KDnuggets Home » News » 2018 » Sep » Tutorials, Overviews » Financial Data Analysis – Data Processing 1: Loan Eligibility Prediction ( 18:n33 )

Financial Data Analysis – Data Processing 1: Loan Eligibility Prediction


In this first part I show how to clean and remove unnecessary features. Data processing is very time-consuming, but better data would produce a better model.



By Sabber Ahamed, Computational Geophysicist and Machine Learning Enthusiast

 

Introduction

 
Financial institutions/companies have been using predictive analytics for quite a long time. Recently, due to the availability of computational resources and tremendous research in machine learning made it possible to better data analysis hence better prediction. In the series of articles, I explain how to create a predictive loan model that identifies a bad applicant who is more likely to be charged off. In step by step processes, I show how to process raw data, clean unnecessary part of it, select relevant features, perform exploratory data analysis, and finally build a model.

As an example, I use Lending club loan data dataset. Lending Club is the world’s largest online marketplace connecting borrowers and investors. An inevitable outcome of lending is default by borrowers. The idea of this tutorial is to create a predictive model that identifies applicants who are relatively risky for a loan. In order to accomplish this, I organized the whole series into four parts as follows:

  • Data processing-1: In this first part I show how to clean and remove unnecessary features. Data processing is very time-consuming, but better data would produce a better model. Therefore, careful and very detail examination is required to prepare better data. I show how to identify constant features, duplicate feature, duplicate rows, and features with a high number of missing values.
  • Data processing-2: In this part, I manually go through each and every features selected from part -1. This is the most time-consuming part, but worth it for a better model.
  • EDA: In in this part, I do some exploratory data analysis (EDA) on the features selected in part-1 and 2. A good EDA is required to get a better knowledge of the domain. We need to spend some quality time to find out the relations between the features.
  • Create a model: Finally, In this last but not the last part, I create models. Creating a model is also not an easy task. It’s also an iterative process. I show how to start with a with a simple model, then slowly add complexity for better performance.

Alright, let’s get started with the part-1: data processing, cleaning and feature selections.

 

Data processing-1

 

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import warnings
warnings.filterwarnings("ignore")


In this project, I used three years of datasets (2014, 2015 and 2017(first-thrid quarter)) and stored in five separate CSV files. Lets read the files first:

df1 = pd.read_csv(‘./data/2017Q1.csv’, skiprows=[0])
df2 = pd.read_csv(‘./data/2017Q2.csv’, skiprows=[0])
df3 = pd.read_csv(‘./data/2017Q3.csv’, skiprows=[0])
df4 = pd.read_csv(‘./data/2014.csv’, skiprows=[0])
df5 = pd.read_csv(‘./data/2015.csv’, skiprows=[0])


Since data are stored in separate files, we have to make sure that we have the same number of features in each file. We can check using the following code snippet:

columns = np.dstack((list(df1.columns), list(df2.columns), list(df3.columns), list(df4.columns), list(df5.columns)))
coldf = pd.DataFrame(columns[0])


The above code is self-explanatory, we first extract the column names the stack them together using Numpy ‘dstack’ object. If you look at the Jupyter-notebook on Github, you would see they are same. Which is good for us. We can move on to the next step. It’s time to check the shape of the data:

df = pd.concat([df1, df2, df3, df4, df5])
df.shape


(981665, 151)


We see that there are approximately one million examples and each of the examples has 151 features including target variable. Let’s look at the feature name to get familiar with the data. It’s imperative to get to know the domain, especially the details of the features relationship with the target variable. It’s not easy to learn overnight, that’s why need to spend some days or maybe a week to get familiar with the data before jumping into further detail analysis. Let’s see the feature names:

print(list(df.columns))


