人工知能
Can AI Warn Investors Before the Next Financial Crisis?
Securities.io を Google の優先ソースに追加
At the core of the recent progress in artificial intelligence is an increasing ability to sort out complex patterns from unorganized data, something previous generations of software struggled with.
This opens many new avenues for the application of AI, from medicine to many kinds of forecasting. So, unsurprisingly, many are starting to apply it to financial questions, especially predicting financial crises.
This is because while financial crises may appear sudden, in hindsight, some data points often reveal the build-up to this point. Financial crises are almost always the result of years of imbalance, poor investment decisions, overvaluation, etc.
So maybe an AI system able to process a massive amount of data and make sense of it could help predict such crises.
A new study by a researcher at the Indian Institute of Technology Roorkee looked at this question. It studies, through machine learning, whether balance-sheet vulnerabilities could be used as a metric to warn about an impending financial crisis.
His findings were published in the International Review of Economics & Finance, under the title “Do balance sheet vulnerabilities predict financial crises? Evidence from machine learning and Shapley decomposition.”
Financial Crises & Predictions
Because of their immense potential for destroying wealth, financial crises are a focal point of economic sciences.
A few key indicators are known to increase the likelihood of financial crises.
The first is “original sin”, or the inability of emerging and developing economies (EDEs) to borrow abroad in their own currency, which forces them to rely on foreign currency-denominated debt (FCD). This, in turn, makes their balance sheets susceptible to external shocks.
Another factor is currency mismatches, where a country, bank, or firm holds assets and liabilities in domestic and foreign currencies in unequal proportions. So, during exchange rate movements, this can create systemic risks that can lead to widespread insolvencies and macroeconomic volatility. For example, such a process was central to the 1997 Asian financial crisis.
Traditional methods of economic sciences, like logistic regression and signal extraction, often fail to properly assess these risks.
A big reason is that they often fail to take into account interaction effects and nonlinearities, where a similar input can suddenly create radically different results once a certain (hard to predict) threshold is passed.
Predicting Crises
Mapping Different Crises
Financial crises can manifest in various forms, with roughly three types generally recognized: currency, banking, and debt crises.
Currency crises can be defined as the “nominal depreciation of the currency vis-à-vis the US dollar of at least 30% that is also at least 10% points higher than the rate of depreciation in the year before”.
Banking crises require two simultaneous conditions:
- Significant signs of financial distress in the banking system, as indicated by significant bank runs, losses in the banking system, and/or bank liquidations.
- Significant banking policy intervention measures in response to significant losses in the banking system.
Sovereign debt crises can be identified through debt default and restructuring events, based on rating agency reports, reserve accumulations, IMF staff reports, and bond market developments.
While different from each other and stemming from different causes, these types of financial crises are not mutually exclusive. For example, a currency crisis can cause a severe recession that can then trigger a debt crisis or banking crisis in an especially vulnerable country.
Measuring Early-Warning Indicators
To overcome the limitations of previously used analysis tools, the author of this study created the broad original sin (BOSIN) and new currency mismatch (NCM) indices to find early indicators of a brewing financial crisis.
BOSIN incorporates both the currency composition of debt securities and cross-border bank loans, and covers 162 countries from 1970 to 2018.
NCM captures external vulnerability by including forex reserves, export openness, and currency composition of bank debt and international debt securities.
This shows interesting potential, as NCM reveals external vulnerabilities ahead of financial crises. For example, Argentina and Brazil exhibit their highest currency mismatch peaks around 2000, coinciding with their 2001–02 crises.
Another case is Mexico’s 1994–95 Tequila crisis. It saw a sharp increase in currency mismatches, due to dollar debt surging from 4% to 75% of privately held public debt within a year.
The final dataset used in the study included 120 pre-crisis observations for currency crisis (5.50%), 95 pre-crisis events for banking crisis (4.64%), and 47 for debt crisis (1.90%).
Crises & Machine Learning
For this study, the author chose tree-based machine learning because it requires minimal parameter tuning, handles large datasets well, and remains simple to implement and interpret.
The primary analysis focused on decision trees, random forests, and extremely randomized trees. The author also tested neural networks, support vector machines, CatBoost, and LightGBM as robustness checks.
The bagging-based ensemble models were the most consistent performers across the three crisis categories, although CatBoost and LightGBM produced competitive results for banking crises.
