Lending Club Loan Data Analysis
Jared Lee
Summary
Lending Club is a peer-to-peer lending company. The purpose of this analysis is to look at loan data from Lending Club to see what factors influence the likelihood of a loan being paid in full. Additionally, this analysis will be used to develop a model to predict the probability of a given current loan defaulting or being charged off.
Preparations
This first chunk loads the necessary R packages.
library(tidymodels)
library(tidyverse)Analysis
Data
The Lending Club data was retrieved May 31, 2020 from Kaggle. These files contain complete loan data for all loans issued through the 2007-2015, including the current loan status (Current, Late, Fully Paid, etc.) and latest payment information. The file containing loan data through the “present” contains complete loan data for all loans issued through the previous completed calendar quarter. Additionally, population data was collected from the U. S. Census to calculate per capita values.
# Read the files
loans <- read_csv("loan.csv")
popDF <- read_csv("nst-est2019-alldata.csv")
# View Structure of Data
str(loans)
head(loans)# Summary of some key character columns
summary(as.factor(loans$term))## 36 months 60 months
## 1609754 650914summary(as.factor(loans$grade))## A B C D E F G
## 433027 663557 650053 324424 135639 41800 12168summary(as.factor(loans$sub_grade))## A1 A2 A3 A4 A5 B1 B2 B3 B4 B5 C1
## 86790 69562 73184 95874 107617 125341 126621 131514 139793 140288 145903
## C2 C3 C4 C5 D1 D2 D3 D4 D5 E1 E2
## 131116 129193 127115 116726 81787 72899 64819 56896 48023 33573 29924
## E3 E4 E5 F1 F2 F3 F4 F5 G1 G2 G3
## 26708 22763 22671 13413 9305 7791 6124 5167 4106 2688 2094
## G4 G5
## 1712 1568summary(as.factor(loans$emp_length))## < 1 year 1 year 10+ years 2 years 3 years 4 years 5 years 6 years
## 189988 148403 748005 203677 180753 136605 139698 102628
## 7 years 8 years 9 years n/a
## 92695 91914 79395 146907summary(as.factor(loans$home_ownership))## ANY MORTGAGE NONE OTHER OWN RENT
## 996 1111450 54 182 253057 894929summary(as.factor(loans$purpose))## car credit_card debt_consolidation educational
## 24013 516971 1277877 424
## home_improvement house major_purchase medical
## 150457 14136 50445 27488
## moving other renewable_energy small_business
## 15403 139440 1445 24689
## vacation wedding
## 15525 2355# Loan status is the most important column
summary(as.factor(loans$loan_status))## Charged Off
## 261655
## Current
## 919695
## Default
## 31
## Does not meet the credit policy. Status:Charged Off
## 761
## Does not meet the credit policy. Status:Fully Paid
## 1988
## Fully Paid
## 1041952
## In Grace Period
## 8952
## Late (16-30 days)
## 3737
## Late (31-120 days)
## 21897Preprocessing
The data appears pretty clean and the summarized variables do not need to be cleaned further. The sub_grade column appears to be a more granular view of the grade column and so only one of them should be used in any modeling. The loan_status column does need to be cleaned up, though, and will be split into three groups: Good Loans, Bad Loans, and Current Loans. Good loans will be the loans that have been fully paid. Bad loans will be loans that were charged off, defaulted or more than a month late on the latest payment. Current Loans will be loans that are currently outstanding, in the grace period, or less than a month late on the latest payment. For the purpose of this analysis, current loans will be ignored.