['id', 'member_id', 'loan_amnt', 'funded_amnt', 'funded_amnt_inv', 'term', 'int_rate', 'installment', 'grade', 'sub_grade', 'emp_title', 'emp_length', 'home_ownership', 'annual_inc', 'verification_status', 'issue_d', 'loan_status', 'pymnt_plan', 'url', 'desc', 'purpose', 'title', 'zip_code', 'addr_state', 'dti', 'delinq_2yrs', 'earliest_cr_line', 'fico_range_low', 'fico_range_high', 'inq_last_6mths', 'mths_since_last_delinq', 'mths_since_last_record', 'open_acc', 'pub_rec', 'revol_bal', 'revol_util', 'total_acc', 'initial_list_status', 'out_prncp', 'out_prncp_inv', 'total_pymnt', 'total_pymnt_inv', 'total_rec_prncp', 'total_rec_int', 'total_rec_late_fee', 'recoveries', 'collection_recovery_fee', 'last_pymnt_d', 'last_pymnt_amnt', 'next_pymnt_d', 'last_credit_pull_d', 'last_fico_range_high', 'last_fico_range_low', 'collections_12_mths_ex_med', 'mths_since_last_major_derog', 'policy_code', 'application_type', 'annual_inc_joint', 'dti_joint', 'verification_status_joint', 'acc_now_delinq', 'tot_coll_amt', 'tot_cur_bal', 'open_acc_6m', 'open_act_il', 'open_il_12m', 'open_il_24m', 'mths_since_rcnt_il', 'total_bal_il', 'il_util', 'open_rv_12m', 'open_rv_24m', 'max_bal_bc', 'all_util', 'total_rev_hi_lim', 'inq_fi', 'total_cu_tl', 'inq_last_12m', 'acc_open_past_24mths', 'avg_cur_bal', 'bc_open_to_buy', 'bc_util', 'chargeoff_within_12_mths', 'delinq_amnt', 'mo_sin_old_il_acct', 'mo_sin_old_rev_tl_op', 'mo_sin_rcnt_rev_tl_op', 'mo_sin_rcnt_tl', 'mort_acc', 'mths_since_recent_bc', 'mths_since_recent_bc_dlq', 'mths_since_recent_inq', 'mths_since_recent_revol_delinq', 'num_accts_ever_120_pd', 'num_actv_bc_tl', 'num_actv_rev_tl', 'num_bc_sats', 'num_bc_tl', 'num_il_tl', 'num_op_rev_tl', 'num_rev_accts', 'num_rev_tl_bal_gt_0', 'num_sats', 'num_tl_120dpd_2m', 'num_tl_30dpd', 'num_tl_90g_dpd_24m', 'num_tl_op_past_12m', 'pct_tl_nvr_dlq', 'percent_bc_gt_75', 'pub_rec_bankruptcies', 'tax_liens', 'tot_hi_cred_lim', 'total_bal_ex_mort', 'total_bc_limit', 'total_il_high_credit_limit', 'revol_bal_joint', 'sec_app_fico_range_low', 'sec_app_fico_range_high', 'sec_app_earliest_cr_line', 'sec_app_inq_last_6mths', 'sec_app_mort_acc', 'sec_app_open_acc', 'sec_app_revol_util', 'sec_app_open_act_il', 'sec_app_num_rev_accts', 'sec_app_chargeoff_within_12_mths', 'sec_app_collections_12_mths_ex_med', 'sec_app_mths_since_last_major_derog', 'hardship_flag', 'hardship_type', 'hardship_reason', 'hardship_status', 'deferral_term', 'hardship_amount', 'hardship_start_date', 'hardship_end_date', 'payment_plan_start_date', 'hardship_length', 'hardship_dpd', 'hardship_loan_status', 'orig_projected_additional_accrued_interest', 'hardship_payoff_balance_amount', 'hardship_last_payment_amount', 'disbursement_method', 'debt_settlement_flag', 'debt_settlement_flag_date', 'settlement_status', 'settlement_date', 'settlement_amount', 'settlement_percentage', 'settlement_term']


Looking at the above features, it may seem scary first. But we will get through every feature and then select the relevant features. Let's start with the target feature “loan_status”

df.loan_status.value_counts()


Current               500937
Fully Paid            358629
Charged Off            99099
Late (31-120 days)     13203
In Grace Period         6337
Late (16-30 days)       3414
Default                   36
Name: loan_status, dtype: int64


We see that there are seven types of loan status. However, in this tutorial, we are interested in two classes: 1) Fully paid: those who paid the loan with interests and 2) Charged off: those who could not pay and finally charged off. Therefore, we select the data sets for these two classes:

df = df.loc[(df['loan_status'].isin(['Fully Paid', 'Charged Off']))]


df.shape
(457728, 151)


Looking at the shape, we see that we now have half of the data point than original data and the same number of features. Before processing and cleaning manually, let’s do some general data processing steps first:

  • Remove features associated with >85% missing values
  • Remove constant features
  • Remove duplicates features
  • Remove duplicate rows
  • Remove highly collinear features (In part 3 EDA)

Alright, let’s get started with the typical data processing:

1. Remove features associated with 90% missing values: In the code below I first use pandas’ built-in method ‘isnull()’ to find the rows associated with missing values. Then I sum them up to get the count for each feature. Finally, I sort the features according to the number of missing values and create a data frame for further analysis.

In the above result, we see that there are 53 features which have 400000 missing values. I use the pandas’ drop method to remove these 53 features. Notice that in this function I set the “inplace” option to True”, which removes these features from original data frame df without returning anything.

2. Remove constant features: At this step, we remove features that have a single unique value. A feature associated with one unique value does not help the model to generalize well since it’s variance is zero. A tree-based model cannot take advantage of these type of features since the model can not split these features. To identify features with a single unique value is relatively straightforward:

In the above code, I create a function “find_constant_features” to identify constant features. The function goes through each feature and sees if it has less than two unique values. If so, the features are added to the constant feature list. We can also find out constant feature looking at the variance or standard deviation. If the feature has zero variance or standard deviation, we are sure that the feature has single unique value. The print statement shows that five features have single unique value. So we remove them using “inplace” option true.

3. Remove duplicate features: Duplicate features are those have the same value in multiple features with the same/different name. To find out the duplicate features I borrowed the following code from this stack overflow link:

We see only one feature which seems to be duplicated. I am not going to remove the feature yet rather wait until we do EDA in the next part.

4. Remove duplicate rows: In this step, we remove all the duplicate rows. I use pandas built-in “drop_duplicates(inplace= True)” method to perform this action:

df.drop_duplicates(inplace= True)


The above four processings are basic which we need to do for any data science project. Let's see the shape of the data after all of these steps:

df.shape

(457728, 93)


We see that we have 93 features after performing the above steps.

In the next part of this tutorial, I will go through each feature, then perform cleaning and remove it if necessary. In the meantime, if you have any question regarding this part, please feel free to write your comment below. You can reach out to me:


 
Bio: Sabber Ahamed is the Founder of xoolooloo.com. Computational Geophysicist and Machine Learning Enthusiast.

Original. Reposted with permission.

Related:


Sign Up