idoaidan/lending-club-eda

Lending Club Loans - Exploratory Data Analysis

Video Presentation

Dataset Summary

The dataset originates from Lending Club, one of the largest peer to peer lending platforms in the United States, and was sourced from Kaggle. It contains 10,000 loan records with 55 features that capture a comprehensive profile of each borrower and their loan. After column selection and data cleaning, the final working dataset contains 9,976 rows and 16 columns.

Research Question: What borrower and loan characteristics influence the interest rate a borrower receives? Specifically, this analysis explores how factors such as income level, debt-to-income ratio, homeownership status, income verification, and credit history relate to the interest rate assigned by the platform.

Target Variable: interest_rate - a continuous numerical variable representing the annual interest rate assigned to each loan, ranging from approximately 5% to 30%..

Dataset Structure

Data Fields

#	Feature	Type	Description
1	interest_rate	Float	Annual interest rate assigned to the loan (target variable, 5.31%-30.94%)
2	grade	Categorical	Risk grade assigned by the platform (A through G, where A = lowest risk)
3	annual_income	Float	Borrower's self-reported annual income in USD
4	debt_to_income	Float	Ratio of monthly debt payments to monthly income
5	emp_length	Float	Employment length in years
6	loan_amount	Integer	Total amount of the loan in USD
7	term	Integer	Loan repayment period (36 or 60 months)
8	homeownership	Categorical	Borrower's housing status (Rent, Own, Mortgage)
9	verified_income	Categorical	Whether income was verified (Not Verified, Source Verified, Verified)
10	loan_purpose	Categorical	Purpose of the loan (debt consolidation, credit card, home improvement, etc.)
11	inquiries_last_12m	Integer	Number of credit inquiries in the past 12 months
12	total_credit_limit	Integer	Total credit limit across all accounts
13	total_credit_utilized	Integer	Total credit currently utilized
14	total_debit_limit	Integer	Total debit limit across all accounts
15	accounts_opened_24m	Integer	Number of accounts opened in the past 24 months
16	account_never_delinq_percent	Float	Percentage of accounts that were never delinquent

Data Cleaning & Preparation

Feature Selection

The raw dataset contains 55 columns. To focus the analysis on the research question, I selected 16 features based on three filtering criteria:

• Excessive missing values - Columns where the majority of records were null (such as delinquency-specific fields that are naturally absent for borrowers who never missed a payment) were excluded.
• Nonsensical or redundant content - Columns with IDs, internal codes, or duplicate representations of the same information were removed.
• Irrelevant categorical data - Features that do not logically contribute to interest rate pricing (such as state or issue date) were dropped.

Missing Values

Two columns in the filtered dataset contained missing values:

• emp_length - 817 missing values (8.2%). I imputed these with the median after verifying that there was no significant difference between the median and mean for this distribution, making either a valid choice. The median was preferred as a more robust measure against potential skew.
• debt_to_income - 24 missing values (0.2%). These records were dropped. At less than 0.25% of the dataset, their removal has no impact on the statistical integrity of the analysis.

After this step, zero missing values remained across all 16 columns.

Duplicate Check

A full check for duplicate rows was performed across the entire dataset. No duplicate records were found, confirming that each row represents a unique loan.

Outlier Detection & Handling

Box plots were used to visually inspect all numeric columns for statistical outliers. While outliers were present in several features, the inspection focused on annual_income and debt_to_income, as these contained the most significant logical and technical inconsistencies. Each case was evaluated individually, distinguishing between two categories:

Technical errors (identified but already removed):

• annual_income - 23 records showed zero income. Further inspection revealed that all 23 also had missing debt_to_income (which is derived from income, so this is expected), and 22 out of 23 had missing employment data. However, all other fields in these rows were fully populated - including grade, loan amount, credit history, and homeownership. This suggests these are not empty records, but rather cases where income data was not captured or reported. All 23 rows were already removed in the missing values stage due to their missing debt_to_income. Only 1 additional row was removed solely for missing debt_to_income.

Logical extremes (capped):

• annual_income - Several records exceeded 1M USD (max 2.3M USD). While these are valid observations, they distort statistical measures such as the mean and inflate variance. Incomes above the 99th percentile were capped to reduce skew while preserving these data points in the analysis.
• debt_to_income - Values exceeding the 99th percentile were capped to contain their effect without removing the records.

Legitimate variation (kept as-is):

• Outliers in columns such as loan_amount, total_credit_limit, and inquiries_last_12m were retained. These values represent natural variation in borrower profiles and do not indicate data quality issues.

