Machine Learning-Based [HR] TSS Estimation - Too Low?

Brutal CX race yesterday. Average HR of 178 (96% of MaxHR) for 49 mins! TR TSS estimation of 64. Seems a bit low to me. Every other app/software estimates the stress/load to be a lot higher. Just seems like the machine isn’t doing as accurate of a job estimating as it should be?

Per the FAQ:

How accurate is Heart Rate-based TSS?

No TSS estimate is perfectly accurate. The machine learning model we use to estimate TSS is significantly more accurate than existing heart rate or RPE-based formulas. For rides of an hour or less, the majority of estimates fall within +/- 7 TSS of power-based TSS. For rides of one-to-two hours, most results fall within +/- 13 TSS. Accurate elevation and speed data also help improve the TSS estimation.

HR is naturally variable and a lagging style indicator compared to actual power data. As such, there are limits to how accurate any estimate could be with that as the foundation. I have no idea which of any of the estimates is most “right” but I’d take all of them with a grain of salt. Picking on the “low” one as the outlier could be valid here, but perhaps one or more of the high ones are just as worthy of questioning?

2 Likes

Perhaps, but it is training stress score, and I do feel like HR is a relatively decent measure of the stress of an activity. 64 TSS is the same as a 90 minute Z2 ride.

  • It’s an estimate, per their own definition “Heart Rate-based TSS Estimation

Per TR’s own info:

  • 64 TSSe value + 7 TSSe tolerance = 71 TSSe max
  • No idea if that gets reasonably closer to the other sources (none specifically listed), but it’s necessary to consider the delta within the tolerance range.

If you have concerns about this specific value, I think it’s best to email them directly so they can review it and perhaps give you better info than the support doc offers.

If it’s using speed data and most of your training is on the road, then wouldn’t a CX race underestimate TSS? I’d override it and just use 83.

I doubt that speed is a factor in the other calculations but we don’t know since there are not references listed to the other sources.

  • Based on what mod to the TR TSSe stated and limited HR/Duration info shared?

Your post says that accurate speed and elevation data help improve the TSS estimation, so wouldn’t that imply they’re used in the model?

I’d assume that the race effort for 49 minutes is at FTP. (50/60)*100 = 83. I did round up one minute.

1 Like

Notice I said “other” above, since we see it in the TR one but have no info on the other method(s) since they aren’t listed.

Gotcha. I don’t know why speed would be included though.

The other methods would most likely be TRIMP based or a simple regression on activities with both power and HR.

1 Like

Yeah, seems like speed would be absolutely useless for TSS calculation. So many variables affect speed(WAY more than HR).