End Date: December 24, 2020 11:59 PM UTC (scroll down for prizes and details)

Overview
In the spirit of the confrontational season that is Thanksgiving, we challenge you to develop a model that can predict whether a box contains a bobcat based on some of our project’s anonymized data!

We want to improve our bobcat detection algorithm, in order to deliver more bobcats to our subscribers. Too many of the boxes we send contain office chairs instead of bobcats, and as you can imagine this is entirely unacceptable.

The contest dataset includes 233,674 boxes, 7.5% of whom contain a bobcat! Each box is richly described with data across three tables that give you more than enough information to estimate which boxes have bobcats in them.

Take a look at our data and join us to make the world a weirder place.

Click here to check out the contest data on GitHub and see what all the fuss is about!

Click here to join our Discord channel and discuss the contest!

Prizes
The first qualifying submission will get an actual, live and ferocious bobcat (pictured below).

The first 50 qualifying submissions will each get $30 credit for surprise packages using our shopping algorithm! (per individual and unique submission) Set keywords to get interesting products delivered to you or friends & family!

The first place winner of the contest will get $500 in surprise packages from Bobcat In A Box AND a choice of one of five fabulous trophies…

–BioVYZR, an outer layer that protects against airborne hazards in our daily environments. With a built in purifier, this is a step above face masks and N95 respirators. Impress your friends and flatten the curve.
-Celebrate your conquest of the data with a life-size Thor’s hammer. Great for cosplay and making guacamole.
-Take a giant golden T-Rex head as your trophy and mount it prominently on your wall. No one will question your data hacking prowess.
-This holiday season, sport an adorable baby Yoda sweater! “Cute I am, Adore me you must”
-On your next hike, sleep off the ground and safely out of reach of prowling bobcats with a lightweight and durable hammock.

Evaluation

Click here to download the contest data on GitHub

To make a submission please send tbl_test_boxes.csv with your best estimates of target_bobcat, the code you used to generate the result and some documentation of what you did to data@bobcatinabox.com

The column target_bobcat should be a decimal value between 0 and 1 that reflects your model’s confidence in whether or not a box contains a bobcat (0 means you don’t think there is any chance of a bobcat, 1 means that you are certain that a box contains a bobcat).

To evaluate your submission, we’ll rank the boxes using your target_bobcat and take a look at how many bobcats there actually were in the predicted top 5% of tbl_test_boxes. The winning submission will be the one that has the largest number of actual bobcats in the 5% of boxes with the highest target_bobcat values in tbl_test_boxes.

To count as a qualifying submission, your predicted top 5% of tbl_test_boxes must contain at least 15% actual bobcats. Across the whole dataset 7.5% of boxes contain bobcats. Please do not submit a large number of high and identical target_bobcat values as this may gum up the works when it comes time to evaluate the top 5%.

Hints: As we are only looking at the top 5%, false positives are a larger problem than false negatives. Bobcats are inherently unpredictable creatures. If in cross validation your model is correctly predicting over 45% of the target_bobcat vales then it has overfitted to the training data and you are doing something wrong.

About Bobcat In A Box
We run a shopping algorithm that scrapes popular websites like Amazon, Etsy and AliExpress for products to buy. Subscribers to the algorithm get surprise packages in the mail based on keywords that they set. Once they have received a package, our subscribers are given the option of ‘liking’ the products that they get on our internal customer dashboard to train the algorithm.

The script and this project was inspired by the below XKCD comic, although it is in no was affiliated with Randall Munroe.