Data Science + Running: PREDICTING BETTER THAN THE GIANTS GARMIN & RACEX

I love running & data science. The week before my last marathon, I was feeling a bit lost about what my race pace should have been, given that I'm a beginner runner.

In the midst of excitement and anxiety for the race, I selected 7 representative runs I've done, put their data together and thoughtfully created models to predict my average race pace. The best models using a cross-validation criteria were consistently Neural Networks & Decision Trees, but they were giving me faster times than what both my Garmin and RaceX were telling me. 

It turns out that my two best models were 99% accurate ✨ , beating Garmin's predictions by ~5min, even though Garmin has super detailed data for +500 runs I've made in my life, in addition to sleep data, etc, plus a huge team of excellent sports data scientists. 

Why Neural Networks & Decision Trees? I had little data (N = 106 miles, from 7 runs) and, among the models that I ran, these were flexible enough to capture all of the data non-linearities (as opposed to restrictive linear regression or k-means), without having a complex data greedy architecture (as opposed to Random Forest or Gradient Boosting, the neural net had a relatively simple architecture). 

Data structure: N = 106 miles, from 7 runs, with carefully selected features:

• Dummy for weather temperature: hot or not 🔥

• Dummy for race or no race 🏁

• How long ago each run was (in days) 📅

• Average heart rate for the mile ❤️

• Average pace for the mile 🏃

• Average elevation gain for the mile ⛰️

• Total amount of miles in the run  📏

• Which mile was it in the run 🔢

Some social media screenshots about this: