Let me know if you run into this again. I’d like to take a closer look before you fix things on your own if there’s time. 
I’m just digging into this one, but one thing I noticed is that the majority of your workouts are scheduled to be outside.
We currently have an open bug report where FTP Predictions are inflated when athletes are doing Outside Workouts. In those cases, it seems like their prediction is always dropping down as they progress toward their FTP detection, but in reality, the prediction should have been closer to that lower number from the beginning.
This doesn’t explain everything you’ve mentioned, but could certainly be playing a part.
I’m also seeing that you may be using Workout Alternates quite often which does impact your training. It’s hard to say exactly how that impacts your future workouts and fitness, but we pretty much always recommend that our athletes follow the plan for best results.
Let me know your thoughts on this and I’ll keep digging to see what else I can find.
Thank you for looking into it.
The first block I did when AI FTP Prediction was first released was almost exclusively inside. And what I saw was FTP moving from 305 to 303 in 3 weeks of build. The prediction was initially showing prediction of 315. So outside does not really explain it.
I was looking into my TSS and was checking how I felt last 2 months and I think I am under recovering a bit. I have recently changed my plan to 2 intensity sessions a week (from 3) and will keep rest endurance or easier. Lets see if that will help with moving the numbers up.
I have a comparison how trainer road worked for me before AI FTP Prediction and it is night and day. I hear lot of praise from Jonatan on the podcast about it but it is just not my experience.
My highest FTP numbers were almost never from AI FTP detection. I can perform well in ramp test and when fresh I can score 5-10 watts more in the test usually. But that I dont really mind. The thing with AI FTP Prediction is that it made FTP front and center of everyones focus. You check what balanced vs aggressive approach would do, you swap workouts to see what would move the number. But then you build plan based on it and start executing and the number goes down. That feels like a failure. And one where you cannot learn anything because you trust the system and algorithm behind it all and there is no explanation. This frustrates me and especially as a software developer.
That sounds very much like the issue I’m having.
Do you have a public bug list for this? There was one during the beta stage. It’d be great to check there and know it’s a bug rather than thinking I’m doing something wrong.
And related, are there release notes so I can check when it’s been fixed?
I have almost the exact same experience as you. I train a fair amount outdoors. Sometimes I choose a workout alternate because the suggested trainer road workout is simply not feasible to complete outside (especially some of the more changeable sweet spot workouts). My FTP prediction keeps decreasing, and I am training 11-12 hours a week with decent 9not perfect) plan compliance only to get FTP decreases each month. As you say, pretty demoralising and definitely demotivating 
Edit to Add: the FTP predictions are always for a higher FTP, often 3%+
My ftp prediction said 230 the day before my ftp updated, and then it updated to 222. (I did not skip or complete any workouts in this time.) I have been doing a mix of indoor/outdoor workouts. Not a huge deal, but it was a bit of a let down. I do enjoy building a plan and seeing how it affects my future ftp.
I have experienced similar issues presented here. FTP estimation just after an update has a very healthy (likely over ambitious) estimated number which tends to increase during week 1 of the block, weeks 2 & 3 it stays the same and then in week 4 it drops 20w+ to end up giving a more ‘expected’ handful of watts increase for that block. This same pattern has happened in both of the first 2 months of Ai FTP and I am a week into the third month and so far the pattern is following the same trend.
In the final week leading up to prediction date, I can’t see enough other variables to lead to such massive decreases if the workouts are performed and rated as expected. It feels like you are being penalised when you are literally doing everything it has asked.
We don’t need to go down the FTP is everything road but there is a relevant case here where these overly high estimations (to the point where each of the last two months would have been lifetime high ftp’s for me - previously 321w had been my peak) are unproductive and demotivating.
