I’m a bit puzzled by how my AI FTP prediction is behaving and would appreciate some clarification.
Right now, my FTP is set to increase by 6 watts (+1.8%) based on a workout I haven’t even completed yet (Threshold 3.2 – “Madison”). That effectively means my FTP for the next 4-week block could differ by 6 watts depending on whether I do that single workout or just rest, which feels hard to reconcile.
I also noticed something similar about 1.5 weeks ago: at that time, with no workouts planned in TrainerRoad, the system estimated my next FTP to be 228W. But even if I hadn’t trained at all in that period, 228W wouldn’t realistically have been my correct FTP either.
So I’m struggling to understand how these predictions are made, especially when they seem to shift significantly based on limited (or even no) actual training data.
I’ve been able to repro this too with different results. I’m able to get an increase of 3% with no workout the day before the prediction and a decrease of 1% depending on the steps I take to get the workout in place or whether or not the previous workout has been ‘marked complete’.
I’m also able to get a shift in 18w based on getting the same workout to appear through two different methods the day before the AI FTP predicts.
I strongly feel like there should be a diminuished weighting throughout the month from prediction to prediction. It’s very unreliable at the moment and I end up having to game the last few days to avoid a big drop from 2 workouts (like, 20w)
Let’s say my AIFTP today is 250 watts. I’ve been doing 5 rides a week for the last 8 weeks. My next workout is tomorrow. It’s 2 × 20 at 98% of FTP. I did an easy endurance ride today. Am I happy doing that ride with the intervals at 245 watts? Bring it on!
Would I be happy doing them at 245 watts if the workout was in 10 days time and I had zero workouts planned between today and 24th April? Not so much.
I suspect the colossal amount of ride data shows a lot of failures in that sort of scenario.
That workout without seeing the rest of your calendar is probably your highest threshold workout of the block. If you remove it you haven’t demonstrated that your AI FTP should be raised to train at a higher FTP.
It is getting more and more strange. Now I have done the workout. Good feeling today, so did a little extra, and got a score of 3.7 instead or 3.2 And now it predicted 281w. 8w more that yesterday, and the only new data is one workout
Unsure why that is strange. It is trying to determine the best FTP for training. That is its number one goal the fact that it is correlated with biology and other FTP test methods is a bonus.
agree, if you prove you can do Madison 4x long LT intervals … plus mark it as hard(could do one more interval but it would suck…) that directly shows then to the system that you could do 5x Threshold intervals… “so then YES, ready for higher power, don’t need to do 5x of that… that would imply closer to sweetspot!..which means their FTP is actually higher!.”
Just like if someone runs in training 23miles at 7:00/mile… most likely that would mean… their Marathon goal pace SHOULD NOT BE 7:00 anymore!!! Don’t need AI for that!
I just noticed that now your next prediction is a 16.4% decrease, which seems reasonable given that there is no planned work between now and then.
That Madison workout was a decent jump from your previous threshold workout (2.9) and much bigger than the one before that (2.0), so it was a relatively weighted workout right in the final days before your FTP Detection.
I get that 6 watts seems like a lot, but you’ve been taking big steps with the last few threshold workouts, and ending with a level 2, or 2.9 workout is different than a 3.7.
I hear what you’re saying, but I’m struggling to trust that the AI really understands what it’s doing.
Two weeks ago I was on holiday around Lake Como (in Lombardy), so naturally I had no scheduled workouts in TrainerRoad. I still rode a bit: two 1-hour rides and one 3 hour 15 minute unstructured ride. One included 8 minutes at 288W, and another 55 minutes at 239W (not all-out efforts).
After these rides, I checked what TrainerRoad predicted my FTP would be in 10 days, and it showed 228W. The issue is that even if I hadn’t ridden at all during those 10 days, 228W still wouldn’t have been a realistic FTP for me to train by.
I didn’t just sit completely idle either—I completed four threshold workouts in the 2.0 to 3.7 range. Solid work, but nothing that should cause a dramatic physiological jump.
Now, suddenly, the AI suggests I should train at 281W—not 228W like it did just 10 days earlier. That swing is hard to make sense of.
For context, I’ve been cycling for 20 years and training with power since 2013, with fairly stable fitness year-round. I’ve also been a loyal TrainerRoad user for years, which is why this concerns me.
If you have the time, I’d really appreciate it if someone could look into what the “new and improved AI” did with my training in February. I fully committed to the plan—no work, good sleep, everything dialed in.
I’d genuinely like to understand what’s going on here and regain my trust in the platform.
I’m struggling a bit to find the issue here, so please clarify where I’m missing things.
It sounds (and looks) like you’ve pretty regularly had an FTP in the ballpark of 280 in the past. Since you haven’t recently been putting in planned workouts on your calendar, at the start of each new FTP phase, you’re seeing a prediction that’s 4 weeks out showing a massive FTP drop (this last one shows -16.4%). Then, as you work toward that detection date, your FTP predictions grow back up toward that 280 figure until you reach your detection date (like the 281 you got today).
Is that what’s happening?
If so, that makes sense to me, as the model doesn’t initially have any evidence of planned training data and therefore assumes you won’t train at all over the next four weeks. This would result in a massive fitness loss (16.4% is our estimate, for four weeks out).
Each time you log an activity, we have more data to use for your actual fitness levels, though, and as you near that detection date, we have much more data than expected, and we can see that you’ve been training to some extent, so we’re able to update where we expect your fitness to be on FTP Detection day.
Let me know if I’m missing anything here, but I think this all makes sense to me given what you’re feeding the software.
I’m happy to dig in some more, so please let me know your thoughts!