# To simplify, loan status will be grouped into 3 categories: Current, Good, Bad
good_loans <- c("Does not meet the credit policy. Status:Fully Paid","Fully Paid")
bad_loans <- c("Charged Off","Does not meet the credit policy. Status:Charged Off","Default","Late (31-120 days)")
current_loans <- c("Current","In Grace Period","Late (16-30 days)")
# Create a new outcome column based on loan_status
loans <- loans %>%
mutate(outcome = case_when(loan_status %in% good_loans ~ "Good Loan",
loan_status %in% bad_loans ~ "Bad Loan",
loan_status %in% current_loans ~ "Current",
TRUE ~ "Other"))
# Make Sure that all loan_status are converted (no "Other"s)
summary(as.factor(loans$outcome))## Bad Loan Current Good Loan
## 284344 932384 1043940# Filter out current loans
finished_loans <- loans %>%
filter(outcome != "Current")Exploratory Analysis
Several plots will be created to explore how different variables affect the quality of a loan. To minimize the amount of code and keep a consistent style, this function generates plots based on a column input. It creates a density plot for numeric columns and two bar plots for catagorical columns, one with the count and one with the proportion of the quality of the loan. It is important to remember density plots do not show the actual counts. Because there are significantly more good loans in the data set, the density plots can make it appear like there are more bad loans than in reality. A bad loan peak that is higher than a good loan peak does not mean that there are more bad loans at that value.
plotData <- function(data, group, label = group) {
colors <- c("#D33016","#1BA9C2") #Bad,Good
if(is.numeric(data[[group]])) {
data %>%
ggplot(aes(x = .data[[group[[1]]]],fill = outcome)) +
geom_density(alpha = 0.5) +
scale_fill_manual(values = colors) +
labs(x = label,y = NULL,fill = NULL) +
theme_minimal()
} else {
data %>%
group_by(.data[[group[[1]]]],outcome) %>%
summarize(Count = n()) %>%
mutate(Proportion = Count/sum(Count)) %>%
pivot_longer(Count:Proportion) %>%
ggplot(aes(x = .data[[group[[1]]]],y = value,fill = outcome))+
geom_col()+
labs(x = label,y = NULL,fill = NULL) +
scale_fill_manual(values = colors) +
facet_wrap(~name,scales = "free_x") +
coord_flip() +
theme_minimal()
}
}plotData(finished_loans,"int_rate",label = "Interest Rate")
This density plot of the interest rate clearly shows that most of the loans have an interest rate of 20% or less. There is also a split between the quality of the loans: Good loans tend to have lower interest rates than bad loans. Most of the good loans have interest rates less than 15% with the most common interest rate being slightly above 5% and about 12%. Bad loans have interest rates that are typically between 10% and 20% with a most common rate of about 14%. Loans with interest rates higher than 20% are rare and more likely to be bad loans.
Key Takeaways
Good loans tend to have lower interest rates than bad loans.
Interest rates are typically less than 15% for good loans and between 10% and 20% for bad loans.
plotData(finished_loans,"loan_amnt",label = "Loan Amount")
This is a similar density plot of the quality of loans by the size of the loan. There are clear peaks at round numbers, which is unsurprising as a loan of $15,000 is much more likely to occur than a loan of $12,345. The most common loan is for $10,000, and amounts less than that are slightly more likely to be a good loan. Loans greater than $10,000 may be more likely to be a bad loan but the ratio of good loans to bad loans does not appear to change significantly above that amount.
Key Takeaways
The most common loan is for $10,000.
The ratio of good loans to bad loans is similar for loans more than %12,000.
plotData(finished_loans,"installment",label = "Monthly Payment [$]")
This density plot shows the distributions of good and bad loans against the size of the monthly payments on the loans. Both distributions look very similar to each other with most payments under $500. The most common payment for either quality loans was around $300.
Key Takeaways
Most payments are under $500.
The quality of the loan does not appear to relate to the size of the monthly installment.
plotData(finished_loans,"term",label = "Loan Length")
The term length for Lending Club loans is either 3 or 5 years. The vast majority of loans are for 3 years and that is the length suggest by Lending Club. When a loan has a term length of 5 years, it is a bad loan approximately 35% of the time. This is much higher than the approximately 11% of 3 year loans that are bad.
Key Takeaways
The majority of Lending Club loans are for 36 months or 3 years.
60 month or 5 year loans have a higher proportion of bad loans.
