> **来源：[研报客](https://pc.yanbaoke.cn)** # HOW TO ASSESS CRISIS RISKS # DEVELOPING A NEW EARLY WARNING SYSTEM AT THE ASIAN DEVELOPMENT BANK DECEMBER 2025 # HOW TO ASSESS CRISIS RISKS DEVELOPING A NEW EARLY WARNING SYSTEM AT THE ASIAN DEVELOPMENT BANK DECEMBER 2025 © 2025 Asian Development Bank 6 ADB Avenue, Mandaluyong City, 1550 Metro Manila, Philippines Tel +63 2 8632 4444; Fax +63 2 8636 2444 www.adb.org Some rights reserved. Published in 2025. ISBN 978-92-9277-531-5 (print); 978-92-9277-532-2 (PDF); 978-92-9277-533-9 (e-book) Publication Stock No. TCS250479-2 DOI: http://dx.doi.org/10.22617/TCS250479-2 The views expressed in this publication are those of the authors and do not necessarily reflect the views and policies of the Asian Development Bank (ADB) or its Board of Governors or the governments they represent. ADB does not guarantee the accuracy of the data included in this publication and accepts no responsibility for any consequence of their use. The mention of specific companies or products of manufacturers does not imply that they are endorsed or recommended by ADB in preference to others of a similar nature that are not mentioned. By making any designation of or reference to a particular territory or geographic area in this document, ADB does not intend to make any judgments as to the legal or other status of any territory or area. This publication is available under the Creative Commons Attribution 3.0 IGO license (CC BY 3.0 IGO) https://creativecommons.org/licenses/by/3.0/igo/. By using the content of this publication, you agree to be bound by the terms of this license. For attribution, translations, adaptations, and permissions, please read the provisions and terms of use at https://www.adb.org/terms-use#openaccess. This CC license does not apply to non-ADB copyright materials in this publication. If the material is attributed to another source, please contact the copyright owner or publisher of that source for permission to reproduce it. ADB cannot be held liable for any claims that arise as a result of your use of the material. Please contact pubsmarketing@adb.org if you have questions or comments with respect to content, or if you wish to obtain copyright permission for your intended use that does not fall within these terms, or for permission to use the ADB logo. Corrigenda to ADB publications may be found at http://www.adb.org/publications/corrigenda. Notes: In this publication, “$” refers to United States dollars. ADB recognizes "Hong Kong" as Hong Kong, China. Cover design by Erickson Mercado. # Contents # Tables and Figures # Foreword vi # Acknowledgments vii # Abbreviations viii # Executive Summary ix # Chapter 1. Introduction 1 1.1 Revamping ADB's Early Warning System amid New Policy Challenges 2 1.2 Crisis Forecasting at ADB: The Status Quo 3 1.3 Innovating ADB's Early Warning System: Machine Learning Models, Shapley Values, and Novel Variables 3 1.4 Early Warning System Design Steps 4 # Chapter 2. Crises Definitions and Dating 5 2.1 Banking Crises 5 2.2 Currency Crises 6 2.3 Sovereign Debt Crises 7 2.4 Fiscal Crises 7 2.5 Twin Crises 8 # Chapter 3. Modeling the Early Warning System: A Literature Review 10 3.1 Signal Approach 10 3.2 Discrete Choice Models 11 3.3 Dynamic Factor Models 11 3.4 Machine Learning Approach 12 # Chapter 4. Standard and Novel Early Warning Indicators 15 4.1 Standard Early Warning Indicators 15 4.2 Novel Early Warning Indicators 16 # Chapter 5. New Early Warning System Workflow Implementation 19 5.1 Data Selection 19 5.2 Model Selection 20 5.3 Out-of-Sample Prediction and Model Evaluation 20 5.4 Computing Shapley Values 21 5.5 Shapley Regressions 22 # Chapter 6. Early Warning System at ADB: Workflow and Preliminary Results 24 6.1 Processing of Data Variables 25 6.2 Machine Learning Model Outputs 26 6.3 Crisis Forecasting Sample 28 6.4 Results of Shapley Regressions 29 # CONCLUSION 31 # APPENDIXES A Financial Crises 32 B Review of Early Warning System Approaches 36 C Standard and Novel Early Warning Indicators 40 D Early Warning System Workflow-Additional Details and Results 61 # REFERENCES 64 # Tables and Figures # Tables 5.1 Out-of-Sample Predictive Area-Under-Curve Values of Machine Learning Models 21 5.2 Shapley Regressions 23 6.1 Summary of Variables in the Database 25 A1 Financial Crisis Dates-Financial Stress Index Versus Other Measures 32 B1 Contingency Table for Crisis Prediction Model 36 C1 Data Description and Sources, Standard Variables 41 C2 Data Description and Sources, Novel Variables 56 D1 Debt Crises Shapley Regressions 62 D2 Banking Crises Shapley Regressions 63 # Figures 5.1 Mean Absolute Shapley Values 22 6.1 Early Warning System Workflow 24 6.2 AUROC and AUPR Curve, Debt Crises 27 6.3 AUROC and AUPR Curve, Banking Crises 27 6.4 Forecasting Crises-Indonesia 28 6.5 Mean Absolute Shapley Values, Debt Crises 29 6.6 Mean Absolute Shapley Values, Banking Crises 30 A1 Crises Interactions 35 C1 Global Factor Timeline 54 C2 Financial Stress Index Timeline 54 C3 Total Spillovers Index 55 C4 Net Spillovers Indexes 55 # Foreword Economic crises remain a persistent threat even as Asia and the Pacific continues to demonstrate resilience to evolving challenges. The scars of economic crises are deep and change the trajectory of an economy's long-term growth. This can lead to sharp income declines due to lost jobs and wealth destruction amid falling asset prices, and stifle developing economies' ability to reach higher income levels, eradicate poverty, and support livelihoods. Previous crisis episodes cast a long shadow of lost economic opportunity. A decade after the Asian financial crisis of 1997, for example, cumulative lost output for Indonesia, the Republic of Korea, Malaysia, and Thailand have been estimated to range between $8\%$ and $14\%$ of gross domestic product. Recognizing the importance of early detection and timely policy responses, the Asian Development Bank (ADB) continues to strengthen its tools for crisis preparedness. This report, *How to Assess Crisis Risks: Developing a New Early Warning System at the Asian Development Bank*, introduces a cutting-edge methodology to predict crises, powered by state-of-the-art machine learning and a