I'll have to read through all that stuff more carefully, but that's really, really interesting.

So, if I'm reading this right, the Neural Network model is better able to predict diminishing returns from training stimulus than the Banister/TSS model is? Not only that, but it can also do so on an extremely personalized level for each athlete?

Exactly!

I'll go a step further and say that the Banister/PMC model, in and of itself, cannot account for diminishing returns *AT ALL*

For a given set of constants at a given time, it will always predict that if you increase load by x%, you will increase performance by y% e.g. if you go from 0-10 CTL, the performance bump is the same as going from 100-110 CTL. Any coach/athlete with any experience knows this just doesn't happen in 'real life'.

Once an athlete reaches a fairly moderate level of fitness, the model falls apart.

Bottom line: In this day and age, there are far better models available.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

Hi Mr. Couzens,

is the Neural Network a model to calculate training load data? And what would be different for example in the chart below to see the right turning points in training loads easier? The chart shows the recovery from a femur fracture and the pure fun of riding with a 1 month break in Mai/June ..

Your article is an very interesting read that challenges my language knowledge, too :-)

---

*
___/\___/\___/\___ the s u r f b o a r d of the K u r p f a l z is the r o a d b i k e .. oSo >>

Hi Alan, interesting stuff for sure (although the math and computer science is a bit much for me.) Could you discuss, in layman's terms, a bit about the predictive capability you've referred to? I've always used the PMC (perhaps ignorantly, I suppose) as a snapshot of where I currently am and where I've come from. Also, is there a way for the layman to put the neural network concept to use, or is programming knowledge required?

Exactly. And it's a huge problem, in my opinion. Because if I look purely at my PMC and pick the times over the past 10 years when it says I was the "most fit," those are the times when I was completely physically shattered and unable to perform well at all.


E.g. at high training loads it becomes a very accurate inverse predictor for me. Which is not at all what one wants in a PMC chart.


I'm really excited about this new method.

This looks interesting. Stupid question: what is the performance (fitness) measure E.F.?

Wonder if Coggan will chime in on this thread.

That was a really interesting read! Is there a widely available system that you recommend which uses NN? Also, what do you see as the period for calibration period of a NN based model to be a reasonable method of prediction?

It has always felt to me that there is a large jump in fitness to be gained in an initial fitness build followed by a plateau that doesn't seem to be accounted for in traditional tracking; however, I don't recall seeing a platform that tracks the diminishing returns and models through a method that "learns"-so to speak--an individuals response to high level training.

---

Blog | Strava

I think he might have been part of the purge.

---

My YouTubes

This is very interesting and of course makes perfect sense once you explain it. I always sort of built these limitations into my thinking when considering the PMC, and only thought of it ever as a "rough check" for where you're at, but not something that actually could predict performance. The idea that there is something that actually COULD predict performance is pretty cool!

Are there any programs or apps that use the neural networks approach to modeling? I'm not sure my Excel or CS skills are up to the task of doing it myself :D

Last logged on Oct. 20 so yeah

---

808 > NYC > PDX > YVR
2024 Races: Taupo

I hope so (purge notwithstanding). But I don't think there's any need to be defensive. I see this as building on the foundation that Coggan helped build, rather than undermining it.

But I do understand there are competitive forces in play given an eventual tool that uses these techniques (if proven effective in mainstream practical use) could undermine licensed products, etc.

Thanks Sausskross!

No, the Neural Network (in this case) is a model that takes training load data and predicts a performance.

Many many years BC (before Coggan :-) dose-response models like those used in your PMC were used to predict actual performances (i.e. the y axis, rather than being arbitrary units of CTL would represent an actual pace run, time swum in a given event etc)

In that sense it is, in my opinion, much more practically useful to coaches and athletes as we're talking in (and assessing the accuracy of) actual performance prediction.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

Hi Mr. Couzens,

with your answer I found some key words to read about .. so Neural Networks are a model to process information to assess actual performance predictions .. nice!

The mathematical theory will stay a mystery for me .. but for Neural Networks signs, switches and paths there is a living imagination ..

Thank you for your reply,

Hanno

---

*
___/\___/\___/\___ the s u r f b o a r d of the K u r p f a l z is the r o a d b i k e .. oSo >>

I agree with that, but never found that the practical limitations of using the Banister model, or any performance model, had to do with the form of the model but rather the difficulties of getting enough performance measures to feel like I could believe the model. Throwing more factors in the model in the form of extra perceptrons doesn't seem to address the biggest problem.

