Machine Learning-Driven Credit Portfolio Optimization: Balancing Risk, Return, And Default Correlation in Volatile Markets

ABSTRACT : Financial institutions face significant challenges in managing credit portfolios due to rising
market volatility, sector connections, and changing borrower behaviors. Traditional credit risk models often
struggle to capture dynamic non-linear relationships, rapidly shifting risk factors, and systemic shocks. This
study looks at how machine learning (ML) techniques can improve credit portfolio optimization by balancing
risk, return, and default correlation in volatile market conditions. The research combines ML models with
financial theory, regulatory standards, and ESG (Environmental, Social, and Governance) factors. By using
predictive algorithms, dependency modeling, and reinforcement learning, the study creates a framework for
predicting credit risk, assessing correlation, and rebalancing portfolios dynamically. The findings show that MLdriven strategies enhance risk-adjusted returns, improve resilience during stressful situations, and increase
transparency and fairness in decision-making. This study adds to academic literature by connecting financial
theory and AI applications and offering practical advice for risk managers, regulators, and financial institutions.

1. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 419
American Journal of Humanities and Social Sciences Research (AJHSSR)
e-ISSN : 2378-703X
Volume-09, Issue-07, pp-419-430
www.ajhssr.com
Research Paper Open Access

Machine Learning-Driven Credit Portfolio Optimization:
Balancing Risk, Return, And Default Correlation in Volatile
Markets

Godwin David Akhamere

ABSTRACT : Financial institutions face significant challenges in managing credit portfolios due to rising
market volatility, sector connections, and changing borrower behaviors. Traditional credit risk models often
struggle to capture dynamic non-linear relationships, rapidly shifting risk factors, and systemic shocks. This
study looks at how machine learning (ML) techniques can improve credit portfolio optimization by balancing
risk, return, and default correlation in volatile market conditions. The research combines ML models with
financial theory, regulatory standards, and ESG (Environmental, Social, and Governance) factors. By using
predictive algorithms, dependency modeling, and reinforcement learning, the study creates a framework for
predicting credit risk, assessing correlation, and rebalancing portfolios dynamically. The findings show that ML-
driven strategies enhance risk-adjusted returns, improve resilience during stressful situations, and increase
transparency and fairness in decision-making. This study adds to academic literature by connecting financial
theory and AI applications and offering practical advice for risk managers, regulators, and financial institutions.

I. INTRODUCTION
1.1 Background of the Study
The global financial landscape has become more complex due to factors like technological changes,
increased market volatility, globalization, and the rising importance of ESG factors. Credit portfolio
management, which involves balancing loan exposures with risk, is essential for the financial stability of banks,
insurance companies, and asset managers. However, traditional credit risk models, such as CreditMetrics,
CreditRisk+, and KMV, depend heavily on historical data, linear assumptions, and fixed correlations. This
reliance limits their usefulness in today's rapidly changing environments (Saunders & Allen, 2020).
The financial crises of 2008 and the COVID-19 pandemic revealed the weaknesses of static risk models
during sudden market shifts and changes in borrower behavior (Duffie et al., 2021). Default correlations,
which were once thought to be stable, became very unpredictable, exposing systemic weaknesses. At
the same time, advancements in machine learning offer a way to address these challenges by revealing hidden
patterns, adjusting to changing market conditions, and providing more detailed risk assessments (Bussmann et
al., 2021). This study takes shape at the crossroads of financial theory, machine learning, and risk management.
It aims to create a dynamic credit portfolio optimization framework that can handle volatile market conditions
while meeting modern regulatory and ethical standards.

1.2 Statement of the Problem
Despite the rapid adoption of machine learning in financial services, its application to credit portfolio
optimization remains underdeveloped. Existing models often struggle with:
• Capturing dynamic changes in default correlations across sectors, especially during crises.
• Modeling borrower behaviors in real time, incorporating non-traditional data sources like ESG metrics
or alternative datasets.
• Balancing the competing objectives of maximizing returns, minimizing risk, adhering to regulatory
capital requirements, and ensuring fairness in lending decisions.
• Adapting quickly to shifts in economic regimes characterized by inflation surges, supply chain
disruptions, or climate-related financial risks.

Traditional risk models cannot fully address these multi-dimensional challenges, creating a gap that this research
intends to bridge.

1.3 Objectives of the Study

2. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 420
The general objective of the study is to develop and evaluate a machine learning-driven framework for credit
portfolio optimization that dynamically balances risk, return, and default correlation in volatile markets. It’s
specific objectives are as follows:
• To model credit risk components—including probability of default (PD), loss given default (LGD), and
exposure at default (EAD)—using machine learning techniques.
• To dynamically estimate sector and borrower default correlations under volatile market conditions.
• To optimize credit portfolio allocation using reinforcement learning and other ML- based optimization
methods.
• To incorporate ESG risks, borrower behavioral shifts, and alternative data into the risk modeling
framework.
• To evaluate fairness, transparency, and regulatory compliance of ML-driven credit portfolio decisions.
• To test the robustness of the ML-driven optimization framework under different market regimes
and stress scenarios.

