Confused about the Variability and Inconsistency of the AIFTP prediction

Training
1

So maybe I am confused and it makes sense to someone else. Here the backstory:

I was under a year-long TR Build-Your-Own Plan type of thing with 3 “A events” on 4/11/2026, 7/11/2026, and 9/27/2026. First one is Paris Roubaix Sportive (107mi) with the other 2 being climbing events ~100mi long. I am a dad to a awesome little guy who is 3.5years old so occasionally time can get tricky and the kiddo brings home random illnesses that may or may not get me sick but can also require a full-time dad mode period that can occasionally wreck training. 99% of my workouts are outside. I live in a FLAT place that would be characterized as a urban environment so there are at times traffic, lights, etc. I am, physiologically more of a punchy rider which TR verifies. With the change I was put on a 3day hard 1 day easy plan and I think it had me at demanding.

Now, onto the issue at hand.

The current plan I am started in October 2025, it was the previous Masters plan on build-your own. Training in November was essentially a wash due to successive weeks of whatever viral plague rolled through the house.

Started back in December and lowered my FTP from its previous 260w to 245w. Workouts felt appropriately challenging. Was pretty consistent for the Month with what I recall and looks like only missing 2 workouts from what my calendar says.

January:

First AI FTP increased FTP to 258w on go-live (1/1/2026). Seemed about right given the work I was doing and that was about where I was before. The Ramp Up for January with new system was pretty insane. as I was suddenly doing VO2 Max and Anaerobic workouts that were Significantly harder than prior. Whatever I worked through it, failed 1 interval workout in the beginning and I think I missed one recovery workout. On 1/10/2026 AI increased the FTP to 265w. Seemed reasonable and workouts after felt appropriate.

Here is where things seemed to start to go left. The Projected next AI FTP was 320w. Which, even in my delusional self-talk I knew was not a thing. Had a 24hr bug 3rd week and it seemed to adjust the calendar and workouts. Ramped me back up and all seemed well. Eventually the Predicted AI FTP adjusted down to 283 or so then dropped me to like 250w from 265w due to ending a 2hr indoor ride 45min early due to Dad-duty and missing 1 60min VO2Max interval session the same week. I was skeptical and did not accept it as that seemed to be a bit odd to go from a 15-17w projected increase increase to a 15w decrease with those data points.

February:

Whatever the above was going on I kept trucking on. The next projected AI FTP had me at 290W on 3/11/2026. Still skeptical of a 25w Increase but whatever we shall see what happens. Did the workouts as directed, they are outside so I just follow the Garmin prompts on the screen. As I said above, ended a 2hr inside workout 45min early on 2/1/2026 (same one mentioned above) and missed a VO2Max 60min session on 2/3/26 (same one as mentioned above). I just finished my endurance week, hit the intervals/durations per my headunit, and until today, the projected FTP was still 290w. Then, today, with nothing on my plan or workout done, its suddenly planned to be 260w. Not sure why the change occurred and also am unsure how compliance with the next 15 workouts consisting of 5-8hrs/wk of AI generated plans will increase my fitness/ftp by 3 watts according to TR, or more importantly, how did the projection shift downward 30w with no really explanation or reasoning that I can see.

Ultimately, while I was initially interested and enthusiastic about the AI FTP prediction, I am now skeptical and reticent to trust it and also, given that my workouts are heavy in the Spicier stuff currently, if this is best preparing me for my planned events. I am aware that I can “turn off the prediction however I am looking for some clarity regarding the above situation. As it stands, my faith in the predictions is currently lacking. If this is somehow due to the initially process over-inflating my FTP fine, but the workouts have seemed appropriate so not sure. If there is a current bug then that is cool too. I am not saying my ftp should jump 25w in 28 days, while I would love that I also know that is less likely. But the idea is that if I am to trust that, then the wide fluctuations without failing or missing a significant amount of the workouts is hard to say. I even played around with the demanding vs balanced vs number of days as well as changing workouts to see if the system would change. Nothing did so it appears TR is set on the FTP no matter what I have as my plan. Which also seems a bit odd but again, maybe I am missing something. TR folks can feel free to go through my calendar or whatever. I do not know how to link it. Thanks in advance. Ask any questions you may have and I will try to answer them.

2 Likes
2

A 55w FTP bump in 4 weeks?!?!? This is the point in the process where you must’ve spoken to someone, right?

3

Not sure exactly what you are saying here. When I started, I was coming off essentially a month off and prior to that my FTP was set at about 260 and felt appropriate. So the 320 was something I literally looked at and was like “yeah that’s not a thing” and chuckled with my friends about. Because while I would love to have a 60 W increase in about a month there’s no way that’s actually real. That was where my skepticism started. The 320 W prediction did average down to about 280 which seemed potentially reasonable given the workouts I was getting were super punchy and definitely more intense than the previous plans I’ve had with TR.

4

I’m saying that if it happened to me you could use the amount of time that passed between me seeing that prediction and me starting to fill out my support ticket as the new definition of a nanosecond.

Was that when it was in beta?

1 Like
5

Looking back at your training history, I think this may be related to your adherence during outside workouts.

Some of your workouts look great, but then there are others where the structure of the workout isn’t followed whatsoever. One good example of this is your level 9.1 anaerobic workout on January 9th. That ride had an IF of .22 and was rated easy.

I’m seeing other cases where it looks like you somewhat followed the theme of the workout, but not the specific structure. December 25th and January 19th are good examples there.

In short, I think the software is seeing all of these really high-level workouts on your calendar, and then sometimes the execution and survey responses don’t fit what we’d expect to see.

Additionally, you’re adding in a fair bit of unplanned, unstructured riding, so we’re having to constantly adjust to that as well.

