AI FTP Detection Key Takeaways

  • AI FTP Detection is available for all TrainerRoad subscribers!
  • 38% less likely to overestimate your FTP vs. the 20-Minute Test.
  • 75% less likely to underestimate your FTP vs. the 20-Minute Test.
  • 40% less likely to underestimate your FTP vs. the Ramp Test.
  • Powered by machine learning, improved through 1,200 iterations.
  • Validated with over 22,000 athletes over 9 months.
  • Takes into account indoor and outdoor rides.
  • No need to do an “all-out” effort.

Today we're making AI FTP Detection available to all TrainerRoad subscribers. AI FTP Detection lets you skip the Ramp Test and simply click a button to get your FTP.

Using AI FTP Detection is as easy as clicking a button. No more stressful tests!

Although the outcome is simple, AI FTP Detection was years in the making. We applied machine learning techniques to data from TrainerRoad’s 150,000,000+ dataset of indoor and outdoor rides.

AI FTP Detection uses demographic data along with ride files from:

  • Indoor workouts with power
  • Indoor workouts with only an HR monitor
  • Outdoor workouts with power
  • Outdoor workouts with only an HR monitor

How to Use AI FTP Detection

You can use AI FTP Detection in two ways:

  1. Anywhere you see the ramp test in TrainerRoad, there’s a button to launch AI FTP Detection
  2. On your profile page on

Isn’t FTP Dead?

Rumors of FTP’s death are greatly exaggerated, but we have come a long way from simply basing workouts off of FTP. AI FTP Detection works with Adaptive Training to break the tight relationship between power zones and FTP.

Instead of VO2 max being a strict “120%” of FTP, Adaptive Training allows your VO2max and all other zones to be personalized. It learns not only an athlete’s power curve, but also their repeatability in their specific zones.

The outcome is a system that takes into account your genetic predispositions to training zones (ie: being really good at short anaerobic work but struggling with sustained threshold) AND how those relationships change during your training plan/phase.

Before we get into how we built AI FTP Detection, we need a brief history of FTP on TrainerRoad.

History of FTP Testing at TrainerRoad

Waaaayyy back when TrainerRoad launched in 2011, we had two ways to get your FTP; the 20-Minute Test and the 8-Minute Test.

The 20-Minute Test was based on the testing protocol employed by Hunter Allen, laid out in the power cycling classic, Racing and Training With a Power Meter.

It had athletes perform a 20-minute all-out effort after a warmup and a 5-minute “clearing” effort to hopefully reduce anaerobic contribution to the test. It took 95% of the 20-minute effort to get an athlete’s FTP.

Although this was “good” for some athletes, it had issues.

  1. Athletes had to pace their 20-minute test correctly. If they started off too hard or too easy, the results could be invalid.
  2. Athletes had to go “ALL OUT!” on this test to get a valid result, which was very hard for many athletes.
  3. The strict relationship of 95% of 20-minute power to FTP didn’t get all athletes into the best training zones.

The 8-Minute test was similar to the 20-Minute Test, except athletes had to complete 2×8 all-out efforts. Although pacing an 8-Minute Test was easier for some athletes, it still had the same issues as the 20-Minute Test.

We needed a better way that didn’t take pacing into account and wasn’t so hard for athletes to complete.

In comes the Ramp Test!

In 2018, TrainerRoad introduced a stepped fitness test that had athletes gradually increase their wattage until they went “ALL OUT!” and couldn’t go any longer.

We then took a percentage of the athlete’s 1-minute max power to set the FTP.

The Ramp Test was a success and:

  1. It was less physically draining and more palatable for athletes.
  2. It was easier to understand and complete without issue.
  3. Had a higher adoption rate compared to the 20 and 8-Minute Tests.

Although the Ramp Test was an improvement over the 20 and 8 Minute Tests, it still wasn’t without issue.  

Our percentage of 1-minute max power on the Ramp Test worked for some athletes but not all athletes. We also still had a strict relationship between FTP and training zones.

We needed to get better…

And then, there was Adaptive Training!

In 2021, TrainerRoad launched Adaptive Training! We introduced Progression Levels that broke free the relationship between FTP and Training Zones.

Fun fact: Adaptive Training started before we even had the idea of the Ramp Test.

Adaptive Training would dynamically learn an athlete’s power curve and repeatability, giving them the correct workout per training zone to make them faster!

It then continually monitors athletes’ performance and changes their training in response.

Adaptive Training considers:

  • Passed workouts
  • Missed workouts
  • Workouts you “struggle” on
  • Workouts that are “too” easy
  • Sickness
  • Vacation
  • Time off
  • Accomplishing stretch/breakthrough workouts
  • Tapering for Races
  • Race-specific goals
  • and more…

Adaptive Training was a huge step forward in cycling training. But there was still a problem.

