine failure percentages. Might be different for different workout types? I’ve got VO2 on/offs today.
I know if I have a long sweet spot workout that goes bad, things tend to feel just fine until they very suddenly don’t. Especially if I do sweetspot or threshold underfueled, it’ll increase risk of failure on what otherwise feels fine.
The 2% seems to apply to any “zone”, endurance, SS, Threshold for me.
I agree that there should be some graduated level where it flags as not recommended 2% feels on the low end of that scale and I feel that there is minimal benefit to the model predicting complete failure based on that threshold. Though it is trying to be an effective “stick” to beat back any attempts to select alternatives.
Right stimulus - I feel like that is another scope of question outside of failure; Why is it picking 2 moderate SS and then a 1h endurance ride on the weekend is another mystery to me.
Edit: to add some data from the alternate to Big Dog; I selected Antelope +1 which was a modest +0.4 SS from prior SS 2 weeks ago (same FTP). Post workout evaluation:
RPE: Hard HR: ~91% of LT2 (3% decoupling) Power: 100% Erg mode
AI FTP Returned to 329 (+3.1%) and it looks like it dropped my overall weekly TSS by 7 by adjusting upcoming workouts down.
In my mind its the same thing albeit with slightly clumsy labeling; just a product perhaps of a level of detail yet to come as TR develops the platform.
If the AI believes your fitness is not yet quite up to the level that the workout is training you at (even if you could technically complete it) I can see why it would swerve you from it.
For example, if it wanted to program 3 x 20min sweet spot for you but it estimated that due to your fitness that would put you closer to your threshold than would be optimal then it would say it was too high a chance of failure even though it’s still rated Moderate. It’s not saying high chance of failure because it thinks you can’t finish it but its a “failure” because you ended up doing a much different workout than planned based on where it believes your zones really are. 3 x20min near threshold is a different ballgame than sweet spot.
Ok but I might say “yeah but my FTP is “x” and these zones are in my sweet spot”. But my FTP doesn’t just JUMP every 28 days, just the training number does. My body is gradually changing throughout that period and the AI is attempting to track that based on performance metrics.
That is precisely my point. And that’s assuming you’re improving at all, you could be detrained or fatigued and your diagnosed FTP wouldn’t tell you that because of the 28 day cycle.
At the point indicated by the arrow I can easily imagine a scenario where a rider might see a Moderate Sweet Spot workout that expects “failure” because at those wattages the rider would not be in their sweet spot according to their current fitness although according to their FTP it would appear to be fine.
Again, not failure in the sense that they couldn’t turn the pedals over but failure in the sense that they spent the entire workout in the wrong physiological zones.
EDIT!!! Sorry, the black line should’ve been labed FTP DETECTION, not prediction. That’ll have caused a nice bit of confusion I bet! Apologies.
I first observed a strong correlation between failed workouts and massive (20%) decoupling and general death level RPE. So then I started changing variables around workouts to see if I could influence a change on it.
What I found was that not only did water help, but adding ice to the water was a significant performance enhancer on the trainer. (despite having massive fans).
So I use it as a performance criteria to evaluate for changing environmental variables in review. I did add it to my head unit, however I didn’t find significant value to it on the road changing power etc.
Example below of a race that went well, kept the lid on my max power and hoovered about 1L of water an hour.
At these 3 points your FAILURE dimension will change however TR has no idea you ate too many tacos nor does it know you had an amazing sleep.
These factors to me are the most likely causes of failure within a target range of a workout; comprising lets say a ±5% change in “fitness”
My experience has been that the Workout Levels are a very good objective measure of the relative difficulty of two workouts. I would expect to fail going from a 4.7 Threshold to a 9.0. However routinely I can select a +1-2 range when I am feeling good (A Days) and have found that to be OK. The AI however doesn’t appear to be looking at the same criteria and doesn’t recommend anything above the level it’s chosen.
As it stands it does not make sense to me how it calculates failure for a given athlete; even Nate mentioned on the pod that DC Rainmaker is basically EASY>MODERATE >>>FAIL “due to all the other things he does” which I agree with 100%.
The concern I have with the Failure ceiling is that it appears to be so avoidant of failure it will fill time with Z2 vs time in zone or intensity. For me workout MAXTime is fixed, however I would rather spend 1h training on Monday and 4h training Z2 on Saturday vs 1h30m on Monday and 1h30m (?!) on Saturday which appears to be what it’s favoring.
For long z2, I’ve been setting it as max length endurance, but I’ll usually tinker with things that are within a few days out. I’ll check what the ai thinks of different durations and different ride timings. Sometimes moving days around makes the prediction go up. I usually let ai pick the ride, I’m just checking how changing the constraints impacts predictions.
This block 3.5 hours has been the best option I found each week.
Apologies.
