AI FTP unable to account for beginner breakthroughs gains?

[symmetry]

Context: Looking to post this before I do a ramp test on Wednesday to see if TrainerRoad AI FTP is too conservative because the machine learning data set possibly does not include enough breakthrough FTP increases in new riders.

FTP Increase Chart

Me: Male 34 6’2" 210 lbs, never did endurance sports only weightlifting, lots of squats etc but definitely not an athlete.

Bought my first real/nice bike + Wahoo Kickr V5 early November (I was 240lbs then and most likely in the worst shape of my life cardio wise, except for lifting some weights) and went straight onto Zwift.

I’ve logged 1700km and 16000m elevation (2d19h) since then and got absolutely addicted to cycling.

First Zwift Ride: November 7th
First Zwift Ramp Test: November 26th 2023 (191 FTP)
Second Zwift Ramp Test: December 8th 2023 (210 FTP)

After the ramp test I started doing structured training with Zwift workouts and then learned about TrainerRoad and did my first TR Ramp Test on January 9th (220 FTP) and put myself on a low volume rolling road race plan (not sure if this is the best since I’m not gonna race but i’ll be road cycling mostly).

First period three weeks went nicely, I added longer Zwift zone 2 rides between TR days and added TrainNow workouts when I felt good as well.

First TR AI Detection: January 27th 2024 (229 FTP)

From here it feels like I unlocked some serious beginer gains suddenly, I realized all the workouts were suddenly too easy and kept increasing progression levels more and more by choosing alternates in stretch/breakthrough and even those were rarely hard and mostly moderate efforts, if not easy.

I then found about intervals.icu and signed up with my strava account, which led to estimated FTP (eFTP detection) from intervals.icu following Zwift Climb Portal activities

First eFTP increase: February 3rd (244 eFTP from 14m @ 264w)
Second eFTP increase: February 21st (268 eFTP from 12m @ 293w) [Zwift estimated FTP aka zFTP 263w]

That second climb especially had felt amazing, I’m on a calorie cut and had eaten a bit more for a few days to fuel up.

AI FTP detection was available last Saturday and I was hoping to see at least 260 which would have been obviously a massive 13.5% increase but TR AI gave me 244 (6.55% from 229 previous AI detection) which was disappointing.

It tried to change Saturday’s 90min workout (Norman Clyde +2 Threshold 7.0) to some 3.3 PL workout after resetting my PLs with the FTP increase. I declined that adaptation, did the 7.0 workout @ 244 FTP and aced it with only the last minute of the last interval set feeling a bit challenging.

Now I realize that FTP in combination with PL levels is supposed to be able to bridge the gap but I would still rather have accurate FTP detection if truly I’m making that kind of progress. I realize it can’t go on forever and that I’ll have to accept much smaller month to month gains at some point.

I’m now in a rest week and feeling good, I’ll do a ramp test either Wednesday or Friday and see if TR AI FTP Detection was correct or if the simple eFTP calculations from Intervals.icu are better. (edit: I realize that intervals requires a short intense effort which AI FTP does not, so there are still different use cases)

[kurt.braeckel]

Based on my experience with it and that of a few of my athletes who came over to me from TR, I would be inclined to trust AIFTPD rather than reach for an aspirational FTP based on 14 or 12-minute efforts.

But if you really want to know, go sit at 268 and see how long you can hold it. That’ll answer this question and probably force AIFTPD to update either way!

[symmetry]

Haha yep doing all this for science

I’m definitely super curious so I’ll do the ramp test to compare but I wanted to post this before I did the test so I can either be humbled by our AI overlords or suggest an improvement to AI FTP Detection based on some factual data.

If I were to actually go all out and do 268w for an hour, would TrainerRoad actually prompt an FTP update immediately? Or are you just saying that it would take it into account for the next detection. Either way thanks for the reply

[MI-XC]

There is little to lose by having a FTP that you think is a bit too low. Just answer the workout surveys honestly and you’ll progress quickly through PLs and workouts by their nature will be more difficult. However, having a FTP a bit too high will make your workouts unnecessarily more difficult and will ultimately hamper your recovery and gains.

