I think this is a general weakness in AI FTP detection at the moment: it seems “sticky” when you stop training for extended periods of time.
I got COVID back in May and stopped training for a while, and when I came back after ~4 - 6 weeks, I only did super easy rides. But AI FTP Detection didn’t drop my FTP at all. I flagged this for Nate. But the more people who flagged this, the more likely TR will look into this issues and work on a fix.
I agree it needs fixing, after 6 months I would think AI should not be available and requires a manual test. Not sure what happens for a new user?
Hey there @Domslot and thanks for posting! Stoked to hear that you’re getting back into training!
In this case, we’d recommend taking a Ramp Test. AI FTP Detection needs more recent training data to produce the most accurate FTP prediction it possibly can.
The good news is that once you take a Ramp Test and complete your first block of training, AI FTP Detection will have enough data to give you an accurate detection the next time a Ramp Test is slotted into your Calendar. Progression Levels will get you locked in and drive your fitness forward in the meantime. 
We’re also working on a version of FTP Detection that’ll remove the need for the Ramp Test after an extended break while still maintaining accurate detections. There are some technical details that are still in testing, but it should be out in the wild soon!
Feel free to ping us back with any additional questions! 
What’s the ballpark cut-off in terms of time away from TR where one should start looking a doing a Ramp Test vs going with the number FTP AI gives? 1 month? 2 months? 3 months?
But the user literally has to set what the fitness decay is out of something like 5 options, otherwise it will be wrong. To me this is the reason the Xert estimate is not useful unless you have done a very recent max effort.
I’m interested in what the longest time off is too. Looking at a few extended work requirements this year that are going to keep me off the bike.
Many users leave Xert on No Decay (i.e. full fitness prediction) and have seen excellent results. Some have validated by testing after many months (some as much as a year) and have seen differences under 10W for predicted vs. actual FTP. Xert relies on good and complete data to establish pre-exisiting training:fitness patterns. If these are good, predictions are good.
Forgive me but isn’t no decay just no decay? I.e your signature just never drops from the last highest breakthrough? My wider point is that the user is left to decide what their decay is and their isn’t really any context about what you should pick other than experience of using it, which to me defeats the purpose. At least, that was my understanding from using it for several months.
No Decay means No Decay relative to the training load. Using No Decay, your FTP will increase/decrease depending on your training in lock step. So to give an example, say your last breakthrough ride was a month ago, your FTP was 300W and your Low Training Load was 50. Since then your Low Training Load has gone from 50 to 70. Today your FTP would be highest with No Decay (say 315W) and Small Decay could show a value of 310W. 315W is what the system would predict a breakthrough would show.
The purpose for having decay relative to expected values is to ensure errors in your historical data don’t cause an overestimation of your FTP (fyi, xert applies this to all 3 signature values not just FTP) and to catch possible negatively-affecting circumstances in your fitness. If for some reason, heaven forbid, you’ve fallen ill for example, your predicted values might need to come down a bit before you’re able to get a breakthrough. Having a decay is very important otherwise you may not ever become aware of the impact the illness has to your fitness.
Didn’t amber say that ftp detection would work after a long break? I seem to remember her saying she used it after a pregnancy and it was spot on. Has ai ftp detection regressed since then if it is no longer reliable for extended breaks?
AI FTP Detection treats each athlete uniquely and each athlete’s AI FTP is constantly evolving, so we don’t have a specific cut-off point for everyone.
@NateP It certainly hasn’t regressed! Each individual case is just different, and the more data we have, the better AI FTP Detection will be. As I mentioned above, we’re working on an updated version of AI FTP Detection that will be even better at dealing with long breaks from training for cases such as this.
Understanding each athlete’s case will be their own, it is interesting to see the OP’s case and have my own as a comparison where AI FTP recommended a bigger drop with less time off the bike. I took about 4.5 months off after some minor surgery and starting a new job at the end of August. I rode twice with power outside during the week of Sept 5th (Weekly TSS 188). Other rides were to pick up my kid from school on an e-bike (no power or HR data, and definitely not a workout). No other workouts or attempt to maintain fitness until December.
AI FTP detection had my FTP at 262 on Aug 1, and then it suggested a new FTP of 246 on Dec 11. I hit Train Now for the first workout after the time off and got Vigo (Threshold 1.9), and I failed to complete the intervals. I started the 4th of 5 intervals about 90 seconds late for a longer recovery, and I eased up 30 seconds early. Last interval was for the full time, but I couldn’t hold the power target.
I rated the workout too intense but kept that FTP and have been hitting the workouts well enough since. I suspect I’ll have an increase in FTP in 2 weeks when I let AI FTP do it’s thing again.