Cleaning Summary

Step	Action	Rows Affected
Feature selection	55 columns reduced to 16	-
Missing emp_length	Imputed with median	817 rows updated
Missing debt_to_income	Dropped rows (including 23 zero-income records)	24 rows removed
Income above 99th percentile	Capped	~1% of rows adjusted
Debt-to-income above 99th percentile	Capped	~1% of rows adjusted
Duplicates	None found	0 rows affected
Final dataset	Ready for analysis	9,976 rows, 16 columns

Exploratory Data Analysis (EDA)

The EDA was structured as a four-part narrative, moving from the platform's institutional risk assessment to the borrower's personal financial behavior. Rather than treating variables in isolation, this approach builds a 'data story' to uncover what truly drives loan pricing beyond the obvious impact of risk grades.

The Grade - Starting Point, Not the Answer

The relationship between grade and interest rate is expected - higher risk grades naturally carry higher rates. However, confirming this pattern is only the starting point. The real goal of this analysis is to look beyond the grade and understand what underlying borrower and loan characteristics drive the grade itself, and whether some factors influence interest rate independently of the grade.

Borrower Financial Profile

Income vs. Interest Rate: Higher income groups receive slightly lower average interest rates, ranging from 13.3% for borrowers earning under 40K to 11.2% for those earning over 150K. While the correlation is positive, it is weaker than one might expect - a borrower earning five times more than another receives only a 2 percentage point reduction in rate. This suggests that the platform treats income as a secondary signal rather than a primary pricing factor, relying more heavily on credit behavior and debt metrics to assess risk.

Income Verification vs. Interest Rate: Counter-intuitively, verified borrowers face higher average interest rates compared to non-verified borrowers. This is a classic example of selection bias - the platform does not verify income randomly. It typically requires verification from borrowers who present higher risk signals or apply for larger loan amounts. As a result, the "Verified" label does not mean "safer borrower" - it is actually a red flag for higher risk. This is an important reminder that correlation does not imply causation, and understanding the business logic behind a variable is just as important as reading its numbers.

Loan Characteristics

Loan Term vs. Interest Rate:

About 70% of loans are issued for a 36-month term, yet 60-month loans carry noticeably higher interest rates. With a correlation of 0.36, loan term is the strongest positive predictor of interest rate in the dataset. The logic is straightforward - a longer repayment period means more risk for the lender, so borrowers pay a higher rate to compensate for that added uncertainty.

Loan Purpose vs. Interest Rate:

The purpose of the loan impacts interest rate, reflecting how the platform assesses the risk behind different uses of funds. Renewable Energy and Vacation loans carry the highest rates (13.3-13.5%), while House and Credit Card loans receive the lowest (11.4%). Interestingly, Small Business loans (13.0%) fall close to the top but not at the peak as one might expect. Debt consolidation, the most common purpose on the platform (over 50% of all loans), sits at 13.0% - right in the middle of the range, representing a known and manageable level of risk that the platform has extensive experience pricing.

Credit Behavior - The Real Driver

The borrower's credit profile turns out to be the most influential factor in interest rate pricing beyond the grade itself. Two variables - credit utilization and total debit limit - show opposite but complementary patterns that together reveal what the platform truly values.

Borrowers who use a larger share of their available credit pay progressively higher rates, climbing from 11.2% for those using under 25% to 14.8% for those above 75%, where the effect plateaus. The total debit limit mirrors this from the opposite direction - borrowers with limits above 30K receive rates as low as 10.8%, while those under 5K face 14.7%.

The insight here is that these two metrics capture different dimensions of the same underlying signal. A high debit limit reflects institutional trust - other banks have already assessed this borrower and extended them significant credit. A low utilization rate reflects personal discipline - the borrower has access to credit but chooses not to overuse it. The platform rewards both, making credit behavior the strongest driver of favorable loan pricing after the grade itself.

Summary & Conclusions

The central question of this analysis - what truly influences loan pricing on the Lending Club platform - reveals a clear hierarchy that challenges conventional assumptions about how interest rates are determined.

Income, despite being the most intuitive predictor, accounts for only a 2 percentage point difference between the lowest and highest earning groups. Even income verification, which one might expect to signal trustworthiness, turns out to be a red flag - a classic case of selection bias where the platform mandates verification precisely for higher-risk applicants.

The real drivers of loan pricing operate on a fundamentally different axis: behavioral and institutional signals. Credit utilization - how much of their available credit a borrower actually uses - and total debit limit - how much credit other institutions have already entrusted them with - emerge as the most powerful predictors of interest rate after the platform's own grade. This distinction is critical: the platform does not primarily price based on who the borrower is (income), but rather on how they behave financially and how existing institutions perceive their reliability. Combined with the loan's structural characteristics (term length and purpose), these behavioral metrics paint a consistent picture - Lending Club's pricing model is fundamentally a behavioral risk engine, not a status-based one.

Full analysis: Open in Colab

Author

Ido Aidan

Course: Introduction to Data Science | Reichman University (IDC)