I replaced a hard workout for today with 103 TSS (Traveler) into 107 TSS (Karakoram +2), and the AI FTP Prediction went from 300 FTP to 283 FTP. I understand I should not do ‘not recommended’ workouts, but a 5% drop in AIFTP sounds for me a bit unrealistic, since the workouts are quite similar. Why this huge drop in AIFTP prediction?
Your workout on Thursday has adpated down. What happens if you put back Haw?
I think it also assumes that all “Not Recommended” workouts end in abject failure…
I was a bit stubborn and I just performed the Karakoram +2 and rated it as ‘Moderate’ (the hard intervals were intense, but doable). Now the AIFTP prediction remains at 301. I hope it will stay at 300+, since it will be my first time hitting 300+ 
Curious did it keep Seewell or did it revert back to Haw?
Also would recommend reading up on what the various RPE survey responses should mean. 90min of that was “Moderate”? looks pretty tough!
My guess on this would be:
- When you chose the harder workout, the AI assumed you would have more fatigue and this would negatively impact your next hard workout
- The AI then downgraded your next hard workout (the Thursday one) to ensure you don’t fail it
- End result, it assumed, was less training across the time period (not just the one workout), hence the lower FTP prediction
When you reported “moderate” it possibly then thought “oh, the fatigue wasn’t so much.” After you entered that RPE, did it then update the (previously reduced) Thursday workout?
I’m starting to think this is true. Can @eddie confirm this?
I think the others are on the right track here.
That Traveler workout was likely going to be the last and most recent threshold workout to feed the model what it needed to settle on a prediction for the 9th. Putting a workout that it thought you’d fail in its place likely made the model think that it would then only have prior passed threshold workouts to reference.
Those other updated workouts also play a small role.
@patrickhill Not Recommended workouts often do have a high failure rate. It’s not always that you’re most likely to fail the workout, but if the failure rate is above a certain percentage, we won’t recommend the workout.
I get that. But scheduling one hammers the prediction, even if the failure rate is still very low (like 2.6%). It looks like the simulation is predicting failure when it’s still highly unlikely. Then based on that assumed failure, the failure rate of future scheduled workouts increases, causing a domino effect that further hammers the prediction.
Hopefully, TR can refine the prediction to more elegantly account for “not recommended” workouts and to also account for the bias of the assumed survey response.
It reverted back to Haw.
Usually I rate these kind of workouts as hard, but for some reason I enjoyed this workout. It was nice weather, my legs felt good and after the workout I had enough energy to do more if I want to.
I see a similar thing, if I replace Obama (2.2% chance of failure) with Red Lake +1 (3% chance of failure) my prediction goes down 6 W (although mine is 24 days away so the butterfly effect is bigger).
I’m kinda tempted to take a stab at Red Lake and see if the 289 W prediction sticks or if it goes back to 295 W if I successfully complete it. For science.
Edit: Failed 
But interestingly, my prediction is back up to 294 W.
lol - yeah it is very interesting, you’d think a fail would be worse… but maybe it failed with lower TSS/IF… so now you are going to be capable of more work sooner (vs burying yourself in a stretch workout). Who knows the black box over reacts a lot, especially until you have done the workout.
Yesterday I downgraded a workout barely, prediction went down… then did workout fairly on point, maybe a hair stronger, and my prediction went up. I always leave a note to self on my prediction date what the original prediction was… quickly downgrades over the weeks… but generally comes mostly back with my final Tuesday workout (prediction date is Wednesday).
I just started the same type of Long SupraThreshold (2.5’ vo2, 30sec ez, 2.5’ vo2)… oh my - some real F’ers! LOL
I was just messing around with my calendar and I actually think this behavior might be a bug. It’s not that adding a “not recommended” workout downgrades your prediction because it thinks you’ll fail, it’s like “not recommended” workouts aren’t considered at all in the prediction for some reason. If I delete a workout from my calendar it causes the same reduction in predicted AI FTP as if I replaced it with a “not recommended” workout; it’s like the algorithm doesn’t see it. Maybe it’s intentional to dissuade you from doing “not recommended” workouts, but who knows. @eddie, would this be worth looking into?
I would assume that may be related to the failure rate. If we assume that you’re going to fail the workout, then it’s not going to give a ton of useful data other than that the workout was too hard, which we likely knew already since it was not recommended…
I’m tagging the team on this, though, to see what they have to say.
A failure rate of just 3% gives an assumed fail?! Was going to post this earlier - comparing my vo2 alternate options, very very similar Hard/V.Hard/Fail numbers… yet one is determined to be “hard” while the other one is “Not Recommended” … seems strange (or that the % hard/VH/fail need to be calibrated a bit better???) . If it’s not recommended and sure i’m going to bomb or fatigue blasted race day effort… shouldn’t it be higher chance of Max Effort, VHard or Fail? 
Like going to hawaii or FL in the wet season and seeing a forecast 2.8% chance of rain all vacation… but then the weather person on TV saying immediately after the forecast… "to bring your umbrella, your going to need it!! " lol
I’m not sure that on the back end it’s as simple as 3% chance of failure as shown in the workout..
That’s something I’d have to dig into.
I’ll follow up with what I hear, though.
Like here is my Thursday’s easiest ‘harder’ option (only three are listed as Similar)… so sometimes I will look at Harder (often barely harder). Easiest ‘harder’ is marked NotRecommended. Interestingly it is pretty close to my “benchmark” FTP test tester workout… 3 or 4x 7min threshold intervals. If you can tough that out, it is probably pretty close to FTP… it says I shouldn’t do it. :-/ bit sad for me (although i bet it would be tough!), especially considering my FTP prediction of 340 coming the next week. lol
BUT-I get it, no need to be a hero in the workouts. BUt in this case the fail % and overall % Hard/VHard are very similar.