Thanks STJ!

Sure thing!

Best way to think of it would be a traditional scatterplot with training load on the x axis and some performance measure on the y - it could be a race, an FTP test etc. If you place all of your tests on there you'll probably have a chart of somewhat escalating dots, i.e. as load goes up performance will tend to generally go up to a point. Importantly, the shape of this 'generally going up' will vary between athletes (& it might even go down at some point!)

If we draw a 'line of best fit' through these points we have a basic predictive model. The accuracy of the model in any complex system (e.g. Human Physiology :-), will in some part be related to the 'bendiness' of the line. A neural network is bendy, a banister model straight.

When we talk about the 'predictive capability' of a model, we're talking about how far each true point lies from the line. The overall model error can be measured in the 'root mean square error', i.e. take the distance of each point from the line, square it (to eliminate negatives), take the average of those and then take the square root of that. When we do this, we have a measure for how good our model is (lower RMSE = better model).

While it varies between athletes (as shown in the data table), the average neural network model (for my data-set) has approximately half the error of the average Banister model, i.e. the model line falls an average of 2x closer to each point. So, if the average predictive error is 10W (from the model line to the actual performance) in a given athlete's Banister model it is likely to be closer to 5W when a Neural Network is used.

As it stands, there really isn't a lot out there in the way of predictive modeling/machine learning software for the endurance sports demographic so, yes, at the moment, Neural Networks take a little DIY coding. Speaking more generally about performance modeling software, Phil Skiba's RaceDay was one implementation of the Banister model and I know Mark Liversedge is also implementing in Golden Cheetah, but nothing that I know of on the commercial front using Neural Networks as predictive models at this point. Rest assured, they're coming though! :-)

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

E.F. is the efficiency factor (Normalized Power (or pace)/HR).

It's a standard metric on Training Peaks (though I think you have to be a premium user now to have it on your dashboard).

It makes for a good regular criterion measure because it's sub-maximal so you can accrue *a lot* of data points.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

A neural network is basically a smart look up table. Looks for correlations and patterns between results and input. We already do this with PMC and our brains. Way I learned to use it was to look at a season or two's data and learn from the patterns. e.g. My best performances were always around a TSS of 80 or so. Much higher and I was stale. Coggan did a good job of pointing out the effect of TSB. For a RR or TT slightly negative TSB was best, slightly positive TSB for a crit. Too negative or too positive and performance suffered. Also saw some good results with a double peak in TSS. Ramp up fitness to 80-85 TSS, ease off a bit to 70 and ramp up to 80 again and I was flying.

We're all different so everybody will have different values that give optimal results and these values might change with age and experience. I know the values that will usually put me in the best position but I've also had some great results at lousy numbers and horrible results at great numbers - so accuracy will be somewhere in the 60-80% range.

Machine learning might be able to discern some other patterns and hopefully make some predictions. Big question is are the model parameters it uses of any meaning or just random ones that give the best results.

Definitely worth exploring.

Thanks Trihawg!

As it stands currently, I don't believe there is a widely available system that utlizes NN's in endurance sport. There is a lot of application in (big money) team sports that will hopefully trickle down.

As far as calibration time, I typically find it takes ~6-12 months before the individual model beats the group model whether Banister or NN. The difference, though, for NN because it's so flexible, it can lead to some really funky patterns until the model gets calibrated! :-)

You're absolutely right. Currently, there really isn't a commercial system in endurance sport that utilizes machine modern learning to "learn an individual's response to the training." It's an area that is ripe for development in the near future.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

I do not know the algorithm used for Xert (https://www.xertonline.com/), but it certainly appears to be a multivariate model that is used to predict training load. Are you familiar with it?

That's the problem with proprietary 'secret sauce' models: NOBODY knows what algorithm they are using - even the people paying for it!! :-)

I suspect it's just a bunch of hard coded if-then statements rather than any true Machine Learning but who knows?

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

Damn! I can see H2O going but sadly we lose so many informative voices to the censor. While arrogant, he offered so much. Too many with thin skin here and we lose scientific and informative thoughts. The banned list has become a "who's who" of experts in the field....

Probably varies over time too, ie. after a few years of training the athlete has changed.

I'd expect the model would be better with raw power data for bike/run.