Love the new XPMC chart – what a story it tells! Wondering if I could use it for simple periodization (such as Hard Block / Endurance Block / Recovery Block). Any thoughts to that? I am suspecting that I am not getting the right mix right now as according to XERT my training load has been going up and my fitness down :-(.
One of the features we’re working on is to be able to schedule future workouts and to see what effect they will have on your XPMC. You should be able to use this plan out your periodization. Note however that what the XPMC doesn’t really help you with, is how your training is affecting your fitness. For that, you’re going to have to wait for Phase II of our plan as we describe in our latest blog.
Looking forward to both capabilities! I’ve read the latest blog post and I am excited about the plans. My - perhaps incorrect - understanding of the PMC from Training Peaks and Golden Cheetah is that they use an EWMA with a set decay window. I’ve always wondered if it where possible to have an adaptive model that learns (in the Machine Learning sense of learns) the appropriate parameters for the impulse function for an individual or a least a class of individuals. Even better is some sort of system (NN or other Machine Learning Model) that could leverage HRV and other factors.
[Actually at one point I was “messing around” with time series models including VAR (https://en.wikipedia.org/wiki/Vector_autoregression) models since we have autocorrelated data and multiple factors mutually influencing each other at different lags. I’d be most curious to know the forms if not the proprietary details of the models you are considering]
Well fortunately I believe we’ve already established causality of training and fitness levels (although perhaps based on your comment above, they might be negatively correlated ). Phase III of our plan will look to take the modeling beyond Banister’s basic Impulse-Response to something more comprehensive. There are other steps we can take still within Phase II. For example, we’re developing new features that show how work affects your fitness signature variables intermittently - the “glycogen depletion” effect. The replenishment can take hours to days and so provides another dynamic to the prediction model in terms of recovery time constants.
We have a lot of ripe information for various ML models to be able to make meaningful inferences. We could base things on our interpreted data (strain by system, recovery times, etc.) or with enough computing power, on the power data itself. And yes, adding in other data, particularly data that reflects what is happening outside of activity data like HRV, would extend the ability to predict fitness values much further i nto the future.
We are planning to publish the details of our models. This will allow our methods to be further validated, improved and extended to other disciplines.
Wonderful! Thanks for all the hard work and the intellect that went into this system. And as an added bonus its easy to use and very good looking
BTW – my comment about Granger Causality is that what underlies it is the extraction of the lagged impulse functions between 2 or more Time Series . So if we had Time Series one (exercise load) and Time Series two (Fitness Metric) we should see impulse functions in both directions (the more fitness the higher the sustainable exercise load and with every bout of exercise we should see the two impulse functions on fitness (immediate initial fatigue and then the delayed overcompensation related fitness gain). But really looking forward to the outcome of your modeling which is obviously far more mature than my musings. Thanks for taking the time to read
Thank you. This is very valuable feedback.
I’m also planning my upcoming training year. This year, I used Training Peaks and periodized time on the bike. On 27 Oct 2016, Training Peaks announced that you can create your training plan using TSS instead of hours per week. Since TSS and XSS are interchangeable, I think I’ll use it but substitute XSS for TSS. I do wish Xert had a weekly sum of my XSS. So, I downloaded my year of data in a comma separated value table, and added up my XSS for each week. The correlation between XSS and Duration was 0.968. The correlation between XSS and miles was 0.836. (I definitely need to throw in a ‘Granger causality’ comment during a training ride break with my friends!)
Had a Garmin failure 6/25, so that week has no data for XSS or Duration.
Was your Focus / intensity consistent throughout the year?
Yes–I’m not allowed to race or redline my heart rate because no one knows how hard you can safely train when you have a minimally positive coronary artery calcium score. But the fastest guys can’t drop me on the flat sections anymore, and they don’t have to wait long on mountain climbs.
This will be my second year of formalized training, and I’m planning on concentrating more on Focus/Intensity
I just watched Joe Friel’s Training Peaks University course titled ‘Season Planning with TSS’ and it was very interesting. One way to help with future planning is to use the XPMC. My Training Load is 86.21. So that means I’ve been training at a weekly XSS of (7 X 86.21) = 604 XSS/week. My Form is a +37 since I’ve been tapering down in season’s end. I shouldn’t have any problem increasing my XSS by 20% to 725 (Friel has a table for this that he developed for a day’s work out). Since my 2017 training is for a multi-day event at a moderate Intensity Factor of 0.66, and I can calculate how many hours to ride at that steady state: XSS ÷ (IF x IF x 100) = 16.6 hours (725 / (.66 x .66 x 100) for that week of training. I highly recommend taking the course ($99).
The XSS/week idea is interesting one to use for planning. An aspect that we will be including in our upcoming Phase II, is to identify not just XSS/week but what mix of low/high/peak will make up that training load. Varying your intensity during training will make sense since each of these systems are affected differently with training, thus need to be managed independently so-to-speak, towards to main goal. There are trade-offs not only in terms of training time but also in certain aspects of your ability to perform (in particular your lower threshold power). In your case, given you’re preparing for a multi-day event, raising your lower threshold power would likely give you best results. This would mean trying to spend more time training but at lower intensity, adding in intensity only near the very end of the cycle.
@ Scott – your Periodization chart is beautiful :-). I also applaud the idea of XSS/week - break it down into the different zones of training load as @Armando suggests. I’m wondering if I can plan my own periodization. I’ve used automated coaching from Polar Personal Training, Training Peaks and 2Peak. A bit frustrated with all of them. TP and 2Peak are perhaps more for folks who are fine tuning the high end of their performance seasonally. Me, at some level, I just need to get “stronger” (I know its more complex than that but hey). I am thinking something more basic that I could manage myself might fit the bill. Also I am very eager to see how the model of XERT evolves and to try to base my plans on it. Maybe I can buy packs of coffee and ship them to Armando and the other Devs.
Could use some right now…