This is quite frustrating if you need that FTP figure to be within the realm of reality (which I did last month) as I had my main ‘A’ event in week 3 of the Ai FTP period which was an 8-9 hour mountainous gran fondo and needed to set my wattage pacing targets for the 4 odd hours of climbing. I chose to ignore the estimate and go off what I felt my ftp was based on past rides and 20min outdoor ftp tests which I have done for the last 10 years as a relative benchmark for myself. At the time of the event, my 3wk old FTP was 311w, it was predicting 335w ftp a week after the event, what actually happened was 307w the week after the event (the decrease likely due to skipping 2 of the recovery week rides post event).
Now a week into the next block it has peaked at 343w estimate and no doubt will end up back around 310-315 once the estimation date rolls around.
Are others getting this roller coaster estimation pattern where it spends a week climbing in week 1 and comes rushing back down in the 4th week to near where you started? Can we turn this 10% swing feature off, surely the Ai doesn’t believe it to be possible that a person could gain (or probably even lost) 10% of their ftp in a single 4 wk block.
@eddie Hi, anything else you could dig up by any chance? I see many people having very similar issues and no official response from TrainerRoad.
Outdoor workouts or indoor workouts?
I normally do 5-6 rides a week and only 1-2 of them is outdoors.
@luke22 Can I take a look at your calendar to see what might be happening with your case and share that with you here?
Same issue for me.
Seems to me like the “AI FTP prediction" is just a very simple general statistical average of what the FTP increase will be of Users that started training with Trainderroad, since new rider will often have 10%+ FTP gain if they didnt train much in the past.
But if you are already training many years this is simply unrealistic and the AI Prediction is simply wrong / expects you to rate each workout as Medium or lower to get the predicted results.
Iam sorry to say it but this feature seems to have “AI” slapped onto it to go with the hype but it isnt really “intelligent” at all and just a statistical average, but feel free to correct me since these are all only assumptions @eddie
Yes, I’d say that those are definitely all assumptions and incorrect. 
I’d push back a little on the “statistical average” take.
The system is trained on a large volume of athlete training data, and it’s looking at your specific training history and performance to estimate where your FTP is heading. It isn’t pulling from a general average of new rider gains. It’s also simulating your upcoming training to factor in how your fitness might evolve. So there’s a lot more going on than a one-size-fits-all calculation.
Of course, I can’t get into the specifics of how the model works – it’s complex, and there are a lot of moving parts we don’t share publicly. What I can say is that it’s designed to respond to your individual training, not just average trends. We see wildly different types of training from our athletes, regardless of whether or not they have a TR plan on their calendar, and each scenario will bring different results with FTP prediction.
Keep in mind that a prediction uses data from what’s in place (yet to come) and what actually happens moving forward. These are both moving targets, so it is tricky to get a perfectly steady prediction unless all of the prescribed work is done perfectly, and we don’t need to adjust your training at all. With that being said, what I’ve seen with athletes who do follow a plan well it’s been pretty accurate.
We have mentioned that there is a bug with inflated FTP predictions for athletes who do mostly Outside Workouts, so keep that in mind while we test out a fix for that.
I know that doesn’t fully address the frustration when the number drops after a good workout. If you’re seeing that consistently, I’d be happy to take a look at your calendar and see if there’s something specific going on. We’ve had a handful of people from the team look at cases where FTP predictions seemed to be jumpy, and in something like 99% of cases, there were valid reasons for it.
I know that there are bugs, edge-case scenarios, etc., so I’m always happy to take a closer look at each unique situation. Let me know if you want me to dig into your case on this thread. Otherwise, shoot me a DM.
5 Likes
Edit: sorry this is a little long and I sort of cover the same information twice, but I am trying to be as specific as possible.
I also have been unhappy with my predicated FTP. Note, I am not unhappy with where I land and my trainerroad / FTP and actually training. It just feels like the AI predication is not reliable.
There are a lot of caveats here, but my predicated FTP was like 267, with 1 workout left. I did an outside ride with HR only data, that dropped my predicated FTP to 260, and created a red day. I decided to attempt my v02 max workout on that day. I failed 45 minutes into the hour. That dropped my FTP to 252.
