1770
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
PREDICTING FINANCIAL CRISIS: THE POTENTIAL OF AI IN GLOBAL MARKETS
Dilshoda Juraboeva Komolxon qizi
Juraboyevadilshoda75@gmail.com
Murodjon Sagdiddinov Rahimjonovich
Students, University of World Economy and Diplomacy
Malikov Numonjon Kamalovich
Scientific supervisor:
Lecturer at the University of World Economy and Diplomacy
e-mail:
University of World Economy and Diplomacy.
https://doi.org/10.5281/zenodo.15284249
Abstract.
This paper reviews how artificial intelligence (AI) and machine learning (ML)
techniques have been incorporated into the prediction of financial crises and evaluates their
performance relative to econometric models. Most of the existing literature has relied on macro-
financial indicators and regression approaches, but the inability to solve nonlinearities, regime
changes, and high-dimensional datasets has motivated the use of AI techniques. The review
surveys the conceptual and empirical literature of financial crisis forecasting, discusses model
classes, and reports new developments in deep learning, hybrid models, and data fusion. Special
emphasis is placed on the use of supervised and unsupervised learning, recurrent neural
networks (RNN), long short-term memory (LSTM) networks, and transformer architectures,
alongside the use of alternative data such as sentiment analysis and media narratives. A
separate section assesses the use of scenario-driven geopolitical stress testing in portfolio risk
management. In conclusion, the review describes the gaps in methodology and develops new
avenues for research in model credibility, generalization across countries and cultures, and
real-time update systems for forecasts. This work enhances the academic discourse around crisis
forecasting while also enabling financial authorities, institutional market players, and policy
decision-makers to devise tools to alert them of potential crises in the context of modern
sophisticated and globalized financial systems.
Keywords:
Financial crisis prediction, Artificial intelligence (AI), Machine learning
(ML), Deep learning, Early warning systems, Scenario-based stress testing, Portfolio risk
management, Recurrent neural networks (RNN), Long short-term memory (LSTM), Geopolitical
risk forecasting, Nonlinear modeling, High-dimensional data, Explainable AI (XAI).
ПРОГНОЗИРОВАНИЕ ФИНАНСОВОГО КРИЗИСА: ПОТЕНЦИАЛ ИИ НА
МИРОВЫХ РЫНКАХ
1771
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
Аннотация.
В этой статье рассматривается, как методы искусственного
интеллекта (ИИ) и машинного обучения (МО) были включены в прогнозирование
финансовых кризисов, и оценивается их эффективность по сравнению с
эконометрическими моделями. Большая часть существующей литературы опиралась на
макрофинансовые индикаторы и регрессионные подходы, но неспособность решать
нелинейности, изменения режимов и многомерные наборы данных мотивировала
использование методов ИИ. В обзоре рассматривается концептуальная и эмпирическая
литература по прогнозированию финансовых кризисов, обсуждаются классы моделей и
сообщаются новые разработки в области глубокого обучения, гибридных моделей и
слияния данных. Особое внимание уделяется использованию контролируемого и
неконтролируемого обучения, рекуррентных нейронных сетей (RNN), сетей с
долговременной краткосрочной памятью (LSTM) и архитектур трансформаторов, а
также использованию альтернативных данных, таких как анализ настроений и медиа-
нарративы. В отдельном разделе оценивается использование геополитического стресс-
тестирования на основе сценариев в управлении рисками портфеля. В заключение обзор
описывает пробелы в методологии и разрабатывает новые направления для исследований
в области достоверности моделей, обобщения по странам и культурам, а также систем
обновления в реальном времени для прогнозов. Эта работа усиливает академический
дискурс вокруг прогнозирования кризисов, а также позволяет финансовым органам,
институциональным участникам рынка и лицам, принимающим политические решения,
разрабатывать инструменты для оповещения о потенциальных кризисах в контексте
современных сложных и глобализированных финансовых систем.
Ключевые слова:
Прогнозирование финансовых кризисов, Искусственный
интеллект (ИИ), Машинное обучение (МО), Глубокое обучение, Системы раннего
оповещения, Стресс-тестирование на основе сценариев, Управление портфельными
рисками, Рекуррентные нейронные сети (RNN), Долгосрочная краткосрочная память
(LSTM),
Прогнозирование геополитических рисков, Нелинейное моделирование,
Высокомерные данные, Объяснимый ИИ (XAI).