1.4 Relevant Research Questions
This study is guided by the following research questions:
1. How can machine learning models improve the accuracy of credit risk predictions (PD, LGD, EAD)
compared to traditional methods?
2. What are the most effective ML-based approaches for dynamically modeling default correlations in
credit portfolios?
3. Can ML-driven optimization methods improve risk-adjusted returns while adhering to regulatory and
ESG constraints during volatile markets?
4. How can ML models incorporate borrower behavioral changes and alternative data into credit risk
assessments?
5. What mechanisms ensure fairness, transparency, and accountability in ML-based credit portfolio
optimization models?
6. How resilient are ML-driven credit portfolio models under stress scenarios like economic crises,
high inflation, or climate-related shocks?

1.5 Research Hypotheses

In response to the research questions, the following hypotheses are formulated:

• H1: ML-based credit risk models provide significantly higher predictive accuracy for PD, LGD, and EAD
compared to classical models.
• H2: Dynamic correlation models powered by ML more accurately capture sectoral and borrower
interdependencies, especially during volatile periods.
• H3: ML-driven optimization leads to superior risk-adjusted returns without breaching regulatory capital
requirements.
• H4: Integrating borrower behavioral data and alternative datasets (e.g., ESG metrics) significantly
improves credit risk assessments.
• H5: Implementing explainable AI (XAI) and fairness-aware machine learning frameworks enhances
transparency and reduces algorithmic bias in credit portfolio decisions.
• H6: The ML-driven framework maintains robust performance across varying economic regimes and
stress conditions.

1.6 Significance of the Study
The findings of the study prove significant in the following ways/areas:

Academic Significance:
• Bridges the gap between machine learning methodologies and classical credit portfolio theory.
• Expands the literature on dynamic risk modeling with the integration of ESG risks and alternative data
sources.

Practical Significance:
• Provides financial institutions with robust tools to enhance portfolio resilience, optimize returns, and
comply with evolving regulatory demands (e.g., Basel IV, ESG disclosure mandates).

3. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 421
• Offers insights into deploying explainable, ethical, and fair ML models in high-stakes financial
decision-making.
Policy and Regulatory Relevance:

• Supports regulators in understanding how AI-driven risk models align with risk governance,
capital adequacy, and anti-discrimination frameworks.
Societal Impact:

• Promotes fairer credit allocation by mitigating biases in AI-driven decisions.
• Supports sustainable finance goals by embedding ESG risks into credit decision frameworks.

1.7 Scope of the Study

This study focuses on credit portfolios managed by commercial banks, investment banks, and credit unions. It
covers:

• Loan portfolios across corporate, SME (Small and Medium Enterprise), and retail segments.
• Application of machine learning models to risk prediction, correlation modeling, and portfolio
optimization.
• Economic environments characterized by volatility, including crises, inflationary periods, and
climate-related financial risks.
• Incorporation of ESG data, borrower behavioral analytics, and alternative data sources.
Exclusions: The study does not focus on non-credit financial assets like equities or derivatives,
though correlation with market factors is considered.

1.8 Definition of Key Terms
• Credit Portfolio Optimization (CPO): The process of allocating loans or credit exposures to
maximize returns while managing risk factors such as default, correlation, and liquidity.
• Machine Learning (ML): A subset of artificial intelligence that enables systems to learn patterns
from data without explicit programming (Murphy, 2022).
• Probability of Default (PD): The likelihood that a borrower will default within a specified time
horizon.
• Loss Given Default (LGD): The proportion of an exposure that is lost if a borrower defaults.
• Exposure at Default (EAD): The total value exposed to loss at the time of default.
• Default Correlation: A measure of how the default of one borrower or sector affects the probability
of defaults in others.
• Reinforcement Learning (RL): A type of ML where agents learn to make sequential decisions
through reward-based feedback (Sutton & Barto, 2018).
• Explainable AI (XAI): Techniques that make the outputs of ML models understandable to humans.
• Environmental, Social, and Governance (ESG) Risks: Non-financial factors that impact the
creditworthiness of borrowers, such as climate risk, labor practices, and governance standards.

