Credit Risk Scoring (Bank loan)

• Normalized the column names of the dataframe.
• Checked for null and unique values in each column.
• As the complete data was given integers, we needed to change the categorical features to string type so that we can apply one hot encoding.
• We checked the statistics of the numerical data, where we observed 99999999.0 as max values which actually were the white spaces and not null.
• Replaced them with np.nan.
• Dropped one row where we didn't knew if the client payed the loan or not.
• Splitted the dataframe into training, validation and test datasets.
• As we replaced some vlaues with np.nan we had to use fillna(0) to make further progress.
• To apply one hot encoding we needed dictionary of categotical features.
• Thus converted dataset into dictionary using DictVectorizer.
• Trained the model on DecisionTreeClassifier.
• Checked roc_auc_score for training dataset, score = 1.0
• Checked roc_auc_score for validation dataset, score = 0.64
• This shows that our model overfitted.
• Applied mx_depth and min_sample_leaf features implied increase in score.
• Aplied RandomForestClassifier to see if we can make for a better score.
• After hyperparameter tuning of max_depth, n_estimators and min_sample_leaf we got the score of 82.3%
• To further increase score we tried if gradient boosting using xgboost can improve the score or not.
• After hyperparameter tuning of eta, max_depth, min_child_weight, objective,eval_metric, nthread, seed, silent, colsample_bytree we got the score of 83.8%
• Refer to the .ipynb file.