Tracking Cycling Training Performance

Hello QS Forum! I’m new here. Not new to self-tracking, but new to QS. I was inspired by a recent DataFramed podcast that featured Gary Wolf of QS on the power of data-driven thinking for everyday life. I wrote a blog about it here, and intend to start some projects on this forum relatively soon.

I have lots of questions, but my focus is on the factors that influence my cycling performance. I would define my cycling performance using a concept called “progression levels” in my cycling training tool, TrainerRoad.

I’m specifically interested in why I can or cannot complete a workout that would’ve advance my progression levels (PL) on the day. TrainerRoad calls these workouts (that increase PL) “productive” workouts. They are hard. Sometimes really hard. But occasionally can feel more doable. There are days I fail them. And there are days I nail them. When I nail them, I call this good performance.

In other words, if TrainerRoad gives me a workout set to advance my progression levels and I nail it or fail it, I’d like to explore the associated features (of my previous training, sleep, nutrition, mood, etc.).

While I clearly have a fair bit more thinking to do on the setup, I’d be interested to hear folks’ thoughts? Has anyone looked at TrainerRoad PL before? What about other measures of cycling performance?

Looking forward to collaborating!

Great blog post, thank you for doing that. I don’t ride so am not familiar with this type of training, but I think it’s an intriguing question. Am I right that you are asking: “What factors most influence my performance on a given day?” That would be a bit more general formulation than “What allows me to nail my ‘productive workouts’ as defined by TrainerRoad?” The reason I’d suggest a slightly more general formulation is that I think uncertainty about the difficulty of productive training sessions will creep into your project that make you doubt your own analysis. Productive training sessions are calibrated using some kind of black box formula that you are unlikely to understand in detail, and so it’s going to add complexity/confusion to your protocol. Is there another measure of performance that would be more transparent?

Excellent point and thank you. I think you’re right that a more general formulation might keep the assessment on track and potentially more interesting. An alternative dependent variable i was thinking might be RPE or something like that. I’ll think some more on it, and thanks.

1 Like

I’m specifically interested in why I can or cannot complete a workout that would’ve advance my progression levels (PL) on the day. TrainerRoad calls these workouts (that increase PL) “productive” workouts. They are hard. Sometimes really hard. But occasionally can feel more doable. There are days I fail them. And there are days I nail them. When I nail them, I call this good performance.

The first question is why are you using PL? Whats model inside it and how it works? What are requirements for that model to be working correctly and do you meet them?
As i know most of end-user companies uses some models derived from scientific research (and implement their own) and i see there is problem that models from research studies have their requirements. For example Critical Power model require few Time-to-exhaustion (TTE) trials to model Power Distribution Curve. TrainingPeaks WKO software uses this concept, but have built model just on cycling training data which is mostly not TTE, but reqular low-moderate intensity training. In that case they took model built on TTE but fill it with data not from TTE and expect same results, which may be not the case.

Another question is what algorithm is used for workout recommendations and why it should be used? Is it proved to give best results in increasing cycle performance? Also cycling performance have different domains (short-term anaerobic, long-term aerobic and a lot between them) and training strategy differs based on goals.

I have indoor trainer and bike and trying to build science backed KPI’s to measure my training progression. What i’ve got for now

FTP. This concept seems to be widely used by athletes but there is few issues with definition of it and it is not always predictive of trial results. Some studies show low correlation between FTP changes and other physiological metrics changes (VO2max, LT, VT etc). Also someone may ask why it uses hardcoded period = 60 mins. Why not 55 or 65? No answers on that.

VO2max. Hard to measure because requires visit to a lab. Some standardized protocols can predict VO2max but accuracy is questionable.

LT. Lactate testing requires fingerpricking during ramp test which is hard to do often.

VT. Usually it requires lab testing (VO2max with measuring ventilation), but there are talk test protocol to roughly determine VT1.

CP. Critical power concept seems be interesting. It requires 3-5 TTE to be done and build power curve. Its seems to be better concept than FTP because you have multiple points (not a single one with fixed period). This model provides 2 variables: Critical Power (power that mathematically can be sustained indefinitely) and W’ which is a measure of Anaerobic Capacity (amount of mechanical job in kJ which can be done when we exceed Critical Power). This model also can predict time to exhaustion on different powers. CP is seems to be connected with MMSS / MLSS because model the same concept.

