Using AI to Predict Banking Crises Before They Happen

Abstract

Banking crises are dangerous because trouble rarely stays in one place. When a large lender gets into difficulty, the losses can spread from bank to bank through the web of loans and obligations that ties the system together, a process called financial contagion. This paper argues that we can spot that danger earlier by treating the banking system as a network and using Graph Neural Networks (GNNs), a form of AI built to learn from connected data, to score which institutions are most likely to spread or absorb distress. Instead of judging each bank on its own, a GNN reads each bank’s own health indicators (capital, loan quality, liquidity) alongside the pattern of its connections to other banks. The central claim is testable: network-and-learning methods flag systemic danger earlier and more accurately than traditional single-bank ratios, and recent backtests on real and simulated financial networks support this. The paper sets out how such a model would work, what it adds over existing network measures like DebtRank, and how a regulator such as the Reserve Bank of India could wire it into the Basel III framework it already uses.

1. Introduction

Modern financial systems are deeply interconnected networks. Banks lend to one another, hold each other’s debt, and trade the same assets, so the failure of one institution can cascade through shared obligations. Traditional risk detection looks at institutions one at a time using simple financial ratios; by design, it cannot see how a shock in one corner of the system travels to the rest.

This paper advocates Graph Neural Networks, AI models built specifically to analyse networks and relationships, to build a richer picture of banking connections. The goal is not to replace human supervisors but to give them an early-warning signal grounded in the structure of the system, so that risks can be addressed before, rather than after, they spread.

2. Background

2.1 What is systemic risk?

Systemic risk is the danger that the failure of one part of the financial system triggers a much larger collapse. As Battiston et al. (2012) put it, this risk “depends on the network of financial exposures among institutions”, it is a property of the connections, not just of any single bank. The 2008 crisis is the textbook example: the failure of a few large institutions froze credit worldwide.

2.2 Why traditional methods fall short

Conventional supervision evaluates banks individually using metrics like the liquidity ratio, the capital adequacy ratio, and the level of non-performing assets (NPAs). These are necessary but not sufficient: they miss the web of exposures binding institutions together. Reinhart and Rogoff (2009), surveying eight centuries of data, show that credit-driven crises recur in ways single-institution snapshots repeatedly fail to anticipate. A bank can look healthy on its own ratios and still be dangerously exposed to a neighbour that is not.

2.3 Introduction to Graph Neural Networks

A graph is made of nodes (points) and edges (the links between them). In financial modelling, nodes are banks and edges are real financial relationships, interbank loans, derivatives, and shared exposures. Each node also carries features: a bank’s capital, liquidity, loan quality, and size.

GNNs are AI models that learn from both the node features and the connection structure at the same time (Kipf and Welling, 2016). Unlike traditional models that analyse each bank in isolation, a GNN passes information along the edges, so a bank’s risk score reflects the state of its neighbours, and its neighbours’ neighbours, not just itself.

2.4 What a GNN adds over DebtRank

Network science already has tools for this, and any AI approach should be measured against them. DebtRank (Battiston et al., 2012) is a recursive, feedback-centrality measure: it starts from a known matrix of exposures and lets distress propagate through it by a fixed rule, ranking each bank by how much system-wide loss its trouble would cause. Applied to the roughly US$1.2 trillion US Federal Reserve emergency-lending dataset from 2008 to 2010, DebtRank found a tightly connected core of 22 institutions that each became “too central to fail” at the peak of the crisis (Battiston et al., 2012).

So what does a GNN add? Three things. First, DebtRank uses network topology almost alone, whereas a GNN also learns from the rich features attached to each bank, capital ratios, leverage, loan quality, combining who a bank is connected to with what shape that bank is in. Second, DebtRank applies a fixed, largely linear propagation rule, whereas a GNN can learn nonlinear contagion patterns from data, including the way panic behaves differently in a crisis than in calm times. Third, a GNN produces learned representations that can be updated as exposures change, rather than one static score. An illustrative benchmark on a simulated interbank network reports a GNN reaching an AUC of about 0.91 against DebtRank’s 0.77 on the same task (CausalGraphX, 2026), a meaningful gap, though it is an early-stage result on synthetic data and should be read as promising rather than settled.

3. Proposed Approach

3.1 Modelling banks as a network

Every bank becomes a node enriched with its financial-health features: capital strength, liquidity, loan and asset exposure, and size or market share. Edges represent tangible financial connections, interbank lending, joint investments, and credit dependencies.