Introduction
Perhaps the most debilitating phenomena in the economy world over is a financial crisis.
This is because they lead to acute exits on macroeconomic activity, wide-ranging
volatility in the financial markets, and long-lasting socio-economic ramifications. As noted by
Reinhart and Rogoff (2009), a financial crisis surfaces with a break down in one or more systems
within the infrastructure network along side a glaring drop in the value of equities, capital market
1772
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
funding, government expenditures, and a rise in the number of bankrupt corporations. This is
synonymous with the International Monetary Fund (IMF)'s definition of the term. As per IMF
(1998), a financial crisis is a situation where one or more components of the financial system
malfunction and require government funding to intervene. Development financial institutions
also take a vantage point as World Bank identifications of financial crises into distinct categories
such as, but not limited to, banking crises, currency crises, and sovereign debt crises deeming
more devastating consequences to the economy whether domestic or international (Laeven &
Valencia, 2020).
The world has suffered because of severe financial crises in the past. Consider, for
example, the Great Depression of 1929. It originated from the U.S. stock market crash and
resulted in a 15% reduction in GDP along with widespread inflation. Another Financial crisis
took place in ‘97 in Thailand, when the Baht collapsed. After its initial failure, the currency
affected the rest of East Asia, leading to a 6.7% reduction in Indonesia’s GDP and 10.5% in
Thailand by ‘98. The Great Recession of 2008 in the U.S. was another financial disaster that
resulted in $15 trillion loss of wealth globally. This also caused a dip in GDP by 2.1% in 2009,
according to World Bank. Moving forward, the COVID-19 pandemic induced the worst global
economic downturn since 1939, shrinking global GDP by 3.1% and leading to unprecedented
financial turmoil (IMF 2021).
An accurate and timely prediction of financial crises continues to be an elusive goal,
despite decades of focus on the area. There is no doubt that traditional econometric models, as
well as early warning systems such as the signals approach (Kaminsky, Lizondo & Reinhart,
1998) and the probit-based models (Berg & Pattillo, 1999) have showed a considerable level of
understanding regarding the sensitivity of economies. They however tend to be overly simplistic
and rigid with regards to underlying framework; model assumptions, variable selection, and
intricate nonlinear relationships characteristic of financial systems. What is more problematic is
the fact that those models tend to produce false positives or, in many cases, not enough time to
preemptively address the situation.
The application of artificial intelligence (AI)for financial forecasting is relatively new,
but it has the potential to auger new prospects in recent years. Machine learning (ML), deep
learning (DL), and natural language processing (NLP) techniques have the capability to analyze
massive, high-dimensional datasets, detect subtle intricate patterns, and adapt to the ever-
evolving economic landscape. Industry heavyweights like JPMorgan Chase and BlackRock are
starting to deploy AI models for market risk analysis. Academic research also emphasizes the
topic, such as studies conducted by Chen et al. (2021) and Alessi et al. (2023), which have
shown that AI-powered early-warning systems outperform traditional systems in terms of
1773
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
predictive performance. These models are capable of incorporating large heterogeneous datasets
ranging from macroeconomic indicators to social media sentiments while improving
continuously through self-learning algorithms.
Conceptual and Empirical Foundations
Due to the disruptive effects of financial crises, accurately predicting them has been an
enduring goal in economic theory and policy. Crises are often described as system shocks that
disrupt the operation of financial markets and institutions on a systemic scale. Reinhart and
Rogoff (2009) identify three types of crises: banking, currency, and sovereign debt, each marked
by plummeting asset prices, bankrupt financial institutions, and severe economic dislocation.
The International Monetary Fund (1998) also defined financial crises as events that
involve the substantial breakdowns of financial intermediaries’ capital allocation functions with
the presence of solvency constraints and policy controls. Laeven and Valencia (2020) carry out
the most recent and comprehensive empirical work by setting up standards for classifying crises
that include deposit withdrawals, substantial deposits from other entities into the institution, and
government control of the entity.
Inherent challenges stem from the nonlinear, infrequent nature of crises, particularly for
those interested in forecasting. These non-frequent events typically happen due to underlying
changes and structural breaks not accounted for in traditional economic frameworks. The lack of
data due to gaps occurring with the slow build-up of systemic risks increases model fragility and
instability. Resulting from the interplay of investor sentiment, asset prices, and leverage, crises
occur through complex feedback loops that are best described by nonlinear systems, rather than
time-fixed or linear designs.
The multifaceted aims of predicting financial crises are as follows. Firstly, the early-
warning detection system makes attempts to identify vulnerabilities within a system before they
lead towards a crisis so that policymakers are able to take action beforehand. Secondly,
estimating the likelihood of a crisis occurrence at a certain predetermined time is referred to as
probability forecasting, assisting in the calibration of macroprudential tools. Finally, systemic
risk detection identifies interdependency networks and dependencies known as contagion that
poses risks to localized shocks which can in turn render global devastation.
Problematical Econometric Models
There are basic approaches towards the traditional models such as structural or statistical
models that make use of a set of macro-financial indicators. Signals approach by Kaminsky,
Lizondo, and Reinhart (1998) lean towards using the most prominent and impactful frameworks
provided. Early warning systems use indicators that possess benchmarks for ranges and limits.
1774
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
The underlying reason as to why empirical crisis research utilize the signals method is
primarily due to its interpretability alongside simplicity.
As mentioned earlier, Berg and Pattillo (1999) used logit and probit regressions to
estimate the conditional probability of crises. These models attempt to explain the occurrence of
crisis episodes with some explanatory factors such as the credit to GDP gap, foreign exchange
reserves, asset price inflation, and fiscal deficits. Like most models, they possess some degree of
statistical credibility alongside policy-relevant conclusions. However, they also tend to have
some shortcomings due to their linearity assumptions and sensitivity to model specification.
More advanced methodologies have also been adopted. Markov-switching models
(Hamilton, 1989) deal with regime shifts within macro-financial time series, marking shifts from
tranquil to crisis states. Other simulations have used Dynamic Stochastic General Equilibrium
(DSGE) models and Vector Autoregressive (VAR) models in order to study the inter-shock
transmission within the financial system and evaluate the policy responses to them. While these
models offer a more endogenous approach in dealing with the dynamics of crises, they tend to be
highly calibrated, which limits their real-time adaptability.
While beneficial, traditional models neglect several areas of concern. They are low-
performing under real-time conditions, highly calibrated which conceals their adaptability,
overly simplistic in design which limits their dynamics and reduces predictive power, reliant on
infrequent and aggregated indicators that swiftly change, and provide delayed warning signals.
The Rise of AI and Machine Learning in Crisis Prediction
Due to shortcomings in conventional models, there is renewed focus on using artificial
intelligence (AI) and machine learning (ML) approaches for crisis forecasting. Such approaches
are particularly effective with high-dimensional, non-linear, or non-static data sets which are
often found in financial systems.
Chenetal. (2021) used ensemble tree-based models constructed from macro-financial data
in an attempt to improve out-of-sample accuracy forecasting banking crises. Their results
indicated marked improvement over traditional techniques. Other researchers have employed
supervised learning techniques like Decision Tree, Random Forest, Support Vector Machine
(SVM), and Gradient Boosting Machine also known as XGBoost to distinguish between periods
of declared financial stability and those marked by crisis.
Time-series problems have seen an increase in the application of deep learning methods,
such as Recurrent Neural Networks (RNN) and Long Short Term Memory (LSTM) networks.
These architectures excel at capturing intricate contextual relationships, enduring lots of
noise where timing and the arrangement of inputs matter.
1775
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
Zhang and Zhang (2022) showcased the capabilities of LSTM for predicting sovereign
debt defaults in emerging markets, significantly outperforming traditional time-series models in
accuracy and lead time.
New Transformer-based approaches are being investigated in the financial domain,
especially in multi-variable forecasting and high-frequency data analysis. The advancements
made in natural language processing using these models indicates their potential in analyzing
complex financial sequences coupled with cross-referenced data.