rich dataset of over 1,500 indicators, and representing a shift toward more adaptive forecasting. The early warning system integrates traditional macrofinancial metrics with emerging risks such as climate change and geopolitical tensions, and broadens the lens for assessing vulnerabilities. An intuitive companion dashboard to support ADB's internal country risk assessments delivers frequently updated forecasts, making insights accessible for policy dialogue with governments and development partners. Early identification of financial and economic imbalances is critical for achieving development objectives in Asia and the Pacific. That is why building capacities for crisis detection and effective resolution remains core to ADB's mission. This new tool kit reflects ADB's commitment to continuously upgrade analytical frameworks for crisis preparedness, and to build regional capacities to reduce vulnerabilities that impede progress toward a more prosperous, inclusive, resilient, and sustainable region. Beyond technical capacities, ADB stands ready to support member economies in addressing economic vulnerabilities through innovative financing facilities, policy advice, and knowledge services. # Albert Park Chief Economist and Director General Economic Research and Development Impact Department Asian Development Bank # Acknowledgments This report is produced with support from TA 6592-REG: Building Financial Resilience and Stability to Reinvigorate Growth and TA 10136-REG: Asian Economic Integration: Building Knowledge for Policy Dialogue, 2023-2025 (Subproject 3) of the Asian Development Bank (ADB). The production of this technical study was led by Alexander Raabe (Senior Economist, ADB) and Mattia Bevilacqua (Associate Professor in Finance, University of Liverpool Management School) under the supervision of Jong Woo Kang, Director of the Regional Cooperation and Integration Division in ADB's Economic Research and Development Impact Department. Mara Tayag (Senior Economics Officer, ADB) provided outstanding coordination support. The following ADB consultants supported the development of the model and preparation of this report: Tim Giffney developed the software code for the forecasting model; Ivan Borja de Leon and Achilleus de Mesa Coronel led the in-house implementation at ADB; and Rainiel Aquino, Pilar Capulong Dayag, and Clarisa Joy Flaminiano constructed the database of predictor variables and provided excellent research assistance. The development of the new early warning system benefited from advice by Adrian Alter (Senior Economist, International Monetary Fund). James Unwin edited the manuscript. Alfredo De Jesus implemented the typesetting and layout. Lawrence Casiraya proofread the report and Ma. Cecilia Abellar provided page proof checking services. Erickson Mercado created the cover design. # Abbreviations ADB Asian Development Bank API application programming interface AUC area-under-curve AUROC area under the receiver operating characteristic CDS credit default swap COVID-19 coronavirus disease DFM dynamic factor model EWS early warning system EWI early warning indicator FSI financial stress index GDP gross domestic product IMF International Monetary Fund LLM large language model MATR Measure of Aggregate Trade Restrictions ML machine learning NLP natural language processing RFE Recursive Feature Elimination ROC receiver operating characteristic SVM support vector machine US United States # Executive Summary The early detection of vulnerabilities leading to the buildup of economic crises is key to preserving growth and prosperity in Asia and the Pacific. This publication lays the conceptual groundwork for a new early warning system launched in September 2025 at ADB. Informed by a thorough review of state-of-the-art modeling approaches, we propose a new, machine-learning-based tool kit. The new tool kit predicts banking, currency, debt, and fiscal crises, with forecasting accuracy improvements by up to $23\%$ to the legacy system. Relative to that legacy system, the tool kit draws on a significantly broader scope of predictor variables at mixed frequencies covering macroeconomic and financial data including international spillovers and asset price volatility, as well as new development challenges ranging from climate change to geopolitical tensions. We employ Shapley values and Shapley regressions to uncover the main determinants of the forecast, allowing identification of levers for preventive and mitigating actions, and to communicate actionable proposals to policymakers. The analytical framework for crisis forecasting presented in this report is designed to be highly flexible, enabling the integration of additional indicators—such as those derived from text-based sources—and the incorporation of methodological advances in machine learning. The crisis chronologies used for calibrating projections are easily updatable, ensuring alignment with the evolving academic literature. # CHAPTER1 Introduction Recent economic and financial crises are a reminder that such events can have devastating impact on long-term growth and prosperity. Interest in identifying vulnerabilities and predicting financial stress has therefore surged. Sufficiently early detection of the buildup of stress gives mitigating measures a greater chance of dampening output losses or avoiding them altogether. Past approaches to predicting crises are challenged by overly adapting models to past observations (model overfitting), in part amplified by the selection of the economic and financial variables used to extract crisis signals. Aggravated by only a few historical crisis observations available and generally hard-to-predict black swan-type events, the overfitting typically results in an unfavorable resolution of the trade-off between missed crises and false alarms. Given this, advances in machine learning (ML) offer a unique opportunity to improve upon the forecasting performance of past models. ML-based models are well suited to identify vulnerabilities in developing members of the Asian Development Bank (ADB), addressing their scarcity of relevant data and pronounced exposure to nontraditional economic risks like climate change