This is on our list of accounts to look at and watch. I’ll follow up as soon as I can! 
No hurry!
On that note … the overnight refresh has just put my FTP prediction back up to 255w
Just completed a workout (predicted as hard … rated it as hard) … FTP prediction drops down to 251w again.
That is in realms of power meter accuracy
My point is that, in my opinion, completing a workout (in erg mode) and rating as per TR prediction… shouldn’t lead to a 5w drop in FTP prediction.
Correct. The message at TR is:
“Follow this plan, rate it as predicted and you will get fitter”.
Instead some (not all) users are seeing:
“Follow this plan. Rate it as predicted and you (a) will get fitter, (b) see no change or (c) lose fitness. Oh, and we don’t know what variables determine whether you will get a, b or c.”
Some suggest disabling predicted FTP. Some suggest that TR FTP isn’t tracking your fitness so don’t worry about it, just keep training. Both reasonable suggestions. OK, then how do we determine if what we’re doing is working?
To be honest I think the plan is working for me (I must be getting fitter as the workouts are having increased periods at higher watts). However the way TR currently rates your progress in relation to a predicted future measurement is a bit wonky.
The best way to train AI is in the wild… so let’s go with the flow for now 
Old fashioned opinion, test your FTP (or whatever power duration you’re working on.) It amazes me how we’ve got to a point where riders are prepared to spend time and money on training, but aren’t willing to do tests, which are actually quite often very good workouts and provide excellent information.
I’ve been coaching High School for 20+ years… and there’s a decent percentage every year that want to train but not race. 
As you said, if you want to know if you’re improving… do a test. A race… a time trial loop on your favorite road/gravel/trails… rip up you favorite local hill. Something. Anything.
Of course. But there is a limit on how frequently you can test and not impact fatigue. suppose the test shows no gains. Now what? What went wrong? Too much? Too little? Poor motivation during test?
Or we could pay for a training package that purportedly uses AI to predict aerobic fitness stimulus-response and can recommend the optimal workouts to get faster
Absolutely, but if you doubt (and it seems from the forum there’s quite a few that do) the AI’s opinion, what other option are you left with but to test? Each to their own, but my opinion is testing every now and then doesn’t create that much fatigue, gives good workouts and provides valuble information, of course other people may have complete faith in the AI, that’s perfectly fine as well.
My FTP prediction when I woke up this morning was 253w in a few days time (0w increase), but having just completed an hour of zone 2 (predicted as easy, rated the same), my AI prediction immediately dropped to 252w.
Maybe because you only did an hour of Zone 2. 
It looks like tomorrow’s workout changed after you completed today’s.
I’m not sure why the suggested change would drop your prediction, though.. 
Ha … I only do what is in the plan! Not a minute more, not a minute less! TR is my master! I even sleep in ERG mode! 
Would be good to get some insight soon on this thread… it’s 4 weeks old.
I’m heading for a month of training and a 1w drop which makes no sense to me. It’s not lack of fuel and sleep as I always rate the workouts as expected.
I wonder whether the way zwift deals with workouts, and ERG mode especially, could be causing issues for how TR sees the completed data? This is not based on any evidence of course… but I thought I’d mention it as all my workouts are via Zwift in ERG.
Well my new FTP has just been detected (251w, down from 253w last month).
My new FTP prediction in 28 days time is… 253w!
I’d love to know why the FTP builder programme… isn’t building my FTP!
Edit to add (1w in 6 months feels underwhelming)
That is only 0.39% change. Perhaps the prediction ought to have a minimum amount it will change at. Or use high confidence intervals etc.
@eddie Any update on this?
I completed week 3 of my block 6 days ago. I’ve skipped a couple of recovery week rides and reduced today’s ride from 1:45 to 0:30.
After my 3hr upper end endurance ride (workout: Surprise) my estimated FTP was 269w.
Today my FTP detection is 265w with no gain for the block even though I successfully completed all my hard rides at moderate perceived effort and the endurance rides at easy.
The estimated FTP detection is now 272w in 28 days. As much as I would love to be capable of gaining +7w in a block, I’m not usually able to. So, the detected AI is clearly going wrong or does missing a couple of endurance rides in recovery week really cause you to lose 4w?
I don’t know what is going on and not sure if it’s the AI or something I’m doing wrong with the training?
p.s. I just realised I hadn’t added any perceived efforts for my short commutes to/from work. No power meter on that bike and sometimes no HR monitor. I think I noticed the FTP prediction was going wrong due to that so perhaps the FTP detection is affected as well? Those short commutes without perceived effort being added never seemed to make a difference in the past. And now that I have adjusted them to easy there is no way for me to trigger an AI FTP detection, so I’m stuck on the lower FTP.
Also, my last over-unders ride was at 4.3 and the next one has dropped to 3.6 for the new block. The third one for the block is listed at 4.5 (I would have thought the last one would be more like 5.2). Make no sense to me. Is there an explanation anywhere why it would drop so much?