I’m guessing it takes a little while to learn (more) about how you respond. I suppose it has all of the back history…..in my case back to 2013 and month 1 went down after not following the plan, looking for a big increase in a couple of weeks here in month 2 and actually hitting my workouts (no less….and maybe more importantly, no more). Just spitballing here!
Joe
I was thinking the same thing, call it a longer calibrate phase.
After one more workout… from 237 with 5 workouts to go to 225 with 4 to go 
. Workout was hard but borderline moderate so who knows…
Your calendar is private. Not much we can say or think without seeing your calendar.
That’s the expectation, but clearly is not what’s happening.
Last block may be a bit wonky because of the Zwift Races. They’re hard to characterize, and I’m sure didn’t contribute as much to the training stimulus as a proper workout on those days would have.
This block, however, I’ve done every workout except one (I’m usually very consistent) and rated them within the expected range. Only one was very hard, and that was a medium percentage prediction. So, missing one I can see taking some off the prediction, but not as much as it did.
It also gave me a stupid easy “threshold” workout as my last hard one in the block. A threshold 0.9, when my level is 3.3, and the threshold part of the interval is solid tempo, not even sweet spot. And the prediction was a high percentage moderate, with equal parts easy and hard on either side. I’m not doing something that easy as 50% of my work for the week (I have 2 “work” days/week), especially on the last one of the block. So I substituted a 3.0 in the same hard starts section. I checked the FTP prediction for funsies, and it took another 2W, and now has me -1W for the block.
This feature is clearly not ready for prime time. I’ll leave it off until it comes out of beta.
How does the output compare to the target levels? Are you consistently below the target power?
I’ve only found the prediction has gone down a little if I’m under the intended target on one workout. If I’m then on target the next time, it doesn’t drop.
This is an interesting one. In my last 30 day window, it predicted me 5% increase, and I did all the work, but I also did more on some days, and had more yellow and red days because of it. Felt really good, but detection ended up with my FTP going down by a watt.
I think an unintended consequence of lowering your training approach (I was on ‘balanced’ vs. ‘demanding’) is that it can make your workouts easier when it sees fatigue and pop more yellow and red days. So, while following a balanced plan pretty closely can be as effective as following a demanding plan, but you won’t have same leeway to err on the side of more training load.
My theory of what happened for me was that in spite of me nailing workouts and rating them generally easier than expected, I also did more training than expected, and maybe the total TSS, or my ramp rates, or the number of yellow/red days being higher than expected, told the AI that I wasn’t absorbing training as expected.
Love to have @Caro.Gomez-Villafane or other TR folks comment.
Go look at the very first workout of that block, Lion Rock +3. It looks like you didn’t get full credit (1.0 vs 1.3). And then you missed your next workout because you were sick.
So, you’re already two steps off the path. And, of course, the earlier you go off course, the further you’re likely to be from your destination at the end.
Couple of watts above
Did you move around a workout? I’ve noticed sometimes right after I move or change a workout type it gives me a 0.9 at first. Then corrects it.
I’m totally mystified as to why it gave me so little credit for that workout (I started a different thread on that one by itself). I felt I nailed the workout, rated it hard but it didn’t give me full credit. I was in erg mode, and the averages for the threshold intervals were a bit low, but not particularly bad. No pauses, or other inconsistencies in the workout.
I rarely miss targets, and not at all this block. I run in erg mode for everything below VO2Max, or longer than a minute or so, and so I’m on target within the erg error which is usually a couple of watts +/-.
There may be something to this for me as well. Historically I’ve had 3 hard days/week, but TR has been insistent that older riders should recover more, and since I’m in that category (57), I figured I’d give it a whirl. I’m “that close” to going back to 3 hard days. 2 hard days just isn’t satisfying either mentally or, apparently, physically. I did game out the more demanding scenarios in plan builder, and it claimed a slightly lower FTP result, but at this rate I don’t see how.
Yesterday, I substituted the 0.9 threshold with a 3.2 Threshold. For reasons unrelated to the workout, i had to stop after the first interval. I got one interval in and it gave me a 1.0 threshold level (?). I got things sorted, but figured it wasn’t worth resuming the workout, so did an endurance workout that matched what was scheduled for today, rated it easy. Now I’m plugging in another hard starts threshold (Whiteleaf +1, PL 2.8). I checked the FTP prediction, and now it has me at a net LOSS of 4W for the block! This makes no sense whatsoever.
So my plea to Trainer Road is to explain the rationale behind the FTP predictions! Tell us why the number changes, in either direction. Numbers without context are noise, not information. We’ve been trained for a very long time to use FTP as a measure (proxy perhaps) for our fitness, and that isn’t going to change just because the new algorithm doesn’t use it when generating our plans. We’re still being given this number so we can track our progress over time and it’s so noisy it’s no longer useful. I understand that this isn’t universal, and other users are having no issues, but there are enough complaints that its obvious this feature needs significant work before exiting beta.
Unfortunately, there are no other metrics we can use in its place right now. PLs change every block, and power records are just that -records, not a measure of overall fitness. Absent a reliable FTP we have to just trust the AI. Without the ability to verify, that trust is … difficult.
Update: I completed Whiteleaf+1, rated it hard (56.7% predicted response) and got full credit for all zones. Now my FTP prediction is 305 (+1 for the block). Why? Shouldn’t the pre-ride prediction have assumed I completed it with the most expected response (hard), and hit the target zones, giving me the same prediction?