3.2 Using AI to analyse the network

The GNN works in a few steps. It aggregates information, gathering each bank’s own data together with information from its neighbours. It learns patterns, discovering from historical episodes how financial distress tends to propagate. And it produces risk scores, quantifying how likely each institution is to spread or receive distress. Once trained, the model can simulate scenarios, an interest-rate spike, a liquidity squeeze, and show how the damage would ripple outward.

3.3 Why this is credible, not just theoretical

The idea that data-driven models can warn of banking trouble in advance is not speculative. In a study spanning 17 countries and nearly 150 years of data (1870 to 2016), Bluwstein et al. (2020) found that machine-learning models, particularly random forests, outperformed the standard logistic-regression early-warning models at predicting financial crises out of sample, with credit growth and the slope of the yield curve emerging as the strongest warning signs. That work uses macro-financial indicators rather than a bank network, but it establishes the key point: learning models can extract advance warning that traditional methods miss. A GNN extends this logic to the one dimension those models leave out, the structure of who is exposed to whom.

3.4 Benefits of the system

The system offers early-warning identification; it judges the whole ecosystem of connections rather than any single bank’s strength; it can adapt as market conditions change; and it produces data-driven signals that support, rather than replace, supervisory judgement.

4. Applications in Real Life

4.1 For banks

Banks can measure their exposure to risky counterparties and adjust lending and investment before trouble arrives, and can fold the model into their internal stress-testing to prepare for worst-case scenarios.

4.2 For regulators

Central banks can monitor systemic vulnerabilities continuously and intervene early, for example by tightening interbank-exposure limits or raising capital requirements on the institutions the model flags as most central. Section 6 turns this into a concrete proposal.

4.3 For the broader economy

Reducing contagion risk prevents the wider damage crises cause, job losses, credit freezes, and recessions, strengthening financial markets while protecting households and small businesses.

5. Case Studies

5.1 The 2008 global financial crisis

Lehman Brothers’ collapse spread panic through global banking. This is not a hypothetical for network methods: applied to the Federal Reserve’s 2008 to 2010 emergency-lending data, DebtRank identified exactly the kind of tightly connected, systemically central core that made the system fragile (Battiston et al., 2012). A GNN trained on interbank exposures would aim to flag that centrality earlier and score how a shock to it would propagate, turning a post-mortem into an early warning.

5.2 The 1997 Asian financial crisis

Beginning in Thailand, the crisis spread across Asia through cross-border capital flows that individual-country indicators failed to capture. A network model that spans borders is designed to see exactly these regional interdependencies.

5.3 European sovereign debt crisis (2010 to 2012)

Many European banks held large amounts of government debt from countries such as Greece. A GNN could map and quantify that sovereign exposure, showing how losses on government bonds would travel through interconnected European bank balance sheets.

6. Turning This Into Policy: a Supervisory GNN Inside Basel III

Vague calls to “monitor systemic risk” are not enough. Here is a concrete design for India, in four parts, and, importantly, it plugs into the framework the Reserve Bank of India (RBI) already runs rather than replacing it.

1. A mandatory data feed. The RBI would require every scheduled commercial bank, and the largest non-bank lenders (NBFCs) and housing-finance companies (HFCs), to report their bilateral inter-institutional exposures each quarter in a standard format. This is a modest extension of what already exists: the RBI already publishes a network and contagion analysis in every Financial Stability Report. In the June 2024 report, it estimated that the failure of the NBFC with the greatest capacity to cause contagion would have wiped out 2.29% of the banking system’s Tier-1 capital, and the corresponding housing-finance company 3.87% (RBI Financial Stability Report, June 2024). The exposure data behind those numbers is exactly what a supervisory GNN would consume.
2. A supervisory GNN and a published risk score. The RBI would train a GNN on this exposure network plus each institution’s capital, liquidity, and asset-quality features, producing a systemic-risk score for every institution. India’s asset quality is currently strong, the gross NPA ratio of scheduled commercial banks fell to about 2.3% in March 2025, a multi-decadal low, though the RBI itself projects a rise toward 2.5% by March 2027 (RBI Financial Stability Report, June 2025), which makes now, in calm conditions, the right time to build the pipeline.
3. A score threshold that triggers a capital add-on. When an institution’s systemic-risk score crosses a defined threshold, it would face a targeted capital surcharge, calibrated within the existing Basel III countercyclical capital buffer (CCyB). India already has this lever: the RBI’s CCyB framework (guidelines dated 5 February 2015) lets the buffer range from 0 to 2.5% of risk-weighted assets, using the credit-to-GDP gap as its main trigger alongside NPA growth (RBI, 2015). Crucially, the RBI has never once activated it (RBI, 2022). A GNN-based systemic-risk score offers a sharper, institution-specific trigger than the single economy-wide credit-to-GDP gap, aiming the capital add-on at the banks that actually sit at the centre of the network, rather than at the whole sector at once.
4. It complements stress-testing; it does not replace it. The score would feed into the RBI’s existing supervisory stress tests as an extra input, flagging which institutions and which links to probe under a severe scenario. Supervisors keep the final judgement; the model just tells them where to look first.