252 feels like the right FTP based off subsequent workouts. My all time high FTP is 265 (trainerroad), I was injured and started the year at a 200. So some amount of very quick ramp is expected.
It looked like the system was heavily weighting a couple outside rides that I had with power (not outside workouts, just outside rides) and was inflating my FTP based on those? I recently deleted those rides from my power records to keep them from interfering with the predication going forward. This power meter used to align with TR exactly, but I think recent TR updates have caused a mismatch between it and TR AI.
So 1 HR only outdoor ride and 1 failed workout was “good” for a 15 watt FTP drop.
Basically realized FTP predication is currently worthless for me if individual rides impact it THAT much. That being said, I am still happy with my training and actual FTP (I can always do a 20 minute ramp test if it feels off too much).
Same information as above - just using my actual calendar for details. This is March.
Successfully completed workout on the 9th - predicated FTP still around 266.
Successfully completed workout on the 12th. This was a true moderate. Still around 266 watts.
Biked outside on the 15th. HR only ride (not workout), FTP dropped to 260. A 6 point drop based on 1 ride with HR only data seemed extreme.
Failed Interval set on the 16th. I was surprised at that drop from my ride on the 15th and decided to try and do my full interval set on a red day (because you never know unless you try). I made it through 6/9 intervals and called it. That dropped my predicated FTP to 251.
So in the course of 1 workout I went from 267 to 251. The system could not accurate predict my FTP from 14 workouts / outdoor rides. It needed 1 HR only and 1 failed workout?
After FTP change to 252
I failed the workout on Friday the 20th in the last 1.5 intervals. But I think that was more disappointment with the AI predication system then anything else. (rides on the 20th and 21st were eBike rides).
I successfully completed the V02 work on the 23rd with my new FTP. I am pretty sure 251 is right (and I did go from 225-251 this block), but the way Trainerroad FTP predication got me there has got to be the most demoralizing way possible.
I contacted support because I was confused what FTP I should use (267 predication from the last successful, or FTP increase of 251). They said 251 is the right call - and it does feel correct.
Note: This is a pretty standard week for me. A mix of inside workouts and 1-2 outside rides. And has been what my calendar looks like since late January. Something that’s not reflected is that every Wednesday is a Red day (and has been since beginning of January).
Complete month is below to give a view into all of it. If my FTP had been slowly decreasing. I wouldn’t have been concerned. It’s the BOOM 15 point drop basically the day before that is so so frustrating.
Hey eddie thanks for your response. I have mixed ~50/50 outdoors and indoor trainings (keeping the outside riding to as close to the TR defined workout as possible) so the outside training bug may very well be the reason for the inflated prediction for me aswell. I just played around with the workout surverys and noticed that if I choose “Medium” for my hard interval session the prediction is mostly on point, and it seems unlikely for a long term rider that a hard / very hard session will become a medium session in a timeframe of 2-4 weeks. (I changed the survey back to my initial answer afterwards ofc). Thats how I got to these assumptions but it makes alot of sense that it may aswell be a bug with outdoor rides.
Interesting.
I know we’re working on this issue as we speak, and I’m hoping for a quick turnaround, but I’ve heard this one is complex.
Stay tuned, and I’ll be sure to follow up when I hear something from the team on that issue. I think it’s likely that it’s having some sort of effect on your predictions.
Sorry for the trouble!
It looks like you gave it two signals that 267 may have been too high. I like to think of the AI FTP value as what is the right value for training. Looking at your workouts after the update, it appears they were appropriate.
oh no doubt - after it was appropriate. But still surprised it couldn’t figure that out from the 9 successful workouts. It took me exercising on a red day for it to figure out what I should be at?
I think the previous workouts are try to confirm a trend upwards. So the last two might cause it to fit a different line.
Just adding my issue with outside workouts downgrading my FTP predcition.
Was predicted 289 in 7 days time. Completed an outside workout sucessfully and imediately went to 283.
Only 6 watts but as I’m coming into summer I’m going to be doing more outside workouts so not epxecting a great AI FTP development 