The study then compared different tree-based machine learning algorithms and their ability to predict crises, as well as their tendency for false positives (predicting a crisis when none happened).
Overall, “extreme” trees performed best, closely followed by random forest algorithms.
Predicting Versus Evaluating Crises
While identifying a potential crisis is important, policymakers and investors also need to understand why a model generated its warning. This is difficult with many machine-learning systems because their internal decision-making can resemble a black box.
To do so, the author used the economic indicators’ Shapley values, which represent their marginal contribution to crisis prediction, averaged across all possible indicator combinations.
This demonstrated that a few indicators are the most important for predicting financial crises:
- Economies with greater original sin exposure face higher currency crisis risks.
- The currency mismatch indicator shows a sharp rise in Shapley values at higher foreign currency liability positions.
- It could help establish specific vulnerability thresholds, providing policymakers with benchmarks to assess risks and guide macroprudential policies.
Different types of financial crises also have, unsurprisingly, different root causes.
- For currency crises, financial openness, exchange rate regimes, and the US Treasury rate emerge as top predictors.
- For banking crises, the US Treasury rate, bank credit, and financial openness are most predictive.
- For debt crises, GDP growth, financial openness, and currency mismatches are top indicators.
Financial Crises & AI Takeaways
This study demonstrates that some financial crises may produce detectable warning signals, despite the complexity and nonlinear behavior of macroeconomic systems.
This can be done using machine learning, especially tree-based algorithms, which proved more consistent across the three crisis categories than the other methods tested.
This works because machine learning can extract the relevant data for its prediction and detect nonlinear warning signals more effectively than traditional logistic regression.
It should also be noted that the “black box” criticism about machine learning and AI in general does not fully apply, as it is possible to demonstrate which of the economic indicators are contributing to the prediction, in this case, foreign-currency debt and currency mismatches.
This proves that such a model could strengthen sovereign-risk analysis and provide a warning about a country-level crisis getting ready to unfold.
However, investors should also be aware that high overall accuracy does not mean it will predict every crisis in time to avoid financial losses. Equally, false positives are still a distinct possibility.
This means that while machine learning can provide an alarm bell, deciding if it is accurate and needs to be acted upon still requires human experience and wisdom. So this technology is best positioned as a decision-support system, not an autonomous forecasting oracle.
Investing In Financial Risk Analysis
Moody’s Corporation
MCO 価格チャート
Predicting the risk of an unexpected crisis at the country or company level has always been an essential part of the work of investors, lenders, and financial experts in general.
In modern times, a lot of this task is delegated to rating agencies who have the expertise, data access, and reputation to make neutral, reliable calls on a company’s solvency, a bank’s balance sheet safety, a sector’s economic health, or a country’s financial stability.
One leading rating agency is Moody’s, which operates across credit ratings, sovereign analysis, financial intelligence, and risk modeling. Within Moody’s Investors Service, the Corporate Finance Group is its largest revenue contributor, accounting for approximately half of the segment’s revenue during the second quarter of 2026.
Source: Moody’s
By reducing information gaps between issuers and investors, widely recognized credit ratings can improve market access and potentially lower borrowing costs. Even a modest reduction in the interest rate attached to a large bond issuance can produce substantial savings over its lifetime.
Source: Moody’s
While AI models can be used to predict risks and forecast eventual crises, the certainty that the prediction is accurate is extremely important when billions are at play. So Moody’s (MCO ) is embracing AI, letting such technological advances be commercialized as interpretable financial-risk analytics, with the backing of an established provider of risk analysis already trusted for decades.
What provides a key advantage to Moody’s in using AI for mapping financial risk is not just its reputation, but direct access to one of the world’s largest curated and controlled datasets covering everything from patents to M&A deals, financial sentiment, economic data, individual company analysis, or even climate analytics.
Source: Moody’s
So overall, Moody’s is a good way to get exposure to the type of risk analysis using AI to predict debt crises and other types of financial crises.
Latest Moody’s Corporation (MCO) Stock News and Developments
Study Referenced
1. Hari Venkatesh. Do balance sheet vulnerabilities predict financial crises? Evidence from machine learning and Shapley decomposition. International Review of Economics & Finance. 2026年10月. Article: 105731. Volume 111. 10.1016/j.iref.2026.105731