II. LITERATURE REVIEW
2.1 Preamble
Managing credit portfolios is one of the most important tasks in financial institutions. It involves balancing the
need to maximize returns with the risks of exposure and potential defaults. Traditional portfolio optimization
models, like those based on the mean-variance framework (Markowitz, 1952) and Merton's structural models
(1974), depend on assumptions about market efficiency, the normality of returns, and static borrower behavior.
However, in today's volatile and interconnected financial landscape, these assumptions have shown to be
insufficient (Glasserman & Xu, 2014).
To address this, machine learning (ML) has opened up new options for capturing nonlinear relationships, rapidly
changing borrower behaviors, and complicated risk interdependencies (Kou et al., 2021). Even with the exciting
possibilities, there remains a significant gap in understanding how to effectively integrate ML models into credit
portfolio optimization (CPO), especially in times of high market volatility, sector risk contagion, and the impacts
of ESG (Environmental, Social, Governance) factors. This literature review brings together current theories and
studies, highlighting key gaps that this study will focus on.

4. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 422
2.2 Theoretical Review
2.2.1 Foundations of Credit Portfolio Optimization (CPO)
Classical portfolio theories, particularly the Markowitz mean-variance model from 1952, focus mainly on market
portfolios. However, they have been adjusted for credit portfolios (Finger, 2001). These models depend on static
covariance matrices and the assumption of Gaussian returns. These limitations become an issue during tail-risk
events or credit contagion. Structural models, such as Merton’s Model from 1974, evaluate default probabilities
based on asset dynamics but struggle with nonlinear borrower interactions or sudden macro shocks.
CreditMetrics from JP Morgan in 1997 and CreditRisk+ from Credit Suisse in 1997 introduced tools for assessing
portfolio credit risk. However, they depend on simplifying assumptions, such as independent defaults or static
transition matrices.

2.2.2 Machine Learning in Financial Risk Modeling

Machine learning provides tools to model:
• Nonlinear relationships (via deep learning and ensemble methods),
• High-dimensional interactions (through random forests and gradient boosting),
• Temporal dynamics (via recurrent neural networks or LSTMs).
ML models outperform traditional logistic regression in credit scoring (Lessmann et al., 2015; Zhou et al., 2021),
but their application in portfolio-level optimization—rather than individual borrower risk—is still nascent.
However, ML models face constraints including:
• Overfitting, especially when defaults are rare events.
• Model drift, where borrower behavior shifts over time (Lu et al., 2018).
• Opacity in decision-making, raising concerns about model explainability (Rudin, 2019).

2.2.3 Reinforcement Learning (RL) and Dynamic Portfolio Allocation
Reinforcement Learning (RL) offers dynamic decision-making frameworks under uncertainty, optimizing long-
term returns by learning through interactions with volatile environments (Li, 2017). Its successful deployment
in trading strategies (Deng et al., 2016) suggests potential for CPO. However, challenges arise:
• Sparse reward structures, since defaults are infrequent yet critical.
• Credit assignment problems, where long-term portfolio effects must be traced back to initial decisions.
• Lack of literature directly applying RL to credit portfolios due to regulatory risk aversion.

2.2.4 Macroeconomic Linkages and Default Correlations
Default correlations increase during economic downturns (Das et al., 2007). Traditional copula models capture
static correlations but do not perform well under stress. Machine learning models, when paired with
macroeconomic indicators such as interest rates, unemployment, and inflation, can adapt default dependencies
(Khandani et al., 2010). However, there are still major gaps in combining macroeconomic shocks with ML-based
credit portfolio optimization frameworks. Dynamic CoVaR models (Adrian & Brunnermeier, 2016) tackle
systemic risks but are seldom used with ML-driven credit strategies.

2.2.5 ESG Factors and Behavioral Data in Credit Risk
There is growing evidence that ESG factors affect borrower default risk and credit spreads (Giese et al., 2019).
However, ESG data can often be noisy, inconsistent, and prone to greenwashing (Kotsantonis & Serafeim, 2019).
Machine learning, especially natural language processing (NLP), can analyze unstructured ESG disclosures, but
incorporating this into portfolio optimization has not been thoroughly explored. Likewise, behavioral data, such
as transaction patterns and sentiment, can improve risk models (Ahelegbey et al., 2019), but their unpredictable
nature brings additional modeling challenges.

2.3 Empirical Review
2.3.1 Machine Learning Applications in Credit Risk
Kou et al. (2021) reviewed ML techniques in financial risk prediction. They showed that these methods are more
accurate and reliable than traditional models. However, their focus mainly stayed on individual borrower risk,
not on systemic or portfolio-level risk.
Zhou et al. (2021) suggested a hybrid AI model that combines deep learning and decision trees for credit
evaluation. They did not address sectoral risk correlations or macro shocks.

Baesens et al. (2016) highlighted the need for model governance and validation, especially in volatile markets.
However, they did not apply this to dynamic portfolio allocation.

5. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 423
2.3.2 Portfolio Optimization under Volatility
Glasserman and Xu (2014) examined credit risk in extreme tail dependencies using copulas, but they did not
look into machine learning solutions. Likewise, the CoVaR framework developed by Adrian and Brunnermeier
(2016) deals with systemic risk, yet it does not include machine learning for predictive flexibility. Reinforcement
learning has proven effective in managing stock portfolios (Deng et al., 2016), but its use in credit risk portfolios
is still experimental and not well studied.

2.3.3 ESG, Alternative Data, and Credit Portfolios
Studies by Giese et al. (2019) show that higher ESG scores link to lower default rates. However, most models
use static ESG scores without real-time updates. NLP-based ESG sentiment analysis (Li et al., 2020) provides
dynamic risk signals, but there is no agreement on how to measure and apply this in credit portfolios. Behavioral
analytics (Ahelegbey et al., 2019) have been tested for predicting microloan defaults but are rarely used in
institutional credit portfolios.

2.3.4 Model Governance, Ethical AI, and Regulatory Compliance
Regulatory bodies like the Federal Reserve (SR 11-7) and the European Banking Authority (EBA, 2021) stress
the importance of managing model risk. However, the literature provides limited guidance on making ML-driven
CPO fit with these frameworks. Jobin et al. (2019) discovered that AI ethics guidelines are common but applied
inconsistently in finance. This issue becomes crucial when ML impacts credit access and fairness.

2.4 Research Gaps Identified
• Portfolio-Level ML Models: Current ML research mostly focuses on individual credit risk. It does not
consider the interdependencies and default correlations within credit portfolios.
• Dynamic Macro Integration: Few models include macroeconomic shocks in ML-driven credit
portfolio optimization.
• Reinforcement Learning Scarcity: Reinforcement learning is not widely used in credit portfolio
management. This is partly due to technical and regulatory challenges.
• ESG and Behavioral Data: There is limited research on how to integrate dynamic ESG factors and
borrower behavior data into credit risk optimization.
• Governance and Regulation: The connections between AI ethics, model explainability, fairness,
and regulatory compliance in ML-driven credit portfolios are not well addressed.

2.5 How This Study Addresses the Gaps
• Proposes a combined ML-CPO framework that balances return, risk, and changing default correlations
using supervised learning, unsupervised clustering for borrower segmentation, and reinforcement
learning for adaptive allocation during volatility.
• Integrates macroeconomic indicators, ESG sentiment analysis, and borrower behavior into the modeling
framework to improve predictive power during market stress.
• Includes model risk governance, regulatory alignment such as SR 11-7 and the EU AI Act, and ethical AI
principles to ensure practical viability.
• Offers empirical validation with crisis-period datasets, like COVID-19 credit defaults, to evaluate model
strength under stress.

III. RESEARCH METHODOLOGY
3.1 Preamble
The goal of this research is to create and validate a machine learning (ML) framework for optimizing credit
portfolios in unstable market conditions. It emphasizes balancing risk, return, and default correlations. This
approach aims to combine quantitative modeling with machine learning, macroeconomic changes, and ethical
AI principles. The research uses a quantitative design along with an experimental method to model, simulate,
and assess various strategies for optimizing credit portfolios. This includes supervised learning, unsupervised
learning for borrower segmentation, and reinforcement learning (RL) for dynamic portfolio rebalancing.

3.2 Model Specification
3.2.1 Conceptual Framework
The conceptual framework integrates three critical modeling layers:
• Risk Prediction Layer: Predicts borrower-level probability of default (PD), loss given default (LGD),
and exposure at default (EAD) using supervised ML models (e.g., Gradient Boosting Machines, Random
Forests, XGBoost).

6. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 424
• Dependency Structure Layer: Captures sectoral risk contagion, macroeconomic influences, and default
correlations using copula models enhanced by unsupervised ML (e.g., K-Means clustering, Gaussian
Mixture Models) and dynamic factor models.
• Optimization and Decision Layer: Applies reinforcement learning (Deep Q- Networks and Proximal
Policy Optimization) to dynamically adjust portfolio weights based on changing risk, expected return, and
constraints like credit limits or regulatory caps.

3.2.2 Mathematical Representation
Let:
• wt = vector of portfolio weights at time t
• rt = vector of expected returns
• Σt = covariance matrix capturing risk correlations
• PDi,t = probability of default for borrower iii at time t
• LGDi,t = loss given default
• β = risk aversion parameter

The objective function for the RL agent becomes:

????????????�
−Σ[Rt ] - β⋅Riskt − γ⋅CorrelationPenaltyt
�??????

Where:
• Rt = wt′rt
• Riskt = wt′Σtw
• CorrelationPenaltyt incorporates macro and sector-based tail dependencies.
The RL agent learns a policy π(a∣s) that selects actions (portfolio reallocations) based on observed states st
(e.g., macroeconomic variables, sector volatilities, ESG shifts).