[kurt.braeckel]

It would feed their model. I don’t know how the black box of AIFTPD works, but I have a pretty good idea (It’s not a new thing to be able to predict FTP based on performance data… intervals, WKO, Xert, TrainingPeaks, etc. etc. etc. have been doing it for a long time).

But seriously, in lieu of a ramp test (which often overinflates your FTP), why not validate Intervals’ number. If you can go sit at 268 for 35 minutes or so, it’s a good number (FTP =/= Hour power!!). You can also manually enter FTP information, and nothing is more accurate than going and riding at FTP for a long time when you’re trying to figure out FTP.

Worst case is you can “only” hold 268 for 20 minutes, and then you know, AND you got a really good work out in AND gave TR a truly maximal effort to work with.

100% agree with MI-XC too… don’t seek the highest number. Seek the REAL number, and train there. Better to be a little too low than a little too high.

[mtbtomo]

I think if you do a work out where the PL is over 9(??? Or is it 8???) for sweet spot or threshold, then doesn’t AIFTP auto detect an(other) increase?? (Could be wrong)

Regardless, it can only go off the data you plug in by doing sessions and recovering.

If it doesn’t pick the noob gains then do a ramp test as you are doing, or you could programme your own work out and set something like 2x20mins or 2x30mins at whatever you perceive your FTP should be. You’d still be waiting 28days for autodetect though. If you manage a 2x30, I’d be tempted to whack in a manual FTP adjustment.

[mtbtomo]

In agreement with @kurt.braeckel

[symmetry]

The only loss here is an ego loss

I’m sad that I didn’t discover cycling when I was younger, I had friends who won world championships and others who did the TDF but I never understood how fun it is. I also saw it as an expensive sport and since I grew up poor I didn’t even consider it.

So now seeing my FTP number go up so quickly is incredibly satisfying and sort of feels like making up for lost time in a way as well and seeing my potential. However I want to do this for the rest of my life, I realize it’s not a sprint.

[MI-XC]

@mtbtomo

I think if you do a work out where the PL is over 9(??? Or is it 8???) for sweet spot or threshold, then doesn’t AIFTP auto detect an(other) increase?? (Could be wrong)

Yes. I completed a 2 hour 9.3 Sweet Spot workout this past Saturday and was immediately prompted by an AIFTP detection request, even though I was within the 28 day limit. I denied it and continued on with my normal programing.

[MI-XC]

@symmetry

I’m sad that I didn’t discover cycling when I was younger,

I started at age 40 having not owned a bike since I was a kid. I’ve seen massive gains with TR.

[symmetry]

Interesting, I did have FTP === Hour Power as an all out effort in mind. Really appreciate the added information.

It feels like the “real” number is higher just because of how a 7.0 threshold workout felt at the new FTP but I’ll validate. I’ll start with the ramp test for consistency and comparability and then later in the week I’ll try manually setting power at 268 and seeing how long I can hold that.

[WindWarrior]

It was posted recently by TR employee, that AI FTP assumes you have good FTP estimates. So if the TR ramp test was a good ftp estimate, then AI FTP is likely giving good estimates and its Intervals eFTP you would want to question. You can tweak the Intervals model - there are 3 or 4 models, and FWIW my experience is that Intervals can at times overestimate if you have some strong shorter efforts.

[symmetry]

Confirming what @mtbtomo was thinking

My next scheduled sweet spot is Tallac -3 (4.1 PL) on March 7th, with alternates I can bump it up to Redondo +3 (9.2 PL) which seems very doable still. Although by that date I’ll have done another FTP test so I’ll most likely be fine with current PLs if really my FTP ends up being higher.

[symmetry]

Indeed I saw a few replies on Reddit threads where people said that eFTP could overshoot and that at the minimum you should set eFTP detection intervals to 600 seconds (10 minutes) instead of the default 300 (5 minutes). In my case both effort are over 10 minutes but it’s still definitely possible that it’s overestimating, happy to add a data point either way.

