The Nubank Puzzle is a short modeling assignment that is typical of some of the work we do at Nubank. It should only take you 1-2 hours, but feel free to spend more time on it if you wish.
We host a ZIP archive here (https://s3-sa-east-1.amazonaws.com/nu-static/workable-data-science/data-science-puzzle.zip) that contains 3 CSV files: train.csv, test.csv, and sample_submission.csv. train.csv contains a column called id which uniquely identifies each row, several columns identified by hexadecimal strings, and a column target which we would like for you to predict. Columns that contain SHA-256 hashes for their values represent categorical variables, while the rest of the variables are numeric. test.csv has the same column names and data types as train.csv, but it is missing the target column. There are no missing values or data corruption problems in either of these files. Do not worry about the meanings of the variables or the metadata -- this is an artificial dataset.
You can use any programming language and modeling approach you choose. You will be evaluated primarily on the R2 of your predictions against the true values in the test set, but the accuracy of your estimate of your R2 will also be considered. You can find an explanation of the R2 metric here (https://en.wikipedia.org/wiki/Coefficient_of_determination).
Please submit your predictions as a CSV file. You should include a header with columns as below and put the id and your predicted target for each corresponding row from the test.csv file. sample_submission.csv gives an example of what your submission file should look like.
id,prediction
Please submit your prediction CSV file and source code we can use to reproduce your submission in a ZIP file. Please put your prediction CSV file in the root of the archive. The only file in the archive with a .csv file extension should be your predictions.
To submit your solution, please email Kevin Bird at kevin.bird@nubank.com.br with a short explanation of what you did and what you expect the R2 to be for your predictions, and attach the ZIP file with your predictions and source code. If you have any feedback or questions, feel free to email Kevin as well.
Good luck and we are excited to get to know you better ;-)
Using a basic approach with XGboost algorithm over linear regression I built a predictive solution that should have R2 value of approximately 0.365 due to the test set evaluation. The train set had a R2 value of 0.883, this higher value may indicate the poor generalization of the built model although R2 may not be the best metric to evaluate a regression algorithm.