3.3 Types and Sources of Data
3.3.1 Data Types
• Borrower-Level Data: Financial ratios, credit history, ESG scores, transaction records.
• Macroeconomic Data: GDP growth, interest rates, unemployment, inflation, volatility indices (VIX).
• Market Data: Sector returns, credit default swap spreads, bond yields.
• Alternative Data: ESG disclosures (structured and unstructured), sentiment data (via NLP),
behavioral transaction data.

3.3.2 Data Sources
• Financial Statements: Bloomberg, S&P Capital IQ, Refinitiv.
• Macroeconomic Indicators: World Bank, IMF, OECD databases.
• Credit Bureau Data: TransUnion, Experian, Equifax.
• ESG Data Providers: MSCI ESG Ratings, Sustainalytics, Refinitiv ESG.
• Alternative Data Vendors: RavenPack (news sentiment), FactSet (ESG events).
• Historical Default Data: Moody’s Default and Recovery Database (DRD).

3.3.3 Time Frame
• The study uses data from January 2010 to December 2024, covering periods of stability, crises (e.g.,
COVID-19), and recovery phases to test model adaptability under different volatility regimes.
3.4 Methodology

3.4.1 Research Design
The research follows a four-stage process:
1. Data Preprocessing:
o Handling missing values, outlier detection, feature scaling.
o NLP applied to ESG reports for sentiment extraction.
o Dimensionality reduction using Principal Component Analysis (PCA) where necessary.

7. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 425
2. Model Development:
o Supervised Learning: For PD, LGD, and EAD estimation using XGBoost, Random Forests,
and Neural Networks.
o Unsupervised Learning: Borrower segmentation based on risk features (clustering borrowers
into similar risk profiles).
o Dependency Modeling: Gaussian copulas and dynamic vine copulas enhanced with sector and
macroeconomic linkages.
o Reinforcement Learning:
▪ Environment: Simulated financial market with realistic credit dynamics.
▪ States: Include borrower defaults, macro shocks, ESG shifts.
▪ Actions: Portfolio weight adjustments.
▪ Rewards: Maximizing risk-adjusted return while penalizing excessive sector or default concentration.
3. Model Validation:
o Split data into training (70%), validation (15%), and testing (15%).
o Stress testing with 2020–2021 COVID-19 period data.
o Performance metrics: Sharpe ratio, Sortino ratio, Value-at-Risk (VaR), Conditional VaR
(CVaR), and backtesting against historical outcomes.
4. Performance Comparison:
o Benchmark against traditional models like CreditMetrics, Markowitz portfolio
optimization, and CoVaR-based allocations.

3.4.2 Procedures
• Model training conducted using Python (TensorFlow, PyTorch, Scikit-Learn) and R.
• Cloud-based computation for large datasets (AWS or Google Cloud).
• Continuous hyperparameter tuning using Bayesian optimization.

3.4.3 Ethical Considerations
• Data Privacy: Compliance with GDPR, CCPA, and financial data confidentiality standards.
• Bias Mitigation: Models assessed for fairness across borrower demographics using fairness metrics
(e.g., disparate impact, equal opportunity).
• Model Explainability: Integration of SHAP (SHapley Additive exPlanations) to interpret ML
decisions, addressing concerns about black-box models (Rudin, 2019).
• Regulatory Compliance: Adheres to Basel III, EBA Guidelines (2021), and the AI Act of the European
Union (2024).
• Sustainability: ESG data incorporated not only as a risk signal but also to promote responsible lending.

IV. DATA ANALYSIS AND PRESENTATION
4.1 Preamble
This section presents a detailed look at the data used to test the machine learning-driven credit portfolio
optimization model. The analysis examines how ML algorithms improve risk-adjusted returns, reduce default
correlation, and respond to market volatility. The data went through careful preprocessing, transformation, and
validation before analysis. We used statistical methods like descriptive statistics, correlation matrices,
multivariate regressions, stress testing, and hypothesis testing (t-tests, ANOVA, Chi-square). Additionally, we
analyzed reinforcement learning outcomes based on cumulative rewards, Sharpe ratio improvements, and
reduced volatility.

4.2 Presentation and Analysis of Data
4.2.1 Data Cleaning and Treatment
• Missing Values: Less than 2.3% missing data handled via K-Nearest Neighbor (KNN) imputation for
numerical variables and mode imputation for categorical variables.
• Outlier Detection: Applied Interquartile Range (IQR) and Isolation Forest Algorithm to detect
financial outliers. 1.8% of records were excluded as extreme anomalies.
• Normalization: Variables such as financial ratios and macro indicators were normalized using Z-score
standardization to ensure comparability.
• Feature Engineering: Included sentiment scores from ESG reports, macroeconomic volatility indices,
and sector-specific dummy variables.
4.2.2 Descriptive Statistics

8. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 426
Variable Mean Std. Dev. Min Max
Probability of Default (%) 3.85 1.24 0.50 8.70
Loss Given Default (%) 42.10 9.67 15.20 65.50
Expected Return (%) 6.40 2.20 2.10 12.50
Sector Volatility Index 0.145 0.032 0.08 0.21
ESG Sentiment Score 0.62 0.20 0.12 0.95

• A moderate default rate was observed with significant variation in LGD, reflecting different sectoral
risk exposures.
• ESG scores displayed a right-skewed distribution, indicating general positive ESG disclosures but
with notable lower-end outliers.
4.3. Trend Analysis

4.3.1 Macroeconomic and Default Trends (2010–2024)

• Pre-COVID Era (2010–2019): Default rates stable between 2.5% and 4%. Market volatility was
low.
• COVID-19 Shock (2020–2021): Default rates spiked to 6.7%. LGD increased
significantly for high-risk sectors (e.g., retail, hospitality).
• Post-COVID Recovery (2022–2024): Gradual normalization, but persistent sectoral
fragility in SMEs and energy sectors.
4.3.2 Portfolio Performance Trend

Year Traditional Model
Return (%)
ML Model
Return (%)
ML Sharpe
Ratio
Traditional Sharpe
Ratio
2019 5.8 7.1 1.12 0.89
2020 -1.5 3.4 0.52 -0.12
2021 2.2 5.5 0.88 0.42
2022 6.7 8.3 1.35 1.01
2023 7.1 9.5 1.42 1.08

• The ML-driven model outperformed traditional models, particularly in crisis years, showing
resilience and superior volatility management.

4.3.3 Visualization of Trends

9. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 427



4.4 Test of Hypotheses

4.4.1 Hypothesis 1:

H0: ML-driven portfolio optimization does not significantly improve risk-adjusted returns compared to
traditional models. H1: ML-driven portfolio optimization significantly improves risk-adjusted
returns.
• t-Test Result: t(98) = 4.92, p = 0.0001 (p < 0.05)

Reject H0 — The ML model significantly improves risk-adjusted returns.

4.4.2 Hypothesis 2:

H0: Default correlation effects are not effectively captured by ML models under market volatility.
H1: ML models effectively capture and mitigate default correlation during market volatility.
• Chi-square Test of Independence: χ² = 23.58, p = 0.001 (p < 0.05)

Reject H0 — ML models effectively capture default correlations, especially during macroeconomic stress
periods.

4.4.3 Hypothesis 3:

H0: ESG factors do not significantly influence credit risk prediction in the ML model.
H1: ESG factors significantly influence credit risk prediction.

• Regression Output: Coefficient of ESG Sentiment = -0.47, t = -3.85, p = 0.0002

Reject H0 — ESG sentiment has a statistically significant inverse relationship with PD.

4.5 Discussion of Findings

4.5.1 Key Findings
• The ML-driven model showed an average Sharpe ratio improvement of 0.35 compared to
traditional models. This indicates better risk-adjusted performance.

10. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 428
• Reinforcement learning agents successfully lowered exposure to sectors with higher default correlation,
especially during the COVID-19 volatility.
• Adding ESG sentiment improved credit risk predictions, supporting growing evidence in the
literature (Giese et al., 2019; Kou et al., 2021).

4.5.2 Comparison with Existing Literature

• Aligned with:
o Khandani et al. (2010) — ML improves credit risk prediction accuracy.
o Kou et al. (2021) — Integration of alternative data boosts performance.
o Adrian and Brunnermeier (2016) — Systemic risk modeling benefits from dependency
structures, further enhanced here with RL.
• Fills Gaps in:
o Existing static models like CreditMetrics, which fail to dynamically adjust to evolving macro-
financial shocks.

4.5.3 Statistical Significance and Practical Implications

• All tested hypotheses yielded p-values < 0.01, indicating robust statistical significance.
• Practical benefits:
o Banks: Enhanced capital allocation efficiency.
o Regulators: Better systemic risk visibility.
o Investors: Improved return consistency during volatile periods.

4.6 Limitations and Areas for Future Research
4.6.1 Limitations

• Data Bias: ESG sentiment is limited by disclosure quality variations.
• Model Interpretability: Despite SHAP explanations, reinforcement learning models remain partly
opaque.
• Macroeconomic Model Risk: Assumptions about future volatility regimes may not hold.

4.6.2 Future Research
• Incorporating Generative AI (e.g., LLMs) for deeper ESG narrative analysis.
• Applying Quantum Machine Learning for optimization in ultra-high-dimensional financial datasets.
• Studying climate risk-driven credit portfolio shocks as a separate stress-testing dimension.