Testing still huuurrtt and wasn’t an enjoyable experience. Testing also could make people anxious and not perform to their capabilities on test day.

So we imagined a world without tests.

We’re not all masochists

There are very few athletes that truly love FTP tests. They are masochists and generally weird people. ;-P

We knew we had an amazing dataset of athletes’ entire cycling history (indoor and outdoor). We also had years of athletes performing FTP tests and then doing structured, power-based workouts directly after that FTP test.

What if we could use machine learning to detect an athlete’s FTP, combine that with Adaptive Training (to take into account individual power zone variation), and let athletes skip the FTP test and just click a button! 😍.

We knew this would be a big, complex problem, so we turned to machine learning.

What is Machine Learning?

Machine learning (ML) is a subset of Artificial intelligence. We used a “supervised learning” approach to detect FTPs.

We wanted to look for patterns and trends in people’s training to determine their FTP on a specific date. To accomplish this, the team built “features,” which are specific and measurable properties of a dataset.

Some examples of features might be:

  • Age
  • Gender
  • Years Training
  • TSS in the last 3 weeks

The team ended up testing over 130 features.

We then took a subset of our data and trained an ML model on athletes’ training history and ramp test outcomes. ML went through that dataset of indoor and outdoor rides and tried to find patterns that lead to specific FTP outcomes.

We then validated the newly trained ML model against a larger dataset of training history and ramp tests and scored it for performance. Or, in other words, how close it was to predicting someone’s FTP.

We took that score and tried to improve it. The team would build new features, train new models, and then rerun them and score the performance.

This is not a small task, as this iterative approach took years. The secret sauce is in our unique dataset of planned vs. actuals and the specific features we wrote to describe our dataset.

We even built ML models to create features to feed into other ML models!

This might sound simple, but it was a very hard problem. We had to worry about things like overfitting, data validation, and dataset quality.

Here’s a good explanation if you want to learn more about machine learning.

Note: We are lucky to have cycling data before, during, between, and after using TrainerRoad. We also have indoor and outdoor data. If we JUST had indoor TR rides, we’d run into issues. But we often have entire ride histories from athletes (or as long as they’ve recorded data).

Note to Engineers: We know there are many more traps and “gotchas” in AI/ML, and this blog post in no way tries to cover all the things we had to watch out for.

Real-World Validation of AI FTP Detection

When we got the model to a state we were happy with, we went to early access for AI FTP Detection, allowing our athletes to opt-in to try the new feature while it was in the beta development phase.

Over 9 months, we had 22,000+ people join AI FTP Detection early access.

During this time, the team continued to check for accuracy by looking at expected workout performance vs. actual post-AI FTP Detection use.

We further iterated on our ML models until we were positive that AI FTP Detection outperformed the Ramp Test by a wide margin.

Is AI FTP Detection magic?

AI FTP Detection might seem magical, but it isn’t. It needs data to accurately detect your FTP…and the more data and the more recent, the better.

For those athletes who have had a recent break from cycling, AI FTP Detection will get you in the right ballpark. Then Adaptive Training will take over and adjust your training zones after a few workouts.

AI FTP Detection is genuinely amazing, and we recommend it to all athletes with recent data.

Wait, don’t other companies detect my FTP?

Other companies do try to tell you your FTP without a test.

A key issue with these companies’ approach is that athletes still have to make an “all-out” effort to get accurate results. Many times at a duration where athletes don’t usually perform all-out efforts (like 20 minutes).

This approach has the same issues that I wrote about above. And what if you never do an “all-out” effort? The systems still assume you did and give you lower training zones than you should have.

AI FTP Detection is different. It doesn’t even need power above threshold to detect your FTP. Athletes can do 8 weeks of traditional “long and slow” training before detection and still get an accurate FTP! 🎉

This is some of the “magic” baked into our ML and dataset.

This means that athletes can focus on the right training for them without having to worry about mentally and physically taxing “all-out” efforts to calibrate their training zones.

Congrats to the team!

The team has done a fantastic job on this and has continued to make it better!

This is a big day for TrainerRoad, but I think we’ll look back at this and see it as just the beginning of the paradigm-shifting advancements coming to cycling training.

Sign up or renew now to get faster

If you’re new to TrainerRoad, go to now to sign up and get faster!

AI FTP Detection FAQ

How is this different from FTP detections on other platforms?

Most other platforms estimate FTP from individual best efforts in your training data. Our model analyzes your training history and personal biometrics for a broader, more nuanced understanding of your abilities than a single effort can reveal.

Is AI FTP Detection just looking at what athletes generally gain in FTP after each block and basing the model on that?

No, we are not estimating FTP based on any static equation. AI FTP Detection uses a model trained on our data set to predict your FTP based on your unique training and biometric data.

Does AI FTP Detection consider rides that aren’t TrainerRoad workouts?

