I moved over from xert to TR but one feature I do miss from xert is having an idea of how much I can increase my FTP after a training block - this would give me some much needed motivation as I’m slogging through the the last sections of a training block wondering what my 30 day FTP AI improvement will be for the effort!
Years ago at the start of TRs AI, this is something they said would be possible. It sounded like they didn’t roll it out because it was difficult to message and was heavily based on compliance. “3 months ago TR said I would gain 20W and I only gained 15!” Hopefully Nate’s non-comment means they figured out a way to overcome those concerns.
1 Like
Maybe some sort of confidence interval. Although everyone thinks they will get 95th percentile results.
2 Likes
Not sure if this is you telling me to ‘stfu that’s the dumbest thing I’ve ever heard’ or hinting something is in the works … please could you clarify 
There’s something new coming that’s currently in beta, so this is very much Nate having fun and not him telling you to stfu.
6 Likes
Let. Me. In. The. Beta!!! 
5 Likes
I think predicted FTP would be really hard because you just don’t know how the athlete is going to rate the workouts a few weeks out. To take a ride from “hard” to “all-out” I’d have to up the intensity by at least 5%.
I would think that would create a pretty large error factor, but I may be underestimating the AI.
I also hope to get a shot at the beta after consistently having success with their plans for 8 years.
You can’t quantify it. Doing a series of workouts <> FTP will increase this much. Physiology doesn’t work that way.
3 Likes
You could look at others who have done a similar series of workouts who are similar in previous training and say that you would expect a change of X% to Y%. The confidence interval could be pretty wide.
But if you’ve ever looked at research where they’ve controlled the variables as much as possible. You see rather a wide range of response. The confidence interval would be very low based on population level trends. It’s be about as accurate as using 220 - age for your max HR.
I should have said wide. But they wouldn’t be narrow or low.
Although… for many of us they have a great deal of personal (i.e. not population level generalisations) data about how we have each responded to previous stimuli. So they could maybe make predictions based on population data augmented with personal factors. E.g. they could say “for someone with these demographic descriptors, this current level of performance and these training history descriptors we would predict x, but for this particular person we have estimated a responsiveness factor of 1.2 so their prediction is x + a bit = y”
1 Like
I use this model, it’s the Bannister IR model (the inspiration for PMC). Xert is quite open with their methodology.
I started running again in late October, so I’m building a new model. I’m using a spreadsheet so I can see my model and manipulate it easily. I can copy/paste weeks of planned training scores, refit the model, and see how the model behaves.
Most here have the necessary data already.
1A - Accurate critical power (from MMP modeling or time trials), which indirectly controls your scoring system. I use GOVSS/XSS for running, and Bikescore (Skiba)/XSS for cycling.
1B - Consistent, repeatable time trial(s), or benchmark. I am currently using 3-minute max effort. This is used for tuning the fitting parameters. Performing a 5, 8, 12, 20, or 30 minute effort is ok, too. Doing routine 3 and 20 minute tests would be ideal, because you get 3 performance data points for fitting (CP from 3 and 20 minute trials and the trials).
2 - Fitting parameters are calculated by minimizing sum of square difference from performance test and model estimate.
3 - Influence parameters that can inform when to maximize training build and optimize taper from from the fitting parameters.
If you notice the “Model” column (1B), there’s an estimate of performance that can be forecast for any date. I would say rolling 4-6 week blocks is the limit to reasonable performance estimates. Trying to use the model for 3-6 months out has too much uncertainty.
IR and PMC models
The green line in the IR model is the performance estimate. You can see that in the future, performance initially increases, as workouts stop and negative stress dissipates. Then, positive training effect decreases from lack of training, and performance fades. Conversely, performance estimates actually decrease in the short-term from large stress increases and the negative training effect (NTE).
One last features of the model is the Influence Curve. This curve “informs” when a workout has the most effect on performance, and when a workout has a negative effect on performance for a race day. From my curve, the workouts around 26 days out have the most influence on performance, and any hard workouts within 7 days of a race will have a negative effect.
4 Likes