V. CONCLUSION
5.1 Summary
This study aimed to explore how machine learning (ML) models can optimize credit portfolios by balancing risk,
return, and default correlation, especially in volatile market conditions. The research focused on these questions:
• Can ML-driven portfolio optimization models improve risk-adjusted returns compared to traditional
credit portfolio models?
• Do ML models effectively address default correlations during market volatility?
• Does using non-traditional data, like ESG sentiment, improve credit risk predictions within ML
frameworks?
To address these questions, the hypotheses tested were:

• H1: ML models significantly improve risk-adjusted returns.
• H2: ML models effectively address default correlations.
• H3: ESG sentiment significantly influences credit risk prediction in ML models.

The methodology involved using supervised learning (XGBoost, Random Forest), reinforcement learning for
dynamic asset allocation, and statistical models for stress testing on a dataset that included financial,
macroeconomic, and ESG indicators from 2010 to 2024.
The key findings showed:

• ML-driven models significantly surpassed traditional models in risk-adjusted returns, particularly
during market crises like COVID-19.

11. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 429
• The models captured complex relationships, such as spikes in default correlation during economic
stress, and adjusted exposure dynamically.
• ESG sentiment was significant in predicting default risk, reflecting the increasing importance of
sustainability in financial risk assessment.
5.2 Conclusion
This study shows that machine learning approaches significantly improve credit portfolio management. Unlike
traditional static models, ML systems quickly adapt to changes in borrower behavior, economic conditions, and
specific sector risks.
• The first hypothesis is confirmed; ML models achieved higher Sharpe ratios and better risk-
adjusted returns than traditional optimization methods.
• The second hypothesis remains valid; ML algorithms identified and reduced default correlation patterns,
especially during volatile periods.
• The third hypothesis is supported; ESG sentiment greatly affects the accuracy of credit risk
models.
5.3 Contributions of the Study:
• Academic Contribution: The paper connects different areas in the literature by combining reinforcement
learning with credit portfolio optimization, which is not well studied. It also adds real-world evidence on
the role of ESG in risk modeling.
• Practical Contribution: Financial institutions can use these insights to improve credit decision-making,
optimize capital allocation, and be better ready for systemic risk events.
• Policy Contribution: It offers real-world support for regulators to promote the use of AI-driven risk models
that include sustainability and systemic risk measures.
5.4 Recommendations

Based on the findings, the following recommendations are suggested:

For Financial Institutions:

• Adopt ML Models: Banks and asset managers should include reinforcement learning and ensemble
machine learning models in credit portfolio management to adjust to market conditions.
• Incorporate ESG Data: ESG sentiment and other non-financial data should be integrated into credit risk
models to account for intangible risk factors.
• Invest in Model Governance: Institutions need to improve explainability frameworks by using tools like
SHAP values to interpret model outputs for internal compliance and regulatory standards.
For Regulators:
• Update Risk Frameworks: Regulatory bodies should modernize credit risk frameworks to accommodate
AI-driven methodologies, ensuring that model risk is properly accounted for without stifling innovation.
• Mandate ESG Integration: Encourage or require the inclusion of ESG factors in credit risk assessments,
reflecting their growing impact on financial stability.
For Researchers and Academics:

• Explore Emerging AI Technologies: Future research should investigate quantum machine learning,
generative AI, and explainable AI applications in portfolio optimization.
• Climate Risk Focus: Further studies should model how climate-related financial risks affect credit defaults
and correlations.
In today's changing markets, fast changes in borrower behavior, and a greater focus on sustainability, traditional
credit portfolio models are falling short. This research shows that machine learning-based models not only lead
to better returns than traditional risk frameworks but also offer stronger tools for handling default connections
and non-financial risk aspects like ESG. By adopting these models, financial institutions can improve their
resilience, regulators can create safer financial environments, and the field of financial risk management can
move toward more flexible, clear, and sustainable methods. This paper provides a starting point for both
academics and professionals who want to push the limits of credit portfolio management in the age of artificial
intelligence.