Yes. AI FTP Detection takes into account all rides with power data—indoor or outdoor, structured or unstructured. This can include group rides, races, and even rides on other indoor training apps synced to TrainerRoad.

Will AI FTP Detection consider unstructured outside rides as well?

Yes. AI FTP Detection takes into account all rides with power data, whether they’re indoors or outside, structured or unstructured. This can include group rides, races, and even rides on other indoor training apps synced to TrainerRoad.

What does AI FTP Detection look at to detect my FTP? What types of efforts or workouts does AI FTP Detection consider?

AI FTP Detection looks at your training history and your biometrics to detect your FTP. Because AI FTP Detection considers so much of your training data, it’s very unlikely a single type of effort or workout would substantially sway the model. For example, the model “knows” when you’re doing an easy ride and won’t confuse that for decreased fitness. You also don’t have to do capacitive or all-out efforts for the feature to work. In fact, it works if you only do sweet spot or aerobic riding and never do any high-intensity intervals!

Will AI FTP Detection detect my FTP if I’ve never used TrainerRoad?

AI FTP Detection requires 10 completed indoor TrainerRoad workouts to begin detecting your FTP. Once you reach this minimum, you don’t need to do any additional TrainerRoad workouts. By enabling RideSync, you can import rides with power data from other services and continue detecting your FTP.

Does TrainerRoad recommend I use AI FTP Detection or do an FTP test?

AI FTP Detection delivers the training benchmarks you need to progress your fitness and allows you to complete a productive workout in place of testing. Because AI FTP Detection gives you productive training stress without the unnecessary psychological and physical stress of an FTP test, we recommend using AI FTP Detection. That said, you always have the option to test!

Can I still complete an FTP Test?

Of course! You can still take a Ramp Test or other FTP test whenever you like. We just don’t know why you’d want to 😉

How often will AI FTP Detection give me a new FTP?

Depending on your current fitness, FTP generally takes several weeks to meaningfully change, so AI FTP Detection generally won’t detect a new FTP more often than this. For now, this is why AI FTP Detection isn’t available more often than every 28 days.

What if my Abilities Change Within 28 Days?

FTP is not the only measure of your fitness. Between AI FTP Detections, your Progression Levels monitor your progress and track subtle changes in your abilities from day to day, even when your FTP doesn’t meaningfully change.

Progression Levels track your capacity to express your current FTP in different training zones at any given time. For example, increasing your ability to hold steady power within your VO2 Max training zone from one minute to three minutes at the same FTP demonstrates significant progress and is reflected by changes in your Progression Levels.

Simply put, your abilities will change between AI FTP Detections in ways FTP alone can’t accurately reflect. Progression Levels have you covered, working hand-in-hand with AI FTP Detection to track every change in your fitness, whether big or small, up or down.

I Didn’t Accept my Recent AI FTP Detection. Can I Get a New Detection Sooner Than 28 Days?

If you keep your current FTP instead of accepting AI FTP Detection’s results, the newly-detected FTP will still be available for you to accept at any time during the next 28 days. Simply click the AI FTP Button from your Account Settings or from any FTP Test, and you’ll again be presented with the option to accept it, along with a countdown to when your next detection will be available.

If you haven’t accepted your newly detected FTP and something happens to significantly decrease your fitness within 28 days (for example, if you take extended time off the bike), AI FTP Detection will automatically adjust your still-available prediction downwards to make sure you get the right workouts and power targets. This will prevent you from accepting a too-high FTP, and it won’t affect the 28-day countdown until your next detection is available.

Does AI FTP Detection predict my hour power or expected Ramp Test result?

AI FTP Detection isn’t meant to predict the results of any individual effort, because all-out efforts can be affected by nutrition, fatigue, and other subjective factors. Instead, it’s designed to give you a training benchmark (FTP) that represents the full scope of your abilities, so you get the most productive training possible.

Isn’t AI FTP Detection just making training easier, and is that not the point of training?

A typical FTP test requires an all-out effort, but due to short durations and high resulting fatigue, these tests don’t generally contribute as much useful training stress as a productive workout. AI FTP Detection lets you replace your FTP test with a beneficial workout and avoid a stressful, less productive effort.

How can AI FTP Detection be accurate if I’m coming back from an offseason or more than one week off the bike?

AI FTP Detection was trained on an enormous data set and can predict decreases in FTP even after long (weeks to months) training interruptions.

How Do I Know AI FTP Detection is Accurate?

The machine learning model powering AI FTP Detection was trained on a dataset with more than 150 million workouts and hundreds of thousands of FTP changes, including athletes of all different ages, abilities, and experience levels. But the best way to validate AI FTP Detection’s accuracy is to try it for yourself! Accept your detected FTP and try some productive workouts. We think you’ll be amazed at the results.