and geopolitical tensions. This study develops a more flexible and comprehensive ML-based framework to predict crises. This new early warning system (EWS) significantly improves the forecasting performance of past models used at ADB. All types of financial crises have huge economic and social costs (Kaminsky and Reinhart 1999; Hoggarth, Reis, and Saporta 2002; Reinhart and Rogoff 2009a, 2013; Laeven and Valencia 2020) as evidenced by rapid declines in output, employment, and asset prices, and accompanied by increases in government debt and inflation (Dell'Ariccia, Detragiache, and Rajan 2008; Reinhart and Rogoff 2009, 2013). Beyond the financial and economic impact, financial crises tend to have profound effects on the broader society, with examples spanning increased mortality rates (Cutler et al. 2002), reduced public spending on education and health care (Knowles, Pernia, and Racelis 1999), and a marked increase in poverty (Cruces and Wodon 2003; Suryahadi, Sumarto, and Pritchett 2003). Importantly, crises can lead to long-term economic depression, leading to a subsequent recession (Estrella and Mishkin 1996; Kaminsky and Reinhart 1999; Farmer 2012; Barro and Ursua 2017). The output often does not fully recover from crisis-induced contractions. Thus, repeated crises can lead to permanent output losses, preventing income gaps between low- and high-income economies to close (Cerra and Saxena 2008). Therefore, an early detection of vulnerabilities possibly leading to crises is paramount to intervening with targeted policies to shore up economic resilience and prevent crises. Crises are often thought of as unpredictable. For instance, senior policymakers took this position for the 2008 global financial crisis (Greenwood et al. 2022). Also, Gorton (2012) states that "crises are sudden, unpredictable events," a view also supported by earlier evidence showing that, while crises are often preceded by weak macroeconomic fundamentals, the predictability is low in light of an element of randomness (Kaminsky and Reinhart 1999; Chari and Kehoe 2003). The claim of non-predictability is rooted in two modeling challenges. First, and more specifically, crises are rare events making it hard to detect patterns and thus model crises. In turn, this can lead to model overfitting, causing often poor out-of-sample forecasting performance. Second, a trade-off between failing to detect crises, and false alarms needs to be managed well (Candelon, Dumitrescu, and Hurlin 2014; Aldasoro, Borio, and Drehmann 2018; Truong et al. 2022). On the one hand, the model needs to be sufficiently accurate so as not to miss crises (type I errors) or signal the crisis only when it is already too late to intervene. On the other hand, the model should not be calibrated too sensitive, leading to excessive false alarms (type II errors). The latter can be costly due to unnecessary policy responses and the buildup of crisis-fighting fatigue. $^{1}$ Third, it is difficult to aggregate relevant data into simple-to-use monitoring systems which remain sufficiently generalizable to be relied upon at any time, for every crisis, and in every economy. Fourth, relevant data are often scarce, especially in developing member economies. Fifth, model complexity often prevents distilling clear leading indicators to guide timely policy intervention. Finally, black-swan-type events (e.g., the coronavirus disease [COVID-19] pandemic) are by definition impossible to predict. While predicting the future remains inherently difficult, recent advances in data science and data availability have eased the pressure on model design. Notably, the literature on forecasting techniques has grown rapidly, improving the accuracy of out-of-sample predictions by adopting state-of-the-art techniques, such as ML (Alessi and Detken 2018; Truong et al. 2022; Casabianca et al. 2022; Liu, Chen, and Wang 2022; Bluwstein et al. 2023). Second, historical records of crises have been collected, allowing new models to be calibrated on more past observations of crises than previously possible (Reinhart and Rogoff 2011; Laeven and Valencia 2013; Jordà, Schularick, and Taylor 2017; Dawood, Horsewood, and Strobel 2017; Laeven and Valencia 2020; Baron, Verner, and Xiong 2021; Moreno Badia et al. 2022). This report provides the context for EWS designs at ADB, outlines general principles, and explains the conceptual steps to construct a new EWS. # 1.1 Revamping ADB's Early Warning System amid New Policy Challenges The 1997 Asian financial crisis was among the worst financial crises of the 20th century for many Asian economies. Indonesia, the Republic of Korea, and Thailand were especially hard hit. Moreover, Asia was strongly exposed to spillovers of the 2008 global financial crisis. The financial distress and volatility were abetted and deepened through a combination of domestic and external factors, including weak internal financial systems, excessive debt, and fixed local exchange rates (Kaminsky and Reinhart 2001; Claessens and Kose 2013; Buckley et al. 2020). Globally, vulnerabilities to crises are on the rise: By 2019, public debt for the average economy stood at $43\%$ of gross domestic product (GDP), up by 10 percentage points since 2008 (Ferrarini et al. 2023). Debt-fueled spending to address the COVID-19 pandemic increased debt levels. As of 2025, public debt to GDP in Asia has reached $47\%$ on average. More recently, supply shocks from Russia's war in Ukraine, geopolitical fragmentation disrupting global supply chains, and aging societies have added to inflation pressures while lowering growth. Climate change and nature loss add to spending pressures, while technological innovations, notably in payment systems, may add to uncertainty. Financial markets' mispricing of such risks can upset financial stability across Asia (te Kaat, Raabe, and Tian, forthcoming). Finally, strong cross-border financial linkages can quickly amplify shocks through global network effects (Park et al. 2020; Rosenkranz and Melchor 2022; Chowdhury et al. 2019). This is the context relevant to the aim of the revamped EWS, which is to produce forecasts that are more accurate than legacy systems. The new EWS expands to