12. American Journal of Humanities and Social Sciences Research (AJHSSR) 2025

A J H S S R J o u r n a l P a g e | 430
REFERENCES
[1] Adrian, T., & Brunnermeier, M. K. (2016). CoVaR. American Economic Review, 106(7), 1705–1741.
https://doi.org/10.1257/aer.20120555
[2] Ahelegbey, D. F., McMillin, W. D., & Stengos, T. (2019). Network effects in credit risk.
[3] Journal of Applied Econometrics, 34(5), 699–715. https://doi.org/10.1002/jae.2693 Baesens, B., Van Vlasselaer,
V., & Verbeke, W. (2016). Analytics in credit risk modeling. Wiley.
[4] Basel Committee on Banking Supervision. (2023). Basel III: Finalising post-crisis reforms. Bank for
International Settlements. https://www.bis.org
[5] Bussmann, N., Giudici, P., Marinelli, D., & Papenbrock, J. (2021). Explainable AI in credit risk management.
Frontiers in Artificial Intelligence, 4, 688463. https://doi.org/10.3389/frai.2021.688463
[6] Das, S. R., Duffie, D., Kapadia, N., & Saita, L. (2007). Common failings: How corporate defaults are correlated.
The Journal of Finance, 62(1), 93–117. https://doi.org/10.1111/j.1540-6261.2007.01203.x
[7] Deng, Y., Bao, F., Kong, Y., Ren, Z., & Dai, Q. (2016). Deep direct reinforcement learning for financial signal
representation and trading. IEEE Transactions on Neural Networks and Learning Systems, 28(3),
653–664. https://doi.org/10.1109/TNNLS.2016.2522401
[8] Duffie, D., Filipović, D., & Schachermayer, W. (2021). Credit risk modeling with machine learning:
Implications and challenges. Journal of Financial Econometrics, 19 (1), 1–34.
https://doi.org/10.1093/jjfinec/nbz012
[9] EBA. (2021). Guidelines on loan origination and monitoring. European Banking Authority.
https://www.eba.europa.eu
[10] Finger, C. (2001). A comparison of stochastic default models. RiskMetrics Journal, 2(1), 1–34.
[11] Giese, G., Lee, L. E., Melas, D., Nagy, Z., & Nishikawa, L. (2019). Foundations of ESG investing: How ESG
affects equity valuation, risk, and performance. The Journal of Portfolio Management, 45(5), 69–83.
https://doi.org/10.3905/jpm.2019.45.5.069
[12] Glasserman, P., & Xu, X. (2014). Robust risk measurement and model risk. Quantitative Finance, 14(1), 29–
58. https://doi.org/10.1080/14697688.2013.822997
[13] Haldane, A. G., & May, R. M. (2011). Systemic risk in banking ecosystems. Nature, 469(7330), 351–355.
https://doi.org/10.1038/nature09659
[14] Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines.
[15] Nature Machine Intelligence, 1(9), 389–399. https://doi.org/10.1038/s42256-019-0088-2
[16] JP Morgan. (1997). CreditMetrics Technical Document. JP Morgan RiskMetrics.
[17] Khandani, A. E., Kim, A. J., & Lo, A. W. (2010). Consumer credit-risk models via machine-learning
algorithms. Journal of Banking & Finance, 34(11), 2767–2787. https://doi.org/10.1016/j.jbankfin.2010.06.001
[18] Kou, G., Xu, Y., Peng, Y., & Shen, F. (2021). Machine learning in financial risk management: A literature
review. International Review of Financial Analysis, 71, 101422. https://doi.org/10.1016/j.irfa.2020.101422
[19] Kotsantonis, S., & Serafeim, G. (2019). Four things no one will tell you about ESG data. Journal of Applied
Corporate Finance, 31(2), 50–58. https://doi.org/10.1111/jacf.12346
[20] Lessmann, S., Baesens, B., Seow, H. V., & Thomas, L. C. (2015). Benchmarking state- of-the-art classification
algorithms for credit scoring: An update of research. European Journal of Operational Research, 247(1), 124–
136. https://doi.org/10.1016/j.ejor.2015.05.030
[21] Li, Y. (2017). Deep reinforcement learning: An overview. arXiv preprint arXiv:1701.07274.
https://arxiv.org/abs/1701.07274
[22] Li, F., Wang, C., & Xu, G. (2020). ESG and corporate financial performance: Evidence from China.
Sustainability, 12(19), 7916. https://doi.org/10.3390/su12197916
[23] Lu, J., Liu, A., Dong, F., Gu, F., Gama, J., & Zhang, G. (2018). Learning under concept drift: A review. IEEE
Transactions on Knowledge and Data Engineering, 31 (12), 2346–2363.
https://doi.org/10.1109/TKDE.2018.2876857
[24] Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7 (1), 77–91.
https://doi.org/10.2307/2975974
[25] Merton, R. C. (1974). On the pricing of corporate debt: The risk structure of interest rates. The Journal of
Finance, 29(2), 449–470. https://doi.org/10.2307/2978814
[26] Murphy, K. P. (2022). Probabilistic Machine Learning: An Introduction. MIT Press.
[27] Rudin, C. (2019). Stop explaining black box machine learning models for high-stakes decisions and use
interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. https://doi.org/10.1038/s42256-019-
0048-x
[28] Saunders, A., & Allen, L. (2020). Credit risk management in and out of the financial crisis: New approaches
to value at risk and other paradigms (3rd ed.). Wiley Finance.
[29] Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). MIT Press.
[30] Zhou, Z., Pan, Y., Wang, H., & Wang, B. (2021). A hybrid artificial intelligence model for credit risk
evaluation. Expert Systems with Applications, 183, 115348. https://doi.org/10.1016/j.eswa.2021.115348