Predictive Modeling of Household Credit Risk and Fear of Denial: A High-Dimensional Analysis Using PCA and XGBoost on the 2022 Survey of Consumer Finances (SCF)

The theoretical literature on credit markets emphasizes information asymmetries, adverse selection, and moral hazard as fundamental sources of credit rationing and market imperfections. In the seminal model of Stiglitz and Weiss (1981), lenders may deny credit to some borrowers even when they are willing to pay higher interest rates, because raising rates can attract riskier borrowers (adverse selection) or induce borrowers to undertake riskier projects (moral hazard). Empirical evidence from the SCF supports the presence of credit rationing, as researchers document that relaxed lending standards in the mid-2000s increased households' ability to obtain desired debt levels, indicating that supply-side constraints had previously limited access (Durguner, 2020). The subsequent tightening of underwriting standards during and after the Great Recession reduced credit availability, particularly for marginal borrowers, consistent with time-varying credit supply constraints affecting household borrowing. Credit access facilitates optimal consumption smoothing across the life cycle, but binding credit constraints can prevent households from achieving desired consumption paths, particularly when they are young or during periods of income volatility (Dogra & Gorbachev, 2016). Empirical SCF analyses document hump-shaped profiles of debt holdings and delinquency rates by age, with younger middle-income households experiencing increased arrears during economic downturns while older households with lower incomes face persistent repayment challenges (Grant, 2021). These age-differentiated patterns suggest that credit access and financial vulnerability vary systematically across the life cycle, with both income trajectories and accumulated wealth influencing borrowing capacity and default risk. From a Behavioral finance perspective, the prevalence of high-cost credit use among credit-constrained households, despite the availability of lower-cost alternatives in some cases, suggests that behavioral factors and limited financial sophistication may influence borrowing patterns (Lee & Kim, 2018).

Demographic Determinants of Credit Access. Empirical evidence from the SCF reveals heterogeneous patterns in arrears dynamics across age groups, with younger middle-income households experiencing increased delinquency in some periods while older households with lower incomes face earlier increases in repayment difficulties (Grant, 2021). These age-graded dynamics indicate that credit access and financial stress do not follow simple monotonic relationships with age but rather interact with income levels and economic conditions in complex ways. Racial and ethnic disparities in credit access and outcomes are well-documented and persistent across SCF waves, reflecting both observable socioeconomic differences and potentially discriminatory practices in lending markets. Anyamele (2018) employs Oaxaca-Blinder decomposition using SCF data from 2007 and 2010 to separate racial gaps in loan delinquency into explained and unexplained components. The analysis reveals that 67.33% of the Black-White delinquency gap can be attributed to observable endowment differences, including income, unemployment rates, credit-constrained status, and payday loan usage, while 33.08% remains unexplained by these factors. For the Hispanic-White gap, 93.03% is explained by observable characteristics, with only 7.36% unexplained. The substantial unexplained component of the Black-White gap, even after controlling for comprehensive financial and demographic variables, raises concerns about potential discrimination in credit markets or systematic measurement issues in capturing all relevant differences. Recent research extending this analysis through 2022 finds that student loans interact with race to affect credit card holdings and credit access patterns, suggesting that the intersection of education debt and racial identity compounds disadvantages in credit markets (Yoo & Felver, 2025). Educational attainment influences credit access through income potential, employment stability, and financial literacy, but the debt burden associated with higher education creates immediate financial stress that can reduce credit availability. The literature provides strong and consistent evidence that student loan exposure creates financial distress and constrains credit access. Bricker and Thompson (2016) analyze the 2007-2009 SCF panel and find that families with student loans in 2007 were 3.1 percentage points more likely to be 60 or more days late on payments and faced a 3% higher probability of credit denial, even after controlling for demographics and comprehensive balance sheet measures including income and net worth. These effects are economically meaningful, representing approximately 20-25% increases relative to baseline delinquency and denial rates. The persistence of student loan effects across multiple SCF waves through 2022 indicates that education debt continues to influence household credit trajectories over time (Hershbein & Hollenbeck, 2015; Yoo & Felver, 2025). While education increases long-term earning potential, the debt burden associated with educational attainment creates immediate liquidity constraints and signals elevated leverage to lenders, reducing access to additional credit during the repayment period. Employment status provides income flow necessary for debt service, signals stability to lenders regarding repayment capacity, and may include benefits such as emergency savings or employer-sponsored credit access. Unemployment emerges as a significant predictor of both delinquency and credit denials across multiple SCF-based studies. Anyamele (2018) identifies unemployment as a key factor explaining racial differences in delinquency rates, while Kim et al. (2017) document that employment status affects debt holding and delinquency patterns among low-income households, with job loss during the Great Recession contributing to elevated default rates. The recession period provides natural variation in employment status, and studies examining this episode consistently find that involuntary job loss increases delinquency risk and reduces credit access, with effects concentrated among households with limited liquid assets and high debt-service burdens. Geographic location affects credit access through local credit supply conditions, regional economic circumstances, housing market dynamics, and state-level regulatory environments. Sant'Anna et al. (2021) examine the relationship between credit availability and land values at the county level, finding that broader credit availability indices correlate with higher land values and that county-level credit supply variation has measurable effects on asset markets. This evidence indicates that local credit market conditions, beyond individual household characteristics, influence both credit access and asset prices.

