Imagine you wake up and your mobile device has a strive.ai notification suggesting you should hit 3 of your starred segments that afternoon because it’s predicted you have a 90% probability of achieving a Strava PR based on loads of information (e.g. your current power curve and the weather forecast).
This is what we’ve been working on for 2 weeks testing Segment Performance regressions and factors for about 15 segments. So far we’re able to accurately predict segment average speed on 60% of the test segments. As you recall we’ve made several improvements to how we test and deploy athlete specific algorithms and now we’re taking that to the next level for the Segment predictors since there is already >100K segments in the system and nearly 2M segment efforts to train from. Regressions (when we try to predict a continuous variable like average speed) are harder to dial in than classifiers (e.g. what bike you rode) so we’ve added more smarts to the system allowing it to find the right algorithm and variables for each segment. We’re about to turn on batch process to gather the observed weather for your activity efforts to train/test the top 25% of starred segments in the system.
We’re working on a story board for this feature to make sure it’s easy to use and fun so we’d love your ideas on how to create a winning experience.
Some of our considerations on this feature:
What power to use on the segment performance prediction? Do we use your best recent observed power on the segment itself, your recent peak power for the closest completion time (e.g. peek 4 minute power) from a power curve or get more advanced and predict what power you can produce based on the weather and gradient on the segment?
We validated that wind speed, wind heading, temperature, cadence, power, athlete weight and even humidity are the highest factors on majority of the segments that aren’t down hill. Other variables impacting the prediction are what bike you’re riding and if you’re in a group. This makes sense if you’ve ever the epic app, Best Bike Split, as your specific drag coefficients play a major role in how fast you go in a specific weather situation. This means we need enough examples of you riding a segment with power to accurately predict your speed in the future. This is why we’re thinking starred segments are good ones to focus on. Eventually we’ll collect more information about you (e.g. height) and your bike (Road, TT, MTB) to estimate your drag coefficients allowing us to do decent predictions without including gear_id in the prediction.
As you can see this is quite involved! As I mentioned above, we’d enjoy hearing your ideas here. We’re hitting this in the following order: test 15 sample segments (done), create a batch process to observe weather, train and test segment performance predictions on top 25% of starred segments (50% done) followed by creating a user experience exposing a rough version of the feature (not started). I’ll follow up with a post in the coming weeks highlighting segments that we’ve been able to achieve a high level of accuracy in our predictions and a mock up of the experience storyboard!