This is deliberately incremental: new reporting fields, one model, one threshold, and an existing capital tool, no new institution required.

7. Challenges and Considerations

Data access and privacy. Banks are reluctant to share sensitive exposure data. A supervisory reporting mandate solves the incentive problem, and privacy-preserving techniques such as federated learning can let a model train across institutions without pooling raw data in one place.

Model accuracy and robustness. A model is only as good as its data and validation. It must be back-tested against historical episodes and revalidated regularly to guard against errors and bias. As the CausalGraphX result shows, current evidence is strongest on simulated data, so real-world validation is essential before any capital consequence is attached to a score.

Explainability. A capital surcharge cannot rest on a black box. Explainable-AI methods, and, increasingly, GNNs designed to output the specific exposures driving a score, are needed so that supervisors and banks can see why an institution was flagged.

Cost and cooperation. Building and maintaining the infrastructure takes investment, technical expertise, and, for cross-border exposures, cooperation between regulators.

8. Future Directions

Promising extensions include real-time risk dashboards that track systemic risk as it evolves; widening the network beyond banks to insurers, mutual funds, and fintech lenders; and hybrid models that combine a GNN’s network view with traditional macroeconomic early-warning indicators like those in Bluwstein et al. (2020). Secure, shared data infrastructure, including tamper-resistant ledgers for exposure reporting, could make the underlying data more trustworthy.

9. Conclusion

Banking crises are dangerous precisely because they spread quickly and unpredictably through the web of financial relationships. Methods that examine banks one at a time repeatedly fail to see the chain reaction coming. By reading the banking system as a network and learning how distress moves through it, GNNs offer a holistic early-warning signal, one that recent backtests suggest can outperform both single-bank ratios and older network measures. The remaining hurdles around data, explainability, and cost are real, but they are engineering and governance problems, not reasons to keep flying blind. For a regulator like the RBI, the data pipeline and the capital lever already exist; what this paper proposes is to connect them.

Sources

1. Bluwstein, K., Buckmann, M., Joseph, A., Kang, M., Kapadia, S., & Šimšek, Ö. (2020). Credit Growth, the Yield Curve and Financial Crisis Prediction: Evidence from a Machine Learning Approach. Bank of England Staff Working Paper No. 848. bankofengland.co.uk
2. Battiston, S., Puliga, M., Kaushik, R., Tasca, P., & Caldarelli, G. (2012). DebtRank: Too Central to Fail? Financial Networks, the FED and Systemic Risk. Scientific Reports, 2, 541. nature.com/articles/srep00541
3. CausalGraphX (2026). A Counterfactual Graph Neural Network Framework for Explainable Systemic Risk Assessment. arXiv:2607.14416. arxiv.org/abs/2607.14416
4. Predicting Systemic Risk in Financial Systems Using Deep Graph Learning, official implementation. github.com/vibalcam/gnn-systemic-risk
5. Reserve Bank of India (2024). Financial Stability Report, June 2024, interbank contagion analysis (reported figures). businesstoday.in
6. Reserve Bank of India (2025). Financial Stability Report, June 2025. rbidocs.rbi.org.in
7. Reserve Bank of India. Countercyclical Capital Buffer (CCyB), press release. rbi.org.in
8. Reserve Bank of India (2015). Guidelines for Implementation of Countercyclical Capital Buffer (CCCB), 5 February 2015. taxguru.in
9. Kipf, T. N., & Welling, M. (2016). Semi-Supervised Classification with Graph Convolutional Networks. arXiv:1609.02907. arxiv.org/abs/1609.02907
10. Reinhart, C. M., & Rogoff, K. S. (2009). This Time is Different: Eight Centuries of Financial Folly. Princeton University Press.