I’ve noticed my predicted FTP drops when I do a harder workout than recommended. Not by much. 1-2 watts.
Also, it seems to take HR and RPE into account. I did an over-under workout yesterday that felt harder than it should have (and my HR response showed) given poor sleep the night before. I rated the workout as very hard and my predicted FTP dropped by 4W. I went back and changed by RPE rating to Hard, and my predicted FTP went back up by the 4 watts.
These of course are examples of me trying to reverse engineer the AI FTP logic based on observations.
I think TR should augment their AI FTP detection engine with an AI chat agent that you can ask questions to about what it’s making the decisions it does. That would be a great feature IMO.
To clarify, I’m not complaining about the predicted AI changes. I have a pretty good idea of what I’m capable of doing and generally TR gets the workouts close. I like seeing how the AI engine changes my calendar and predicted FTP in response to the workouts I do and how I rate them. I think there’s a lot of insight to be gleaned from that - I only wish TR would create the aforementioned AI chat agent to help me interpret vs me guessing for myself.
yeah, that’s interesting. I’ve seen the same, where I expected prediction to stay the same, maybe even go up, and it dropped. I wondered if it was either HR response, or plateauing at a higher HR for the effort than expected.
Like you, if that were the case, I’m not complaining. That would be really cool.
Hey @dsacasa 
Sorry for the late reply here! I was off for the past week.
I took a look at your account, and I don’t see something in particular standing out. As Nate has mentioned in the past, the closer you are to your prediction, the smaller the change will be. When the prediction is farther out (6 workouts to go which in your case were 2 weeks and a half till prediction), there’s more of a butterfly effect, and small changes now compound over time. I wouldn’t fret about it either way; it’s more informative and meeting you where you’re at rather than judging you as bad/good.
You went from a 219 February 18, 2026 to a 226 FTP on March 18, 2026 which is an increase overall. Nice! While this is important to look at, I also see that you’ve rated almost every interval workout as “Hard” in between detections, which is exactly how we want you rating workouts as this means we’re serving you the right intensities to progress your fitness!
I see this FTP prediction issue as a HUGE marketing issue. If I think about the early days of TR the sales slogan was “making you a faster cyclist.” Moreover, they convinced us, just like pavlovian dogs, that the way to measure whether or not you were getting faster was with a rising FTP. TR conditioned its user base to think this way. Now folks are training their guts out but are being told they are not getting faster, even though compliant because their FTP has not moved to the predicted result. The marketing has sucked because the product / customer messaging has become confusing and inconsistent with the initial brand promise…”making you a faster cyclist.” TR got way to focused on the tech and lost sight of why people were attracted to the platform in the first place.
I posted some of this in an other thread but will summarize here.
Like others, I followed a build plan to the letter. Nothing skipped, nothing extended, all targets hit. Workouts rated honestly (VO2 “hard”, endurance “easy”) always falling in the peak of the predicted rating of that workout. Variables: I did many workouts outside (hitting right in the middle of targets) and measured HR (so if that factors into the calculation I can’t control it). After 4 weeks of build the predicted FTP went from +9W to -1W.
On the last day of the block I did a (pre planned) B road race and had the best result in this race I’ve had in 10 years. Felt great. No power PRs, however.
I then did a 20’ FTP test outside on a climb (so not a realistic prediction of what I can do for 20’ on a trainer) and got a FTP (=20’ power x.95) that was 15W higher than TRs predicted FTP.
My conclusion is that the training is working with respect to my events. I honestly don’t know why predicted FTP starts high and then drops despite following and rating a plan according to what I “expected” TR wants. I just will ignore the number. I do share everyone else’s frustration with the program which boils down to “I’m doing what you ask and it’s not working. WTF am I doing wrong?”
I do think that if you just want to see the number go up, just rate everything “easy” and maybe don’t measure HR .
This past saturday - FTP detection 273w
Next prediction - 288w - this feels unreasonably high given I’m starting specialty phase for XCO races but whatever, I’ll just roll with whatever happens
Sunday - unplanned/unstructured outdoor ride - prediction goes up to 289w
Monday - completed a VO2 workout
Tuesday - 60’ easy spin (lazy mountain -1)
Wednesday - complete an anaerobic workout, prediction goes down to 280w ¯\_(ツ)_/¯
Fully expecting this to continue trending down because that’s what has happened the past 2 cycles. I also don’t really expect to gain FTP during XCO specialty phase so this recent 273-289 prediction felt ambitious. At this point I should just disable it, in its current state it doesn’t feel very useful.
Edited to add: just added a survey response for an easy outdoor ride last week (.41 IF, 14 TSS), and my next prediction just jumped from 280 to 292.. what
This will be my second training block under the new AI. Block one had one failed Vo2- first Vo2 of the year. I did a vo2 the next day more in line with my levels. Predicted ftp dropped 6 watts ultimately ended up being-1 watt 269 to 268. New 4 week block AI Predicted ftp = 267. 8 weeks of training for -2 watts. What am I paying for? Seems like a waste of time and hard earned money. I can get slower on my own. Lastly we pay how much a year for them to harvest our data to work out bugs on a new model that seems incredibly flawed? Frustrating .
You can turn it off and not have to worry about all the potential ups and downs.
Which isnt good ofcourse. You not following the plan 100% does not equal you getting slower.