strive.ai’s vision was founded on two key pillars:  Enhance the Strava Experience and Enhance Athlete Performance through advanced analytics.

Impactful data science requires enough data to be successful so what we did to start was offer some interesting features that created data for us to leverage down stream: Thresholds and Segment Watts/KG Achievements.    While neither of these are products of data science, both are great sources for training interesting algorithms.  The thresholds data has been largely used to analyze athlete performance and segment watts/kg data has been used to focus on the Strava Experience.

I’m happy to report that we pushed our first Strava Experience related feature into beta today. We’re calling it Segment Performance and it’s a result of many iterations of training and testing regression algorithms that can successfully predict three key things related to a Strava Segment:

  1. Wind: We gather the observed weather for every segment effort with power two weeks after each activity.  We do this for the top 10K segments in the system.  This information is used in the next two algorithms so we gain a real sense of what weather related variables have a material impact on your segment’s performance.  This variable is a product of many variables (wind heading, wind speed, humidity, pressure, etc) represented as a z-score where a positive number is good and a negative number is bad for the weather’s impact on segment performance.
  2. Speed:  In this prediction we take the weather data and 100% of your 8 week power curve at your current PR time for a segment and provide a prediction of how close to your PR speed expressed as a percentage (i.e. >100% is a PR).
  3. Power: In this prediction we flip the target variable from speed to power to predict what power (watts/kg) would yield a Strava PR based on the weather.

Right now we’re able to get ~70% of the segments with sufficient training data to yield predictions that are pass our tests.  The Wind factor works on any segment where we gather the weather observations.  One of the ways we found to improve accuracy of these algorithms was to train the predictors off of solo rides only so the speed and power predictions are more conservative than if you can sit in behind a fast wheel.  We may try to include group rides eventually but for now we’re testing it this way.

Right now we serve this up in our iOS mobile app on the Starred Segment view through an action we call “Watching” a Segment.  Our backend takes all watched segments and pulls the weather forecast once per day.   Like our athlete centric algorithms we train, test and save the model every two weeks to improve accuracy.

We offer two views today that I’ve attached below:

Figure 1: Segment Summary

Segment Summary depicts the three predictions for each of your watched segments with a date time picker control allowing you to see the predictions for anytime from now until the end of forecast window gathered for the segments.

Figure 2: Segment Detail

The Segment Detail drills into a single segment to show all of the predictions as well as showing you a bit more information (e.g. number of athletes watching the segment).

Our plan is to add push notifications of either good wind index or possible speed > 100% predictions.  We’ll also start to compare forecast to the observations we gather post activity find tune when we pull each daily forecast.

The objective of this feature is to help you identify good opportunities to earn a PR so we’re also going to track how many PRs are achieved by athletes since they started watching a segment.  Additionally, we’re currently brain storming various segment related notifications to help athletes understand what it takes to deliver on their watched segments.

Please note the following caveats/points related to Segment Performance:

  1. Segment must be in our top 10K in terms of volume
  2. Segment must be on your Strava starred segment list which we sync each time you have a new activity
  3. Predictors are trained off of solo rides only
  4. Athlete must choose bike they wish to use in our predictions.  We use athlete id if you don’t track your gear in Strava
  5. Better results if you ride the segment frequently
  6. Algorithms are trained every 2 weeks

There will be more Strava Experience features as this is just the begining!

 

Segment Performance