FAQs

Get the answer to your question below. Please let me know if you're missing a question 🤠

I wanted to somehow combine two of my main interests (climbing and data) and this website is the result of it.
I got some inspiration from the 9b counter introduced by EpicTV.
It might look pretty trivial at first sight, but I hope after reading through these FAQs you'll get a better understanding of the underlying techniques and technologies. Although I definitely do not claim that the predicted result is not predictable by a normal human being with some climbing knowledge 😉
Probably most frequently asked question! The 2 main reasons are:
  1. Your 8a.nu profile has been created after 12 Sep 2017;
  2. Your 8a.nu profile has been marked as suspicious, incorrect or irrelevant during the data cleaning stage (see tab DATA)
Good question! In order to create an accurate predictive model "more data" is usually preferred (under the assumption of good data quality, variety, etc.). As there are not that many people climbing 9a/9b compared to 8a I decided to go for the 8a-predictor.
There are the top 3 male and female climbers according to our model including the highest ranked Dutchies 💪

# Name (male) Country
1 Nicolai Užnik Austria
2 Tanguy Merard France
3 Loris Veneault France
5 Mark Brand Netherlands
# Name (female) Country
1 Magdalena Trzemzalska Poland
2 Margo Hayes United States
3 Giorgia Tesio Italy
20 Mirthe van Liere Netherlands
Haha, my favourite question. Although I'm a big fan of Magnus, unfortunately he hasn't reported any climbs softer than 8a on his profile 😅
  • Layman's explanation: according to Adam's 8a.nu profile, he didn't climb any boulders before climbing his first 8a whereas Margo did report boulder ascents before sending her first 8a.
  • Technical explanation: see tab DATA
No data, no model, no website. A big shout-out to David Cohen for making this dataset available on Kaggle 🙏
After cleaning the dataset, I ended up with information about ≈26k climbers.
I tried to remove as much suspicious and incorrect data as possible in order to reduce the noise to a minimum. For example, some profiles reporting ascents before they actually started climbing or they claim to have climbed very hard non-existing routes/boulders. I also looked at reasonable bounds for things like age, height, BMI, etc. In additional, all 8a.nu profiles reporting solely 8a (or higher) ascents have been marked irrelevant for the purpose of this project.
No, it does not! Once I trained the model, all 8a.nu profiles have been scored and the results stored in a database. Afterwards, all 8a.nu related data has been removed.
When plugging Adam's 8a.nu profile into the model, the feature with the most negative impact on his likelihood of climbing an 8a sport route, was his bouldering grade. Any significant boulder ascents would have increased his likelihood, approaching the likelihood of Margo.
I generated around 60 features in total before performing feature selection and reducing this number to 7.
Gradient Boosting Machine (GBM) as implemented by H2O 💚
Pretty accucate when looking at performance metrics. For example, the AUC and PR-AUC on the test set are ≈0.99 and ≈0.95, respectively.
All raw data predictions are real-time. Predicitons based on 8a.nu profiles have been pregenerated and stored in a database.

Go to home page