There are also a lot of other metrics, some of them is sport-specific. Like for long distance runners fat oxidation rate is pretty important.

I have indoor trainer and bike and for now i’m testing CP model. I’ve estimated my VT1 by doing talk test for a few times and using my HR at CP as a VT2 (it somewhat a proxy to real VT2 / MMSS / MLSS / AnT). I’m spending 80% time just below VT1 and 20% after VT2. I’m aware that VT from these tests is just estimations but since i’m cycling in recreational/health purposes i’m fine with that. Every 2-3 weeks single TTE can be done to update CP model and compare PD curves to see the response to training.

Environmental factors influence performance for sure. Lets assume recommended trial is 200W for 20 mins and you cant do it on monday, but can do and wednesday. That doesnt mean your performance changed from monday to wednesday, more likely the performance were same. But results arent same because of some other factors (sleep, stress, recovery etc).

To asses what factors influence your actual trial results (not base perfromance, because base performance doesnt change that too fast) you need to measure these factors and trial results and build some dataset. It’s better to go with simple variables which can be measured with lowest error rates (each device or subjective feeling expressed on a scale have a margin of error). After you have built dataset - some statistical skills are required for analysis. Some of examples you can find in my blog.

It’s not an easy task to measure nutrition, sleep, activity, psychological state. It’s time consuming. And it will take some time to build protocols / model / analyse data and find out causal relations between variable of interest and lifestyle factors.

1 Like

Hi @Max_Eastwood , thanks so much for that input. This is great information, and glad to see that you’re grappling with this as well.

My choice of TrainerRoad progression levels is tentative at the moment.

TrainerRoad, if you haven’t heard of it, is a cycling training platform, not unlike TrainingPeaks, Zwift, etc. I’ve been using it consistently for many years and it has taken me from a very slow, amateur bike racer, to a much faster (still amateur) bike racer. I love the platform, would recommend it to anyone, and will stick with it for life.

That said, they’ve fairly recently released an AI system that called “Adaptive Training” that leverages the platform’s million + ride dataset to come up with a tool that slowly and deliberately progresses a rider forward, in the energy systems demanded for their discipline and race calendar. So this concept of progression levels is one of the main outcomes of that. While it’s a proprietary black box to me (and I’m ok with that), my understanding is that the AI will recommend rides to me (based on it’s vast dataset and my own recent performance with workouts) and I’ll either accomplish them and advance my progression levels, or I won’t. Sometimes the AI recommends “Achievable” rides that it thinks I should have no problem doing. Sometimes it recommends “Productive” rides. And sometimes the AI recommends “Stretch” rides, a notch above “Productive”. Any ride that I complete at “Productive” or above gets me to a higher progression level.

It’s a giant black box for sure, but I’ve personally had success with it. Interestingly, there are times when the AI recommends I do a Stretch ride for a certain energy system (ex VO2 max), and I nail the ride. These are rare as these workouts tend to be devastatingly hard in the threshold or above energy systems. There are other times it recommends an Achievable ride and I fail miserably, also rare. But the Productive workout, the “tweener”, is interesting and more prevalent. Sometimes I fail them, and sometimes I complete them. And my question is, why? Especially for energy systems that I’m accustomed to.

Anyway, hope that clarifies some of my earlier points and thanks again for the thoughts.

Adaptive training on a complex process is pretty difficult. The best example I know of is SM2, the predictive algorithm for memorization practice that came from Supermemo and has been widely used and adapted in other learning software. The software is tuned to train to 90% success on the set of cards presented each day, and with daily training is pretty accurate. That’s a big accomplishment when you consider that the difficulty of individual items is not standard, so has to be estimated from past performance. There is some error, however, which is unsurprising since all kinds of things influence memory performance aside from how hard the item is for you and how much you’ve practiced it before.

Considering that the memory phenomenon SM2 is trying to model is much less complex than energy/capacity/performance on a ride, it’s unsurprising that TrainerRoad is not completely accurate. In fact, it could be even less accurate than it seems, as you don’t typically ride to failure and so your full capacity at that moment is not known.