mirror lendingclub repo from github

Justin Hsi e51f5f7291 finished fast run of catboosth both 12 hours ago
.circleci f073d4d660 trying circleci 3 months ago
.dvc ebce979db2 full dataset catboost clf 2 weeks ago
data 0e8a97c9e9 simple combined with > pred 95% non def rate 12 hours ago
lendingclub e51f5f7291 finished fast run of catboosth both 12 hours ago
notebooks e51f5f7291 finished fast run of catboosth both 12 hours ago
requirements 3b3a925421 commit all before full run for logistic regression 3 weeks ago
results 229096f48e reran on master, full run of catboost_regr 1 day ago
results_all e51f5f7291 finished fast run of catboosth both 12 hours ago
run e51f5f7291 finished fast run of catboosth both 12 hours ago
.bashrc cda27ae492 alter dockerfile to clone all branches since should be super light 3 weeks ago
.dockerignore 3b3a925421 commit all before full run for logistic regression 3 weeks ago
.gitignore f725d73e73 pulled all results into different dir 1 day ago
.travis.yml 59392f2e09 changed permissions recurisvely 5 months ago
Dockerfile bfad13bdb8 alter dockerfile and requirement.txt 6 days ago
README.md c28e7f20e8 reverse grade/subgrade mangling, verified that it is still used internally, See readme notes 1 month ago
environment.yml 3b3a925421 commit all before full run for logistic regression 3 weeks ago
requirements.txt bfad13bdb8 alter dockerfile and requirement.txt 6 days ago
setup.py 0ba9c974a7 save before reinstall ubuntu 1 month ago
test_push.txt 2d2f1c2ed8 try push again 2 weeks ago

Data Pipeline

Legend
DVC Managed File
Git Managed File
Metric
Stage File
External File

README.md

A temporary Readme

lendingclub

For data driven loan selection on lendingclub. Important packages are sklearn, pandas, numpy, pytorch, fastai.

1) Current model is RF (sklearn) + NN (pytorch). Performance was compared against picking entirely at random and picking at random within the best performing loan grade historically. 2) Investigative models are trained on old done loans and validated on newest of old done loans. 3) Models used in invest scripts are trained on all available training data.

Notes about the csvs/data

1) Even though LC only issues loasn A1-D5, they still internally have A1 - G3/5 in the loan info. I checked the interest rates and grades with the information at https://www.lendingclub.com/foliofn/rateDetail.action

Git Tags

Various git tags for navigating between datasets and dev/full datasets datav0.0.0 <- model/scorer.dataprocessingtype.raw_data_csvs? Each redownload of new data, increment rightmost #?

DVC Stuff

1) when want new raw_csvs: python lendingclub/csv_dl_archiving/01_download_LC_csvs.python but beware: https://dvc.org/doc/user-guide/update-tracked-file

Usage:

Advisable to set up an environment After cloning: in root dir (lendingclub) with setup.py, run pip install -e . properly setup account_info.py in user_creds (see example)

Run order (all scripts in lendingclub subdir): 1) python lendingclub/csv_dl_archiving/01_download_and_check_csvs.py 2) python lendingclub/csv_prepartion/02_unzip_csvs.py 3) python Before running clean_loan_info, have to cd to lendingclub/csv_preparation python setup.py build_ext. Will make a build dir in cd, copy the .so (unix) or .pyd(windows) to cd

Notes to self:

j_utils is imported and use in several scripts. See repo https://github.com/jmhsi/j_utils

To fix permissions troubles, I ended up adding jenkins and justin to each others groups (sudo usermod -a -G groupName userName) and doing chmod 664(774) on .fth and dataframes or other files as necessary.

Current jenkins setup runs in conda environment (based off https://mdyzma.github.io/2017/10/14/python-app-and-jenkins/) Considering moving to docker containers once I build the Dockerfiles?

Made symlink: ln -s /home/justin/projects to /var/lib/jenkins/projects so jenkins can run scripts like the actual projects directory

.fth to work with after initial data and eval prep:

'eval_loan_info.fth', 'scaled_pmt_hist.fth', 'base_loan_info.fth', 'str_loan_info.fth'