Financial Factors Influencing Credit Access. Choi et al. (2015) provide detailed analysis of unsecured credit limit determinants using SCF data and instrumental variables to address endogeneity between net worth and credit limits. Their findings indicate that income level is the primary determinant of credit limits, with income growth strongly predicting higher limits and income volatility affecting lender risk assessments. Wealth provides collateral for secured borrowing, buffers against income shocks, signals creditworthiness and financial management capability, and offers an alternative to borrowing that reduces credit demand. Despite these theoretical roles, Choi et al. (2015) find that income dynamics dominate wealth in explaining unsecured credit limits, which may reflect that unsecured credit relies more on income flow than asset stock for repayment assessment. Kim et al. (2017) examine debt indicators among low-income households before and after the Great Recession using SCF data from 2007, 2010, and 2013, employing Heckman selection models to separate the probability of holding debt from debt outcomes conditional on holding debt. Their findings reveal that the lowest-income households, those below the poverty line, were less likely to meet debt-to-income guidelines typically used in underwriting. After the recession, the poorest households were more likely to have higher debt burdens and elevated delinquency rates, indicating that low-income households face both credit access challenges and repayment difficulties.

Credit Denials, Delinquency & Constraints. Credit denials occur when households apply for credit but are rejected by lenders, while credit constraints encompass both actual denials and self-imposed limitations where households do not apply because they expect to be denied. The SCF collects detailed information on credit applications, denials, reasons for denial, and whether households refrained from applying due to expected denial, enabling comprehensive analysis of credit constraints. Lee and Kim (2018) analyze the relationship between credit constraints and payday loan use across SCF waves from 2007 to 2013, finding that credit-constrained households were significantly more likely to use payday loans and that this relationship strengthened after the Great Recession. Macroeconomic shocks and economic downturns significantly affect delinquency and default rates, with the Great Recession providing a natural experiment for examining these effects. Multiple SCF-based studies document that delinquency rates rose sharply during and after the recession, with particularly large increases among low-income and minority households (Kim et al., 2017; Grant, 2021). Grant (2021) employs decomposition analysis to separate compositional changes in who holds debt from behavioral changes in repayment patterns, finding that most temporal variation in arrears stems from behavioral changes in how borrowers respond to given characteristics rather than compositional shifts in the borrower pool. This finding implies that repayment dynamics and post-origination factors, not just initial screening and selection, drive delinquency trends over time.

Literature Gaps and Research Direction. Despite substantial progress in understanding household credit access using the SCF, important research gaps remain. The literature lacks integration of machine learning methods for credit prediction and risk assessment, relying primarily on traditional econometric approaches. Machine learning techniques may offer superior predictive performance and ability to capture complex non-linear interactions. Accordingly, this research aims to address these gaps by evaluating the predictive value of advanced machine learning methods—specifically an XGBoost model augmented with PCA—and benchmarking their performance against the more interpretable and widely used Logistic Regression framework.

This study addressed the following research question: What household financial and demographic characteristics predict credit access concerns, specifically the experience of being turned down for credit or fearing credit denial? Based on prior literature on financial inequality and credit market disparities, as well as preliminary exploratory analysis of the SCF 2022 data, three primary hypotheses were tested.

Data Source and Overview. This study utilizes data from the 2022 Survey of Consumer Finances (SCF), a comprehensive triennial cross-sectional survey administered by the Federal Reserve Board. The SCF is designed to provide detailed information about the financial characteristics of U.S. households, including their balance sheets, pension holdings, income, and demographic characteristics. The dataset was obtained directly from the Federal Reserve's public data repository. A distinctive feature of the SCF data is its use of multiple imputation to handle missing values and measurement error. Each household in the survey is represented by five complete observations (implicates), resulting in a dataset where the total number of records is five times the number of unique families surveyed. For the 2022 SCF, this translates to 22,975 total records representing 4,595 unique families. The SCF 2022 summary extract contained hundreds of variables capturing diverse aspects of household finances, including:

(a) income measures (total household income, wage income, business income, investment income);

(b) asset holdings (value of primary residence, other real estate, vehicles, financial assets such as stocks, bonds, and mutual funds, and retirement accounts);

(c) debt obligations (mortgage debt, credit card balances, education loans, other consumer debt);

(d) financial behaviors (savings patterns, investment decisions, credit usage); and

(e) demographic characteristics (age, education, race/ethnicity, family composition, employment status).

I also extracted the SCF Variable names that I used for building the data dictionary mapping for variable codes (University of California, Berkeley, n.d.)

Data Demographic Composition. The 2022 SCF sample of 4,595 unique families exhibits a mature demographic profile relevant to credit access research. The sample is predominantly married or partnered (63%), with nearly 60% of households headed by individuals aged 55 or older. Nonwhite or Hispanic households comprise 40% of the sample, providing substantial racial and ethnic diversity critical for examining credit access disparities. Educational attainment skews higher, with two-thirds having at least some college education, potentially underrepresenting the most financially vulnerable populations who face the greatest credit barriers.