Blended approaches, or hybrids, that include AI methods with econometric structures
attempt to resolve the tension between inferential statistics and predictive power. Some
researchers structure integrate macroeconomic predictors using VAR models and then classify
them using AI methods, dynamically preserving the economic hierarchy while increasing
flexibility.
In AI driven studies, the datasets utilized are becoming more eclectic. Traditional
economic data is now being integrated with nontraditional indicators such as ESG metrics,
sentiment analysis, media coverage, and Google search trends. These models have also been
applied during various crises like the global financial crisis of 2008, the European sovereign debt
crisis from 2010 to 2012, and during pandemic-related financial turbulence.
AI Models Traditional Models Evaluation
Cross-sectional assessments have shown that AI and machine learning are more efficient
than traditional economy-based models. In comparison to their traditional econometric
counterparts, AI models yield better results when it comes to accuracy, recap, and precision
metrics. The rate of false positive results particularly in tests outside the sample range is also
much lower. It has been proven that AI models provide much greater advanced warning periods,
enabling earlier detection of a financial crisis (Alessi et al., 2023).
Furthermore, AI models are more capable of dealing with economic systems that are
high-dimensional and complex, making them more adaptable to changes within the economy.
Unlike traditional models that work using a predetermined set of functions, formulas, and
variables, machine learning models can find patterns and interactions that no one has manually
coded using predetermined boundaries.
Regrettably, these benefits come with a set of equally balancing disadvantages. Machine
learning algorithms, for instance, are often viewed as “black boxes” that offer little to no
explanation on how decisions were reached, hindering the adoption of such frameworks in
decision-making contexts like policymaking. In addition, overfitting, which is more likely to
happen in small-sample settings with observations of few crises, can severely affect the
generalizability of a model.
1776
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
There are also issues with data dependence, particularly, the type of data that many of
these AI models need to work with, such as vast amounts of historical or high-frequency data,
which is a luxury that not all economies or types of crises can afford.
Imposing benchmark studies has become commonplace for international institutions like
the IMF, BIS, and the OECD as they attempt to juxtapose AI and traditional models. Yet, there
continues to be a lack of established guidelines to guide these assessments in a coherent and
unified way. These comparisons tend to be extremely one-sided, analyzing single countries or
conduct policy-irrelevant retrospective simulations that diminish their overall usefulness.
Thematic Gaps and The Next Steps in Research
Even with the developments in methods, there continues to be a lack of literature in
certain areas. To start, the analyses differ significantly when it comes to defining and dating a
crisis, making repetition and generalization more challenging. Next, most models tend to
underperform cross-country; they do well in one region but poorly when used in a different
context. Finally, there is a lack of real-time forecasting; most research done relies on ex-post
evaluations that are not bound by the constraints present in reality.
On the other hand, the more advanced the AI models become, the more behavioral and
qualitative components like sentiment, media, and trust in institutions get overlooked. By adding
these factors, the ability of these models to respond to severe damage will dramatically increase.
It's important to note that there is little to no work done using AI with formal delineation
explainability frameworks such as SHAP (SHapley Additive ExPlanations) or LIME (Local
Interpretable Model-agnostic Explanations), which are crucial for regulatory and policy spend
trust. Notably, there is an absence of analysis from multiple countries that work alongside
multiple crises and consider the level of integration present in the global financial markets.
1777
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
With these gaps in mind, further research should concentrate on the following:
- Increasing validation done across different countries and outside the scope of the study
-Adding datasets that are not strictly economic and alternative behavioral datasets with
high frequency.
- Utilize XAI tools for explainability.
- Create flexible models that deal with numerous types of crises in different regions and
periods.
In fixed-income markets, scenario impacts frequently change sovereign risk spreads,
particularly in emerging markets that are vulnerable to shifts in commodity exports or
geopolitical alignments.
For this latter situation, tail-risk behavior tends to dominate. Unlike traditional othering
approaches, which are based on the assumption of stable correlations, trying to analyze
relationships under geopolitical stress reveals correlation breakdowns and tail clustering
evasions.