The grade of the loan is calculated by Lending Club based on several factors, the largest of which is the applicants credit score. The Lending Club data does not include the applicants credit score so the grade will be used as a proxy. The grade of most of the loans are B or C and the distribution is slightly skewed toward lower grades.There is a very clear and strong relationship between the grade and the quality of a loan. Loans with higher grades are more likely to be good loans, and loans with lower grades are more likely to be bad loans. However, There are a low number of F and G loans and so the expected proportion of the quality of the loans may have more uncertainty. Even ignoring F and G loans, the relationship between grade and quality of loans is still very clear and strong.
Key Takeaways
Grade is a proxy for the applicant’s credit score.
The grade distribution is skewed toward lower scores with the most loans occurring at grades of B or C.
Loans with higher grades are more likely to be good loans.
finished_loans %>%
mutate(emp_length = fct_relevel(emp_length,"10+ years",after = 10)) %>%
mutate(emp_length = fct_relevel(emp_length,"n/a")) %>%
plotData("emp_length",label = "Years Employed")
The plurality of completed loans are given to people who have been employed for 10 or more years. As the length of time employed increased there is a very slight increase to the quality of the loans. When employment length data is not available, the quality of the loan is much lower.
Key Takeaways
Most loan recipients have been employed for 10 or more years.
There is a small positive relationship between the number of years employed and the quality of the loan.
This bar plot shows the count and proportion of the quality of the loan split by the type of home ownership. The “other,” “none,” and “any” levels of home ownership are negligible and can be safely ignored in analysis and modeling. Almost all loans went to people who rented, had a mortgage, or owned their own home, although the number of people who owned their own home is far fewer than the number in either of the other two categories. Renters repaid their loans approximately 75% of the time compared to about 78% of the time for home owners and about 81.5% of the time for people with a mortgage.
Key Takeaways
The vast majority of loans went to people renting or who had a mortgage.
People with a mortgage were more likely to fully pay their loan than people who rented.
finished_loans %>%
mutate(purpose = str_to_title(str_replace(purpose,"_"," "))) %>%
plotData("purpose",label = "Purpose of Loan")## `summarise()` regrouping output by 'purpose' (override with `.groups` argument)
This plot also shows the count and proportion of good and bad loans broken down by the stated purpose of the loan. The majority of loans were given for the purpose of debt consolidation or paying off a credit card, although most were for debt consolidation. The debt consolidation loans had a lower proportion of good loans than loans for credit cards. Small business loans are typically the lowest quality as they have the highest proportion of bad loans, however, small business loans make up only 1.2% of all loans.
Key Takeaways
Most loans were given for debt consolidation (58%).
Loans to pay off a credit card were typically higher quality than loans for debt consolidation.
map <- map_data("state")
state_link <-tibble(abbr = state.abb,name = tolower(state.name))
state_centers <- bind_cols(as_tibble(state.center),state_link)
popDF <- popDF %>%
transmute(name = tolower(NAME),pop = POPESTIMATE2019)
state_counts<- finished_loans %>%
count(addr_state, outcome) %>%
group_by(addr_state) %>%
mutate(prop = round(n/sum(n)*100,1),
n_tot = sum(n)) %>%
filter(outcome == "Good Loan")
midpoint <- (max(state_counts$prop) + min(state_counts$prop)) / 2
labelDF <- state_counts %>%
left_join(state_centers,by = c("addr_state" = "abbr"))
state_counts %>%
left_join(state_link,by = c("addr_state" = "abbr")) %>%
right_join(map,by = c("name" = "region")) %>%
ggplot(aes(x = long,y = lat,group = group,fill = prop)) +
geom_polygon(color = "white") +
geom_text(data = labelDF,aes(x = x,y = y,group = NULL,label = paste0(prop,"%"))) +
scale_fill_gradient2(low = "red",mid = "gray50",high = "blue",midpoint = midpoint,
labels = function(x){paste0(x,"%")},
n.breaks = 4,name = 'Percentage of \n"Good Loans"') +
theme_void() +
coord_fixed(ratio = 1.3)
This map shows the percentage of good loans split up by state. States colored red have worse quality loans than states colored blue. States in the Southeast region of the country, with the exception of Georgia, tend to have a slightly lower percentage of good loans. This is especially true in Mississippi, which has the lowest percentage of good loans out of every state at only 71.5%. Additionally Nebraska had the second worst percentage at 72.6%. The combined Mountain and Pacific Northwest regions (including states like Washington, Montana, Colorado, and Utah) along with New England tended to have the highest percentage of good loans, with most states in these regions having percentages above 80%. Vermont, Oregon, and Maine were the best states with percentages of 84.7%, 84.5%, and 84.4% respectively.