[kurt.braeckel]

@symmetry

I did have FTP === Hour Power as an all out effort in mind. Really appreciate the added information.

From the guy who defined FTP, Andrew Coggan, Ph.D, “FTP is the highest power that a rider can sustain in a quasi-steady state without fatiguing for approximately one hour.”

He’s gone on other places to define this as ~40-70 min. FWIW, most athletes that I’ve taken straight from a TR plan have a TTE (time to exhaustion) at FTP of 35 minutes or so, and that can be really common.

Note: none of this means that your 35 min power is your FTP, nor that FTP is your 70 min power. The BEST way to determine FTP (IMO) is with an accurate power profile including long, sustained “max” efforts, and then identify that bend point on the log scale graph where fatigue becomes pretty obvious. Basically any FTP test you do other than actually “go ride and find FTP by feel and hold it” is providing an “estimate”, and some of those tests are better than others.

Sorry to go full nerd.

[Gnome]

change the minimum ftp detection inintervals to 20-25min. that will change your graph

[symmetry]

Sorry to go full nerd.

Full nerd is the only way to nerd properly. I’m trying not to make the mistake of thinking I know best based on reading things rather than experiencing them first hand. My best friend is a former olympic level athlete (endurance sport as well) and he constantly reminds me how much he does by feel and that I should avoid being rigid in the way I do things.

[ZackeryWeimer]

Hey there and welcome to the TR community!

Glad to hear it sounds like you’re enjoying cycling! Looks like you’re seeing some awesome gains from the start here.

It looks like your recent AI FTP Detections reflect that hard work you’re putting in. Going from your first TR Ramp Test of 220 watts to your most recent AI FTP Detection of 244 watts is a serious jump – especially given the timeframe!

We like the advice @MI-XC gave you here already – Adaptive Training will get you dialed in as you work through your plan via your Progression Levels. AI FTP Detection will continue to receive all of the data you upload from your training to adjust your FTP with future detections.

@kurt.braeckel also gave you some great info on FTP itself. If you continue to question whether your FTP is accurate, you can always test it out in the real world on a training ride. Since you’re new to cycling and it doesn’t sound like you’re one to shy away from a test effort, I’d even say there can be some good value in doing so to get an idea of how such an effort feels so it can inform your future training. As Kurt said, you’ll figure out how hard you can go over a longer duration, and even if you go too hard and blow up, that’s still useful data to have and learn from!

[symmetry]

Was impatient and did the test today instead of resting one more day.

I think with one more rest day and a bit more mental fortitude than today I could have matched the 263 estimate from Zwift, maybe not the 268 from intervals.icu

That’s still a 6.6% gap between AI FTP detection and a ramp test.

Again I imagine the dataset the model is trained on doesn’t have a lot of noob breakthrough data points so I presume for more seasoned riders the margin of error will be smaller.

If workouts feel easy at least the progression levels can be adjusted which helped for last month since it was too conservative for my first detection as well.

[huges84]

The fact is that there’s a lot of variation from person to person on what their power profile is. And every (good) model requires enough max efforts at a few very different power levels to get a decent estimate of your profile.

Most likely you just haven’t given AIFTP a good enough set of data to work with. But there’s no way to know because TR doesn’t disclose what data it needs besides 10 TR rides.

As a counter example, I’m also in the noob gains phase (I’m not a true noob but returning to training from a several month break and little riding the last two years).

I’ve used AIFTP to detect the following:
1/28: AIFTP 205 to 218 (+6%), intervals.icu said 216
2/27: AIFTP 218 to 235 (+8%), intervals.icu said 242, my estimate from the Kolie Moore FTP test was about 235-240.

So AIFTP did a great job for me. I am using 235W as my FTP. However, I have fed the models good data for max efforts of various time frames (5s, 30s, 5m, 20m). This really is the key factor.

I personally don’t trust the ramp test. It works well for some pale some times, but it doesn’t work often enough that it’s best to avoid once you’ve got decent data to work from.