Consider for instance:
- An exacerbated Taiwan Strait conflict podría lead to region-wide sell-offs in East Asian
stocks and US tech stocks that simultaneously inject cashells into US Treasuries and gold
making them “safe-haven assets.”
- An embargo on Middle Eastern oil is likely to create enormous inflationary difficulties
for equities and long-duration bonds which undermines assumed historical hedging relationships.
These various approaches suggest implementing geographical reallocations, realigning
sector and hedge positions, and shifting national currencies, as well as adding real assets,
commodities betting, or volatility indices based on (geo)politics into recess. Advanced election
cycle dynamic models capture these observations by adding hypothesis “alert thresholds” like
NATO DEFCON levels or sanctions on energy market frameworks within their exploiters.
Methodological Strengths and Structural Constraints
The use of scenario-based geopolitical stress testing offers several methodological and
strategic advantages. First, it enables forward-looking risk management that anticipates rather
than reacts to emerging threats. Second, the approach is tailored to capture asymmetric risks—
events with significant downside potential and limited historical precedent. Third, it provides a
mechanism to incorporate expert knowledge, essential for risks that are poorly understood by
markets or inadequately priced.
1778
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
Nevertheless, the framework has its weaknesses. Scenario development is still an
estimation exercise based on some predefined behavioral heuristics regarding the actor, pathway
of escalation, and institutional reaction. It is difficult—if not impossible—to validate these
scenarios, as estimating their financial impacts is intricate and often devoid of historical
precedent.
Additionally, presenting the results of the scenarios to internal audiences or regulatory
boards is often difficult because the output requires telling a story that is inherently fuzzy in
defining probabilities.
From a practical point of view, the application of such models requires understanding
both the geopolitical arena and quantitative modeling, which means they can only be used by
large asset managers, central banks, or sovereign wealth funds. Smaller institutions may find it
challenging to implement the required analytical architecture without outside help or
collaborations.
New Boundaries: Emerging Trends and Institutional Adoption
With geopolitical risks becoming more systemic and multi-faceted, scenario-based stress
testing is expanding to capture new domains such as competition between states and intertwining
dependencies across the globe. Some noteworthy new areas include:
-Climate and water diplomacy: There is increasingly aggressive geopolitical competition
over transboundary water rights (like the Nile Basin or Himalayan glaciers). Portfolios with
agriculture, infrastructure, or EM debt exposure are being stress-tested against resource-driven
instability.
1779
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
-The geopolitics of rare earth and strategic minerals: Gainful competition for lithium,
cobalt, and rare earth elements necessary for energy transitions are adding geopolitical
uncertainty to ESG portfolios and green technology supply chains.
- Cyber sovereignty and military application of artificial intelligence: The DoD’s
increasing cyber warfare initiatives accompanied by the decomposition of AI and data
frameworks present new tail end risks to the financial, defense, and technological industries.
- Diplomacy of currency fragmentation and debt: The proliferation of alternative financial
systems like China’s belt and road lending or BRICS digital currencies pose new risks regarding
currency order, capital flow control, and global reserve currency frameworks.
Adoption by institutions is on the rise. BlackRock developed proprietary dashboards
measuring geopolitical risk and scenario trees for tactical asset allocation. Gulf, East Asia, and
Scandinavian SWFs are setting up internal geopolitical analytics units aimed at bolstering
informed long-term investment planning. The UN Principles for Responsible Investment (UN
PRI) is starting to acknowledge geopolitical sustainability risks by urging for scenario planning
in ESG strategies framework peripheral to investment politics.
These changes underscore the growing recognition that financial risk models require a
change from historical correlations and reliance on statistical inferences. Combining strategic
foresight with probabilistic modeling through scenario-based geopolitical stress testing creates
opportunity for preemptive adaptative risk management.
Conclusion
Over the past thirty years, the foresight of financial catastrophe has shifted from
econometric models to more complex and machine learning approaches. While the first
generation of predictive devices including advanced signal extraction techniques, early warning
systems, logit and probit models pioneered forecasting financial crises, they were unable to
provide timely crisis detection due to the need for real-time adaptability, complexity, and non-
linear dynamic handling.