include a larger set of financial crises and a significantly increased set of explanatory variables. In turn, the new EWS strengthens macroeconomic stability in Asia and the Pacific, and regional policymakers' capacity to identify priorities to safeguard growth. While the new EWS equips policymakers with a better understanding of emerging vulnerabilities, and allows for a timely policy response, the design of appropriate macroeconomic, fiscal, and financial mitigating policies remains outside of the scope of the new forecasting model. # 1.2 Crisis Forecasting at ADB: The Status Quo ADB's legacy crisis forecasting framework was built in 2005 as a statistical tool kit along with a proprietary software, called VIEWS. The tool kit aimed at building regional capacity to detect emerging finance sector vulnerabilities and crises, and to inform the choice of mitigating policy actions. It relied on the signal approach of Zhuang (2005), based on the earlier works of Kaminsky and Reinhart (1999) and Goldstein, Kaminsky, and Reinhart (2000). Relative to the earlier literature, Zhuang (2005) innovated the modeling by (i) including leading indicators tailored to selected Asian economies, (ii) constructing six sector-specific composite indexes for crises prediction, and (iii) adopting a stochastic trend to proxy for the long-term equilibrium level of the real exchange rate. The VIEWS software was used to identify variables strongly associated with crisis episodes and to conduct scenario analyses (i.e., stress testing and forecasting). The framework's latest version covers currency and banking crises, and macroeconomic imbalances. Details on the methodology and steps of the signal approach are discussed in Chapter 3. However, new sources of financial and economic vulnerabilities, as outlined earlier in this introduction, coupled with advances in modeling techniques, suggested a need to revamp ADB's new EWS. # 1.3 Innovating ADB's Early Warning System: Machine Learning Models, Shapley Values, and Novel Variables The new EWS relies on state-of-the-art ML models and innovates the legacy system along several dimensions. First, we augment past datasets by incorporating both standard EWS variables frequently used in the established literature on crisis forecasting such as macroeconomic, fiscal, and balance of payments indicators, and novel variables that reflect broader financial, economic, political, and demographic concepts—capturing new policy challenges not yet considered in established forecasting models. Second, beyond these standard data dimensions, we include cross-border financial linkages and international factors, acknowledging the importance of crisis dynamics and the potential contagion across borders. Third, the ML approach allows us to build in flexibility to navigate nonlinearities, interactions, multicollinearity, and mixed frequencies. Finally, relative to the legacy system, we achieve a significant increase in forecast performance, as documented in Chapter 6. Moreover, running the model requires minimal manual intervention relative to the legacy framework, largely limited to updating some of the data feeding into the forecasting process. Producing forecasts, and the display of results on a dashboard, are fully automatized, allowing for quick, reliable, and transparent access. The strength of ML models lies in capturing nonlinearities and incorporating information from a significantly expanded set of explanatory variables without specifying functional forms beforehand. The advantages of ML techniques become particularly valuable when forecasting crises in developing member economies, since they offer needed flexibility to manage sparse, unbalanced, and mixed-frequency datasets more effectively. However, the improvement in predictive accuracy and efficiency comes at the expense of model interpretability, as ML models obscure direct associations between outcome variables (crisis probability) and explanatory variables. As a remedy, we draw on so-called Shapley values, a concept recently adopted from game theory, and apply even more recently developed Shapley regressions to ensure that the ML model output remains interpretable, as elaborated in Chapter 5. This allows for a timely design of mitigating measures. # 1.4 Early Warning System Design Steps Next, we provide guidelines for designing effective EWSs. First, it is important to accurately identify the historical crisis chronology. In all the approaches used to classify a crisis, the resulting variable is either discrete, binary, or multinomial. This variable is then generally employed as the outcome variable of the predictive models in the EWS literature (Chapter 3) and used to train ML-based models in-sample (Goldstein, Kaminsky, and Reinhart 2000). Successfully identifying financial crisis episodes involves extending the analysis beyond the most recently recorded crises, encompassing a broader sample (Chapter 2). Second, EWSs require a set of explanatory variables to obtain the crisis prediction. The choice of the explanatory variables is usually guided by economic theory, and it differs according to the type of crisis to be predicted. However, as explained in Chapter 4, it is crucial to include a wide array of early warning indicators (Zhuang 2005; Lo Duca and Peltonen 2013; Truong et al. 2022). Third, the frequency of the data should be selected in line with the research endeavor. For instance, monthly data on several variables are available for fewer economies relative to annual data, and would result in a smaller sample. Conversely, monthly data provide more timely information. Finally, the accuracy of the out-of-sample model predictions is informative about the added value of specific explanatory variables to be included (Chapter 5). # CHAPTER 2 # Crises Definitions and Dating The early warning system (EWS) model performance crucially depends on proper identification of the crisis dates serving as model input. Hence, when building an EWS, the first step is to exactly date crises (Boyd et al. 2019; Baron, Verner, and Xiong 2021; Truong et al. 2022). Adopting wrong dates may either hide the relationship between a crisis event and other variables or generate