Key Takeaways
The Southeast had the lowest quality loans, with Mississippi being the worst state in the country at 71.5% of loans paid back.
New England had the highest quality loans, with Vermont being the best state in the country at 84.7% of loans paid back.
state_counts_2 <- state_counts %>%
left_join(state_link,by = c("addr_state" = "abbr")) %>%
left_join(popDF, by = "name") %>%
ungroup() %>%
mutate(perCap = (n_tot/pop)*100000)
labelDF_2 <- labelDF %>% left_join(state_counts_2,by = "addr_state")
state_counts_2 %>%
mutate(perCap = ifelse(perCap<1,NA,perCap)) %>%
right_join(map,by = c("name" = "region")) %>%
ggplot(aes(x = long,y = lat,group = group,fill = perCap)) +
geom_polygon(color = "white") +
geom_text(data = labelDF_2,aes(x = x,y = y,group = NULL,fill= NULL,label = round(perCap,0))) +
scale_fill_gradient(low = muted("red"),high = "blue",
labels = function(x){paste0(x)},
n.breaks = 4,name = 'Loans Per \n100,000 people') +
theme_void() +
coord_fixed(ratio = 1.3)
This map shows the per capita amount of loans for each state using population estimates from the U.S. Census for 2019. Lending club loans are unavailable for people living in Iowa and so there is no data for the state. There doesn’t appear to be any type of regional difference in the amount of loans that are taken out. Nevada does have the most loans (649) by a large margin, almost 100 more loans per 100,000 people than the second highest state, New York (559). Considering the state is known for gambling and Las Vegas, and that the purpose for most loans is for debt consolidation, it seems reasonable to conclude that the two are related. However, further analysis would be needed to support that relationship. Idaho has the fewest amount of loans (89) of any of the states.
Key Takeaways
There does not appear to be significant regional differences in the amount of loans taken out.
Nevada has the most loans per capita at 649 loans per 100,000 people.
library(plotly)##
## Attaching package: 'plotly'## The following object is masked from 'package:ggplot2':
##
## last_plot## The following object is masked from 'package:stats':
##
## filter## The following object is masked from 'package:graphics':
##
## layoutg <- list(
scope = 'usa',
projection = list(type = 'albers usa'),
showlakes = TRUE,
lakecolor = toRGB('white')
)
finished_loans %>%
count(addr_state,outcome) %>%
group_by(addr_state) %>%
mutate(prop = round(n/sum(n)*100,2),
n_tot = sum(n)) %>%
filter(outcome == "Bad Loan") %>%
plot_geo(locationmode = 'USA-states') %>%
add_trace(locations= ~addr_state,color = ~prop, z= ~prop,hoverinfo = "text",
text= ~paste0("Number of Loans: ",n_tot,"\nGood Loans: ",100-prop,"%\nBad Loans: ",prop,"%"),
colorscale = "RdBu") %>%
colorbar(title = "") %>%
layout(
title = 'Percentage of Loans that are "Bad"',
geo = g
)## Warning: `arrange_()` is deprecated as of dplyr 0.7.0.
## Please use `arrange()` instead.
## See vignette('programming') for more help
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_warnings()` to see where this warning was generated.Finally a nice interactive map to explore the state data.
Modeling
This analysis will include two different types of machine learning. An unsupervised machine learning algorithm will be used to segment the loans into a number of different clusters to try and find additional patterns in the large dataset. Additionally, a supervised machine learning classification model will be constructed to predict the likelihood of a loan being paid back.
Model Preperation
Most machine learning algorithms require a matrix of numbers to run. The Lending Club dataset, however, includes many features that are categorical. This code will turn those catagorical features into dummy variables of ones and zeroes as well as scale and center the numeric features to improve the speed and performance of the algorithms.