In this scenario, artificial Intelligence and Machine Learning have proven to be effective
adjunct approaches. Diverse datasets are now handled with ease by supervised learning models
such as Random Forests and XGBoost, deep learning architectures like LSTMs and
Transformers, and hybrid econometric-AI frameworks which have significantly improved
precision in prediction and time. Particularly, these enhancements are critical considering the
fast-paced, interconnected, rich in data, nature of world financial systems.
Despite those advancements, the literature is still incomplete. From a regulatory and
policymaking perspective, AI tends to lack design interpretability which significantly impedes its
applicability.
1780
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
So many studies seem to be constrained by retrospective evaluation designs, which raises
issues around real-time usability. The gaps in the definition of crises, absence of data in
infrequent contexts, and the insufficient behavioral and geopolitical integration tend
Recommendations for the Future
1. Achieve Generalizability Across Borders and Multi-Crisis Scenarios
Attention should be devoted to developing training datasets from multiple countries and
spanning varying periods of crisis for future AI-based models, as such measures would mitigate
the current overfitting to particular economic environments.
2. Combine Behavioral Data with Other Non-Traditional Data Sources
The models could benefit from the monitoring of real-time public sentiment through
news outlets, social media, search engines, and ESG activities. These factors often accompany or
precede financial market distress. Qualitative and quantitative data streams need to be blended
for effective modeling.
3. Enhancing Explainability for AI Models Used in Forecasting
More advanced deep learning models are bound to require tools for explainable AI,
SHAP and LIME, which help explain the model’s predictions based on a subset of features.
These defined borders will increase the utility of such models in under scrutiny
governance and policy decision frameworks.
4. Integrated Systems for Dynamic and Real-Time Forecasting
Using high-frequency and streaming data, future systems should shift focus towards real-
time monitoring capabilities and dynamically updating model parameters for pre-emptive
warning during heightened systemic risk.
5. Scenario-Based Simulation Platforms
By incorporating AI-based predictions with scenario planning stress testing tools, it is
possible to capture the economically damaging effects of geopolitical events (like energy
embargoes, armed conflicts, cyber attacks and other disruptions), which these platforms should
incorporate expert judgment such as probabilistic reasoning through Bayesian networks or fuzzy
logic paired with stochastic simulation techniques.
6. Regulatory Collaboration and AI Governance Frameworks
Cooperative work of policymakers and financial institutions is necessary in the context of
developing standards of AI governance such as validation requirements, ethics policies, and
systemic risk oversight boundaries for responsible utilization of predictive technologies.
1781
ResearchBib IF - 11.01, ISSN: 3030-3753, Volume 2 Issue 4
REFERENCES
1.
Beutel, J., List, S., & von Schweinitz, G. (2019
Does machine learning help us predict
banking crises? - ScienceDirect
? Journal of Financial Stability, 45, 100693.
2.
Zhang, H., & Zhang, Y. (2020)
Predicting systemic financial crises with recurrent neural
3.
Li, J., Xu, C., Feng, B., & Zhao, H. (2023).
Credit Risk Prediction Model for Listed
Companies Based on CNN-LSTM and Attention Mechanism
4.
Alessi, L., Detken, C., & Lo Duca, M. (2023).
Credit growth, the yield curve and financial
crisis prediction: Evidence from a machine learning approach - ScienceDirect
International Money and Finance, 132, 102759.
5.
Predicting Fiscal Crises: A Machine Learning Approach in:
IMF Working Papers Volume 2021 Issue 150 (2021)
IMF Working Paper No. 2021/150.
6.
Zhang, W. (2020).
Thesis_Final_Version_Wenke_Zhang.pdf
7.
Han, J., & Zhang, H. (2021).
A survey on deep learning for financial risk prediction
Quantitative Finance and Economics, 5(3), 500–522.
8.
Ghosh, S., & Ghosh, S. (2020).
Predicting sovereign debt crises using artificial neural
networks: A comparative approach | Request PDF
9.
Fitch Ratings. (2023).
EM Sovereign Defaults at Record Level, but Rating Outlooks More
10.
Cheng, D., Niu, Z., Zhang, J., & Jiang, C. (2023). Critical Firms Prediction for Stemming
Contagion Risk in Networked-Loans Through Graph-Based Deep