spurious linkages between crises and these variables (Laeven and Valencia 2020). Records of crises, called chronologies, have expanded significantly. $^3$ A common approach to date crises is to use macroeconomic and financial data to construct EWS indicators, and denote a crisis as a period when these indicators exceed specific threshold values. $^4$ Values above the threshold signal upcoming distress and heightened probability of a crisis (Zhuang 2005; Truong et al. 2022). In the next sections, we review the literature on dating crises for each specific type of crisis: for banking, currency, sovereign debt, fiscal conditions, and their interactions, focusing on definitions and state-of-the-art chronologies. # 2.1 Banking Crises The 2008 global financial crisis sparked new interest in dating and predicting banking crises. Notable contributions came from Demirguc-Kunt and Detragiache (1998), Reinhart and Rogoff (2009a), Schularick and Taylor (2012), Romer and Romer (2017), Laeven and Valencia (2013, 2020), Baron, Verner, and Xiong (2021), and Ahir et al. (2023). Most of these studies have taken different approaches to identifying and dating banking crises, encompassing mainly narrative approaches, policy interventions, quantitative research, or a mix of these. The narrative approach looks at the narrative sources of events, such as bank runs, policy intervention, or equity declines. Pioneering works are Caprio and Klingebiel (1996); Bordo et al. (2001); Caprio and Klingebiel (2002); Reinhart and Rogoff (2009a, 2009b, 2011, 2013); Schularick and Taylor (2012); and Jordà, Schularick, and Taylor (2017). Authors from the International Monetary Fund (IMF) refined earlier banking crisis chronologies by expanding the criteria to identify substantial policy interventions and adopting quantitative variables (Laeven and Valencia 2013, 2020). The rationale of these studies, with hindsight of the “too big to fail” bank bailouts, was that policy interventions can mitigate bank losses and quantifying the depth of the crisis without them would be difficult. Therefore, a banking crisis is defined as an event meeting not one, but two conditions: (i) significant signs of financial distress in the banking system (e.g., significant bank runs, losses in the banking system, and/or bank liquidations), and (ii) significant policy intervention measures in response to significant losses in the banking system (Laeven and Valencia 2020). When banking sector losses or liquidations are severe, the first criterion is a sufficient condition to date a banking crisis. The second criterion is also considered when it is not easy to quantify the degree of financial distress in a banking system. The first year when both criteria are met is considered the crisis start date. This approach ensures that crises are dated at the early signs of significant problems in the banking system. Furthermore, Laeven and Valencia (2020) update the comprehensive global database on banking crises put forward in Laeven and Valencia (2013). The update covers all crisis episodes during 1970–2017. There are advantages and disadvantages of using these approaches separately (Romer and Romer 2017). Statistical measures are objective and capture variations in financial indicators across crisis episodes in real time but are typically limited to advanced economies and cover short time horizons. Hence, exclusive reliance on them may miss financial disruptions or misidentify crisis episodes. At the same time, one drawback of a focus on a narrative approach is that the source could be idiosyncratic or biased and might miss fundamentals shocks. For these reasons, the most broadly used financial crisis chronologies are based on historical analyses of events combined with statistical indicators (Caprio and Klingebiel 1996; Reinhart and Rogoff 2009a; Laeven and Valencia 2013, 2020; Baron, Verner, and Xiong 2021; Ahir et al. 2023).<sup>5</sup> Table A1 in Appendix A reports several banking crisis episodes selected for some ADB regional members, taken from Ahir et al. (2023), and compares their index with other banking chronologies. We observe instances in which their index does not capture any financial stress episodes while other approaches do, and vice versa; this discrepancy is dependent on economy coverage, frequency, and time coverage across measures. Overall, several of the studies on dating banking crises note significant similarities with the crisis chronology by Laeven and Valencia (2013, 2020), when the sample overlaps. Given the comprehensive coverage of economies and historical sample period studied in this work, Laeven and Valencia (2020) can be considered among the best state-of-the-art banking crises chronologies available today. # 2.2 Currency Crises A currency crisis could occur for several reasons (Kaminsky 2006). They are often defined as episodes of significant currency depreciation (mostly against the United States [US] dollar), foreign reserve losses, or short-term interest rate hikes (Zhuang 2005). In practice, two methods have been used to date currency crises. The first considers month-on-month changes in exchange rates, foreign reserves, and interest rates in defining crises separately and adopting fixed and subjectively defined thresholds (Frankel and Rose 1996; Zhuang 2005). Many studies adopted this approach. For instance, Laeven and Valencia (2013, 2020) adopt the above criteria considering two thresholds for depreciation: (i) a year-on-year depreciation of at least $30\%$ , and (ii) a depreciation of at least $10\%$ extended from that in the year before, based on the official nominal bilateral exchange rates from the IMF. The second involves constructing an exchange market pressure index (Eichengreen et al. 1995; Kaminsky et al. 1998). A crisis episode is considered to occur in a particular month if this index exceeds its sample mean by a certain number of standard deviations, the choice of which remains arbitrary. Other studies have adopted alternative approaches to dating crises (Zhang 2001), and a modified version of the Kaminsky et al. (1998) pressure index, adding more economies and variables (Edison 2003; Lestano and Jacobs 