The project solved the problem of predicting risk of client's default, based on credit history data

Used data from https://www.kaggle.com/datasets/mikhailkostin/dlfintechbki

3 prediction models were tested: LightGBM, HistGradientBoosting and a neural network implemented on Pytorch

Final accuracy of the best model on the test sample: 0.7625 (roc-auc)

/data - project data (download from source)

• test_predict.csv - predicted values for test_data

/jupyter - jupyter-notebook files:

• collect - data loading and aggregation
• modeling - model selection (LightGBM, HistGradient)
• modeling_mlp - Pytorch neural network
• pipeline - creating a pipelines with the final model
• pipeline_test - test predict on the pipeline
• pipe_utils.py - utilities and classes for work of the pipelines
• features.pkl - list of final features of the model

/model - dumped model files

Final pipeline model/pipe_11.pkl
Import all the functions from jupyter/pipe_utils.py
To deserialize the pipe_11 model use dill (see example in pipeline_test.ipynb)