Target Variable: Credit Access Concerns. The primary outcome variable for this analysis is TURNFEAR, a binary indicator that captures whether a household has been turned down for credit or feared being denied credit in the past five years. Approximately 84% of the unique households in the sample report indicated credit access concerns in the past 5 years.

Data Preprocessing. Several data cleaning steps were implemented to prepare the dataset for analysis. First, variables were systematically classified as categorical or continuous based on cardinality thresholds. Features with fewer than 25 unique values were treated as categorical and converted to factor variables. Second, variables directly defining the target outcome (TURNDOWN and FEARDENIAL) were removed from the predictor set to prevent information leakage. Third, the dataset was assessed for missing data; no missing values were found in the summary extract, eliminating the need for imputation at the initial stage. Fourth, identifier and weight columns (X, WGT, YY1, Y1) were removed as they were irrelevant for modeling. Finally, features with zero variance (constant values across all observations) were systematically identified and removed, as they provided no discriminatory power for prediction.

Analytical Approach. This study employed a quantitative predictive modeling approach to identify household characteristics associated with credit access concerns. The analysis proceeded through four main stages: (a) feature engineering and multicollinearity assessment, (b) model development and training, and (c) model evaluation and comparison.

Three key theoretical insights emerged from the analysis. First, the dominance of asset-based and behavioral financial indicators (equity income, homeownership, credit history, emergency savings, leverage ratios) over simple income and wealth measures supports multidimensional theories of financial stratification. Credit access is not merely a function of current economic resources but reflects accumulated assets, investment capacity, payment behaviors, and prior credit market experiences. The prominence of EQUITINC representing investment income and capital gains suggests that credit markets differentially reward households with capital ownership versus labor income, potentially perpetuating wealth-based inequality. The exceptional model performance (AUC = 0.9885) indicates that credit access patterns are highly systematic and predictable from observable household characteristics.

The analysis provided strong support for all three hypotheses with important nuances. H₁ predicted that households with lower income and wealth would exhibit higher credit access concerns. This was strongly confirmed, through asset-based indicators equity to income (EQUITINC), homeownership (HHOUSES1), and emergency savings (EMERGSAV1) proved more powerful than simple income or wealth measures. Descriptive analysis in Figure 2, Figure 3 and Figure 4 reveals that households experiencing credit denial or fear demonstrated substantially lower net worth, normalized income, and liquid assets compared to the overall population, supporting this expectation. This extends H₁ by demonstrating that credit access depends specifically on asset ownership, investment capacity, and liquid reserves rather than economic resources generally.

H₂ predicted that demographic factors, particularly race and age, would demonstrate significant associations with credit access after controlling for economic resources. This received partial support with qualifications. Age ranked seventh among predictors, confirming importance and aligning with descriptive findings. However, race ranked 14^th substantially lower than anticipated given dramatic descriptive disparities (64.5% of affected households were nonwhite/Hispanic versus 40.2% overall). This suggests that observed racial disparities operate substantially through differential economic circumstances particularly homeownership, investment income, and payment history rather than as direct independent effects. Race's persistence in the top 15 nonetheless confirms remaining structural barriers. Exploratory comparisons in Figure 5 shows that nonwhite or Hispanic households comprised 64.5% of those with credit concerns despite representing only 40.2% of the sample, while younger households (under age 45) were disproportionately affected in Figure 6.

H₃ predicted moderate-to-strong predictive performance (AUC > 0.75). This was dramatically exceeded, with Model A achieving AUC = 0.9885, demonstrating that credit access concerns are highly systematic and predictable from observable characteristics rather than idiosyncratic.

Policy Implications. The results suggest four policy directions. First, asset-building policies targeting homeownership and investment income generation such as down payment assistance programs, community land trusts, and expansion of matched savings accounts for securities investment represent fundamental strategies for improving credit access. Second, emergency savings programs and financial resilience initiatives addressing material hardship indicators could reduce credit vulnerability by enabling households to avoid payment delinquencies and credit denials during temporary income disruptions. Third, the importance of payment history and credit application history highlights need for credit repair programs, alternative credit scoring models that weigh recent versus historical behavior, and policies reducing barriers to credit access after past delinquencies. Fourth, addressing racial disparities requires targeted interventions through strengthened fair lending enforcement.

Limitations. The study is not without certain limitation. Key limitations include (a) sampling error (our study is based on a survey with 4595 unique families), (d) absence of credit bureau data (credit scores, detailed credit histories) that lenders actually use; (e) no examination of variation across credit product types (mortgages, auto loans, credit cards, small business loans); (f) unique 2022 economic context following COVID-19 pandemic; and (g) limited exploration of interactions among predictors (e.g., how race moderates relationships between financial factors and credit access).

The complete R Code used for this analysis is available at https://github.com/jainrachit1008/Predictive_Modeling_of_HouseholdCreditRisk-_and_FearofDenial.