2004; Candelon, Dumitrescu, and Hurlin 2014). More recently, Goldberg and Krogstrup (2023) propose a new index that combines pressures observed in exchange rate adjustments with model-based estimates of incipient pressures that are masked by foreign exchange interventions and policy rate adjustments. However, crisis periods identified through these thresholds have to be cross-validated against the historical evidence of currency crises in each economy (Candelon, Dumitrescu, and Hurlin 2014; Lestano and Jacobs 2004). # 2.3 Sovereign Debt Crises During the 2008 global financial crisis, many governments chose to bail out failing banks, jeopardizing fiscal sustainability (Dawood, Horsewood, and Strobel 2017). This context revitalized the importance of constructing monitoring tool kits for sovereign debt turmoil. Monitoring largely elevated sovereign debt levels relative to prepandemic times has become a pressing in Asia and the Pacific (Ferrarini et al. 2023).7 Manasse et al. (2003) propose an approach to date sovereign debt crises aiming to capture both actual and potential defaults on sovereign debt. This classification has become a standard in the literature on forecasting sovereign debt crises, including Ciarlone and Trebeschi (2005), Fioramanti (2008), Manasse and Roubini (2009), and Dawood, Horsewood, and Strobel (2017). Laeven and Valencia (2020) construct a database of historical debt crises dates. They date episodes of sovereign debt default and restructuring by relying on information mainly from Sturzenegger and Zettelmeyer (2007) and Cruces and Trebesch (2013), and also report from rating agencies and the media. Adopting this approach, they identify 79 episodes of sovereign debt crises during 1970-2017, 12 of which have taken place since 2007. The study by Laeven and Valencia (2020) is recognized as one of the most thorough chronologies of crises available, making it suitable as the main reference chronology in ADB's new EWS. Laeven and Valencia (2020) also provide an overview of the global coverage of the database covering banking, currency, and sovereign debt crises during 1970-2017, summing to 151 banking crises, 236 currency crises, and 79 sovereign crises. # 2.4 Fiscal Crises Finally, another important set of events includes fiscal crises. These typically entail a significant loss of annual output (Medas et al. 2018). Since the aftermath of the global financial crisis and the European sovereign debt crises of 2010, there is greater interest in how to detect and avoid fiscal crises (Hellwig 2021). However, the previous literature on different EWS for fiscal crises, in general, relies on relatively small samples of advanced and emerging market economies. Only a few studies include low-income economies (Cerovic et al. 2018), even as fiscal crises are found to be more frequent and associated with larger output losses in these economies (Gerling et al. 2017; Moreno Badia et al. 2022). Fiscal crises have been traditionally associated with sovereign debt crises triggered by external default episodes. While sovereign defaults represent one category of significant shocks that can lead to fiscal crises, they do not capture all fiscal crisis episodes. In fact, an economy may experience fiscal distress when large imbalances emerge between inflows (revenues and financing) and outflows (primary expenditures and debt service). These imbalances may lead to a fiscal crisis if the economy cannot adjust its fiscal position sufficiently and quickly (Gerling et al. 2017). Recent literature has started to model fiscal crises independent from defaults (Medas et al. 2018). Over time, the definition has extended to cover large-scale official financing, domestic public debt defaults, and higher inflation (Manasse et al. 2003; Reinhart and Rogoff 2009b, 2011; Erce et al. 2022). In general, the term fiscal crisis describes a period of heightened budgetary distress, resulting in the sovereign taking exceptional measures (Gerling et al. 2017; Hellwig 2021). Recent studies produced chronologies of fiscal crises (Baldacci et al. 2011; Bruns and Poghosyan 2018; Medas et al. 2018; Moreno Badia et al. 2022).<sup>9</sup> By adopting the above criteria, Medas et al. (2018) expand the economy coverage to 188 economies, over 1970–2015, identifying 439 fiscal crises. Moreno Badia et al. (2022) update the IMF’s dataset by Medas et al. (2018) during 1980–2018 by (i) expanding the coverage to sovereign debt yields, (ii) improving the information on domestic arrears, and (iii) checking the quality of data through IMF economy teams. The number of crises closely matches previous data from the IMF, identifying 384 crisis episodes for a sample of 188 economies over 1980–2018. This constitutes the most comprehensive and up-to-date studies of fiscal crises, and is thus used to train ADB’s new ML-based EWS. # 2.5 Twin Crises In line with the Anna Karenina principle, $^{10}$ idiosyncratic factors tend to favor the buildup of vulnerabilities leading to crises. Still, crises tend to occur simultaneously. $^{11}$ Financial crises tend to begin with a private debt overhand, leading to banking crises, and eventually morphing into sovereign debt and/or fiscal crises as private losses are socialized. Laeven and Valencia (2020) document such twin crises. Figure A1 in Appendix A shows an example of crises interaction drawn from Laeven and Valencia (2020) during 1970–2017. Among twin crises, the currency–banking (31) and currency–debt (20) crisis pairs tend to be more common than the banking–debt (3) crisis pair. Triple crises (i.e., simultaneous banking, currency, and debt crises in a given economy) are less common (11) (Laeven and Valencia 2020). $^{12}$ Moreover, close to a fifth of fiscal crises happen at the same time as either a banking or currency crisis, and can coincide with both (Medas et al. 2018). The interaction between fiscal and banking crisis can lead to larger economic losses characterized by a deeper decline in growth than stand-alone crises (Laeven and Valencia 2013; Romer and Romer 2017). Moreno Badia et al. (2022) find that in about a third of cases there is an overlap between fiscal and currency crises, mostly in emerging markets or low-income economies. Finally, all types of crises tend to happen around the same time, not only within the same economy, but could also spread to others in "waves." Especially, banking crises are rarely single-economy events, and spread across borders (e.g., across Latin America in the early 1980s; in Asia during the 1997 Asian financial crisis; and more recently, the 2008 global financial crisis). # CHAPTER 3 # Modeling the Early Warning System: A Literature Review Significant advances have been made in improving crisis forecasting models. This chapter reviews the commonly adopted methods in the literature, contrasting their advantages and disadvantages. # 3.1 Signal Approach Seminal studies on forecasting financial crises relied on the signal approach (Kaminsky et al. 1998; Kaminsky and Reinhart 1999). The approach starts with a selection of indicators that are conjectured to predict crises. For each indicator, a specific threshold value is set based on a specific percentile of each indicator's sample distribution. If an indicator crosses this threshold, a signal is issued, suggesting that a crisis is likely to occur in the forecast time window. Leading indicators are converted into binary variables which take a value of 1 if the actual value of the leading indicator crosses the threshold (warning signal), or 0 if the actual value does not cross its threshold (no warning signal). For technical details on the mechanics, refer to Appendix B.1. The signal approach is a nonparametric and simple method to use in different contexts (Kaminsky and Reinhart 1999). Its strengths lie in a straightforward assessment of indicators' predictive power, and its flexible use in several crisis prediction applications (Kaminsky et al. 1998; Goldstein, Kaminsky, and Reinhart 2000; Borio and Lowe 2002; Davis and Karim 2008; Baldacci et al. 2011; Savona and Vezzoli 2015; Cerovic et al. 2018). Despite its simplicity, flexibility, and ease to compute and interpret, the signal approach entails several drawbacks. The performance of the approach depends on the leading indicators chosen for each type of crises (Frankel and Saravelos 2012). Modeling the preferences governing the trade-off between false alarms and missing crises can be difficult. Choosing the "optimal" threshold that best balances both types of errors; that is, false alarms and missing crises can be challenging and discretionary (Dawood, Horsewood, and Strobel 2017). Moreover, although the economy may be vulnerable to an impending crisis, many of the indicators may not signal the possibility of distress, either jointly or in good time. A key reason lies in the ad hoc choice of the thresholds, which directly affects the distinction between crisis and noncrisis episodes (see Appendix B.1). Finally, this approach provides no statistical inference. ADB's legacy EWS incorporated in the VIEWS software relies on the signaling approach. # 3.2 Discrete Choice Models Discrete choice or limited dependent variable (logit/probit) models are frequently used. The most common form is a regression with the outcome variable being binary (crisis/noncrisis) and the probability that the event (crisis) occurs being estimated as a function of the factors adopted. From the estimated coefficients of the model, it is possible to retrieve the estimated probabilities of the crisis. An early study adopting this approach is Frankel and Rose (1996), who use a probit model to estimate the probability of currency crises using annual data. The first application of this methodology to banking crises is Demirgüç-Kunt and Detragiache (1998). The literature using this methodology to analyze sovereign debt crises includes Manasse et al. (2003), Fuertes and Kalotychou (2007), and Dawood, Horsewood, and Strobel (2017), and for financial crises comprises Davis and Karim (2008), Barrell et al. (2010), Gourinchas and Obstfeld (2012), Schularick and Taylor (2012), Duca and Peltonen (2013), and Caggiano et al. (2014). Discrete choice models are widely employed for their ease of estimation and interpretation. Moreover, their forecasting ability usually outperforms the signal approach (Berg and Pattillo 1999; Kumar et al. 2003; Fuertes and Kalotychou 2007). Some disadvantages are that they heavily depend on data availability. Also, these early models do not consider that the chosen indicators could behave differently during tranquil times or postcrisis periods. For instance, the behavior of the indicator is affected both by the crisis itself and the policies undertaken to mitigate it. Combining observations of tranquil periods with those of postcrisis ones into a single (zero) group can lead to "postcrisis bias" (Bussiere and Fratzscher 2006). To overcome this bias, some studies adopted a multinomial logit where the crisis variable is modeled to reflect all three states—i.e., tranquil times, crisis, and after the crisis (Bussiere and Fratzscher 2006; Caggiano et al. 2014). Consequently, some "hybrid" methods combining approaches have been proposed (Fuertes and Kalotychou 2007; Schularick and Taylor 2012; Duca and Peltonen 2013; Savona and Vezzoli 2015), along with Markov switching models to analyze time series data in different regimes or states (Abiad 2003). Overall, discrete choice models are limited by data availability, crisis modeling, and, out-of-sample forecasting performance. # 3.3 Dynamic Factor Models Other advances in EWSs rely on dynamic factor models (DFMs) introduced by Forni et al. (2000) and Forni et al. (2005). In general, DFMs postulate that a small number of latent factors explain the common dynamics of a larger number of observed time series (Stock and Watson 2016). An important development in the literature on DFMs is also the popular mixed frequency generalization and mixed data sampling, or MIDAS, regression models (Banbury and Modugno 2014). Hence, DFMs are useful to combine information from rich but unbalanced mixed-frequency datasets. This approach is very flexible and allows the assimilation of newly available data, such as financial high-frequency data (Andreou et al. 2010, 2013). For instance, Truong et al. (2022) add data credit default swap into a DFM model, accounting for mixed-frequency data and external, domestic, and global factors applied to the Asian context. However, DFMs also come with limitations. One criticism is that the factors are estimated from extensive panel data and do not take full advantage of data hierarchy—for instance, ignoring economy-specific structures, heterogeneity between income levels, and nested relationships (e.g., indicators within economies over time) that may be crucial for interpretation (Wang et al. 2022). To overcome this problem, Truong et al. (2022) develop EWSs for financial crises with a focus on small open economies using DFMs with a multilevel factor structure applied to panel data. However, DFMs suffer limitations as overly simplified models, with the exact structure of factor interdependencies being chosen by prior knowledge rather than analytical detection, leading to potential model mis-specification (Wang et al. 2022), significant estimation bias, and poor model fit (Francis et al. 2017). In addition, DFMs rely heavily on the availability and quality of economic and financial data, and are based on assumptions about the underlying structure of the data. Also, they typically use lagged information to make predictions but do not consider contemporaneous shocks. Moreover, some DFMs can be computationally intensive, particularly when dealing with large datasets. Finally, in attempting to capture all potential sources of variation, DFMs can become overly complex, risking model overfitting and thus weak out-of-sample predictive performance (Stock and Watson 2016). # 3.4 Machine Learning Approach More recently, a large effort among academics and policymakers has seen the development of machine learning (ML) methods. ML can be defined as a subset of artificial intelligence in the field of computer science that draws on statistical techniques to give the model the ability to "learn" from the data without functional forms between variables being explicitly programmed (Samuel 1959). In general, ML is a data-mining tool kit able to analyze complex datasets, fit multifaceted and flexible functional forms to the data, and find functions that perform well out of sample (Mullainathan and Spiess 2017). ML approaches have been growing fast and are applied extensively for predicting any type of financial crisis. Among several, Bluwstein et al. (2023), Samitas et al. (2020), and Tölö (2020) for financial crises; Duttagupta and Cashin (2011), Alessi and Detken (2018), and Wang et al. (2021) for banking crises; Manasse et al. (2003), Manasse and Roubini (2009), and Savona and Vezzoli (2015) for sovereign debt crises; Hellwig (2021) and Moreno Badia et al. (2022) for fiscal crises; and Lin et al. (2008) and Sevim et al. (2014) for currency crises. More details on some of the most common ML approaches were reported in Appendix B.2. The appeal of ML methods is that they overcome several well-known challenges in the literature on EWSs. First, they allow us to analyze complex and large datasets, circumventing the risk of model overfitting. $^{17}$ Second, ML approaches allow the inclusion of a large number of variables also taking into account crisis interactions, and even transformations of these variables and/or their lagged values, thus including predictors which could be redundant and highly correlated. $^{18}$ The more candidate predictors we add, the lower the risk that variable selection is biased by our own judgment. Third, the dynamics of financial crises are mostly complex, not well captured in linear models (Demirguc-Kunt and Detragiache 1998; Hellwig 2021; Bluwstein et al. 2023). The clear advantage of ML approaches is their ability to learn nonlinearities and interactions directly from the data without pre-specification (Cesa-Bianchi et al. 2019; Goulet Coulombe et al. 2022; Bluwstein et al. 2023). In addition, the advancements of ML techniques pave the way for a recent strand of economic literature which leverages the power of large language models (LLMs) and natural language processing (NLP) for various applications, including economic time series forecasting (Carriero et al. 2024). Finally, the overall consensus in the ML literature is that most ML models outperform several other methods, including the signal approach, discrete choice models, and DFMs in out-of-sample predictions (Manasse and Roubini 2009; Sevim et al. 2014; Alessi and Detken 2018; Beutel et al. 2019; Tölö 2020; Fouliard et al. 2021; Hellwig 2021; Liu, Chen, and Wang 2022; Casabianca et al. 2022; Bluwstein et al. 2023). However, ML techniques can be typically complex and difficult to interpret, especially when interpreting the importance of explanatory variables (Bluwstein et al. 2023). This limitation makes ML models appear as "black box," obscuring insights for policymakers on which actions to take to avert crises. For the same reason, economic mechanisms underpinning the prediction cannot be easily inferred, although none of the statistical techniques to forecast crises can prove causality. To overcome these difficulties in interpretation and classification of important variables, several authors recently started to augment ML methods with Shapley values (Liu, Chen, and Wang 2022; Buckmann et al. 2022; Chan-Lau et al. 2023; Bluwstein et al. 2023; Casabianca et al. 2022). Use of Shapley values for ML models interpretation is borrowed from the cooperative game theory literature (Shapley 1953; Strumbelj and Kononenko 2010). In that literature, Shapley values are used to calculate the payoff distribution across a group of players, whereas in the context of crisis prediction, they can be used to calculate the marginal contribution from including different predictors in the models. This allows crisis probability to be decomposed into the sum of contributions from each predictor (Shapley values) (Bluwstein et al. 2023). Augmenting the ML output with Shapley values allows us to identify which variables are driving the model prediction, to quantify their marginal contribution and the direction of the effect on the crisis probability. Shapley values are superior to other importance measures given the set of appealing analytical properties thanks to their origin in game theory (Lundberg et al. 2020