These are my initial thoughts as to why you’ve seen significant changes to what we’re predicting your FTP to be in the future. Your training hasn’t been super stable, and sometimes the data is very off.

Let me know if this makes sense and if you have any questions. I’m sure we can find a way to get things working a bit more smoothly if you’re interested. :+1:

1 Like
6

Honestly, this sounds more like an algorithm overreacting to incomplete data. Ending a long ride early and missing a VO2 session might have been interpreted as reduced capacity. I doubt your real FTP fluctuated that much, it’s probably the prediction being unstable.

1 Like
7

First, that ride on 1/9 was unlinked with the scheduled ride that day but for some reason your system still labeled it as that. I have had that happen before but not often as I tend to just do the rides as scheduled. I am not sure how I can do more on my end than unlink the ride so it isn’t considered the plan. Previously this was fine before the AI plan update.

Second, as for the unstructured parts, those occur, in addition to the structured plan, and isn’t that the point of your system to account for whatever happens and meet me where I am as well as show me where I should be? Not being argumentative as I am trying to see what the limits are because the predictive system seems to be rather inconsistent. Due to my location there are times when before or after then ride it requires travel to the area where it is safe to do the intervals hence usually 20-30min of riding before the workout starts as well as start/stops depending on traffic and time of day. Previously this wasn’t an issue obvious as adaptive training is different.

Despite that, how is that the prediction went from 320w >> 280w, then from 280w >> 257w, then recently 290 >> 260w. Either the prediction was wildly (read: delusionally) optimistic or it was wrong. I get that there is a level of “trust in the process” but in order to do so, there has to be a baseline level of trust for the process. At this point, the workouts and numbers seem either appropriate or good compared to prior years. However, while I feel like I am improving, analytical data is by its nature not subjective, or at least less so. Ergo, I would like to know if what I am doing is actually making me faster by empirically measured data and not just vibes.

I have been a TR user for quite some time through a host of various levels of available training time. I pay the money because I appreciate the product. I have listened to the podcast and consumed the metric tonne of information being handed out since back when it was Amber, Chad, and Pete regularly on the in person podcast and you used to breakdown Nate’s race footage. My issue is not with the product as a whole, it’s that I am not convinced that the predictive ai ftp is accurate as the fluctuations are rather big. If you look at my training since aiftp as “not stable” I would like to understand how when I think I have had 3 missed workouts (30min end of the week recovery spin, 2hr recovery week ride, and 1hr VO2 max interval session). That is essentially 83% compliance since AI FTP Adaptive training went live for me on 1/1. What is considered stable if that is not?

From what I am seeing, the algorithm does not really seem to work with me currently for predictive FTP. From what I can tell, the prediction algorithm was potentially inflated in the early part and is walking it back, it is taking unstructured rides and treating them the same as efforts/capability and weighting them equally (which seems like a very flawed premise given a family bike ride with my kiddo is nothing like a 180w endurance ride), or it honestly has a much longer on-ramp than it seemed and need/needed more data somehow. At the end of the day, explain to me what the metric I should use (data not vibes) to determine if I am getting faster and the whys behind it. My thought about the new process is that it, and the system, still doesn’t really handle outside rides as well in its current iteration and likely would be more accurate with indoor riding (which I frankly loathe). Also, still not sure how it went from 290w predicted several days ago to 260w currently. Those measurements are not precise even if one is supposed to be more accurate than the other.

I do appreciate you responding and again, swallowing as much ego as possible on my end, looking forward to trying to rectify the issues or at least address the data that is relevant. I really have tried to focus and be consistent with my training so understandably this matters.

8

This both my thought and concern. If the algorithm is not reading/applying correctly then my faith in the prediction fails. Further, it begs the question of every ride/activity is weighted the same. Also, if the system only see watts without context, how does that not potentiate a situation where it does not in fact improve as much as they might otherwise be able to?

9

Ah, I see. I am not sure if I was in the beta as it seemed to go live around the same time I had the AIFTP on my account, but the variability has happened since so it seemed to be consistently inconsistent. I did not fill out a ticket as the first prediction seemed off and I chocked it up to needing more data. As it went down to the 280w I was questioning but seems plausible. Then when it dropped several days ago or so and then it projected 290w for the next ftp check in 28days I was even more skeptical. When this was followed by a sudden predicted drop from 290w to 260w in 28 days without a real understanding of what precipitated that; especially given I would somehow not make any gains despite projected interval growth in WLs, it prompted the post.

2 Likes
10

@eddie

So, I am not sure if this is a bug but I am thinking it may be. Previously I did not accept the AIFTP that was published on 2/11. It was for 257w (down from 265w that was previously the FTP given to me by AI). Suddenly a few days ago the predicted FTP on 3/11 dropped from 290w to 260w as I noted above. Well, today, I said fine and accepted the prior AIFTP. Now, suddenly my predicted AIFTP for 3/11 is 282w… It seems that perhaps the act of not accepting the predicted FTP somehow can cause the prediction algorithm to go wonky. Conversely, if it is working appropriately, can someone explain how it somehow decided I am going to get nearly a 10% increase when it 1. just decreased my most recent FTP, and 2. suddenly made my gains so much better because my threshold is set 8w lower… The adaptive training process I am digging but I am definitely not sure about the validity of the predictive modeling. Not sure if @Nate_Pearson or @Jonathan is better able to explain how the numbers math out when they seem to be not in my head. Again, I appreciate all that you guys do and I am sure the chaos that is happening behind the scenes is a bit like herding cats. IF it is a bug, please let me know if only to satisfy my own curiosity. Thanks.

11

Sorry for the late response here.

We’re actually digging into a few cases where FTP predictions had swings in one direction or another.

I’ve added you to that list to investigate, and I’ll follow up when I know more.

Thanks for bringing this up and for your patience! :handshake: