Bivariate sudden stop analysis of equity and bond fund flows to emerging markets using isolation forest

8.6.2026

Susanne Griebsch, Andreas Röthig  | Deutsche Bundesbank

What triggers sudden stops in capital flows to emerging markets? Using cutting-edge machine learning techniques, the authors introduce the isolation forest algorithm to uncover anomalies in equity and bond fund flows. Their innovative bivariate approach reveals new insights into the interconnected dynamics of global capital flows, particularly during periods of financial stress like the COVID-19 pandemic and recent interest rate reversals. 


What drives sudden stops in capital flows to emerging markets?
Emerging markets are highly vulnerable to abrupt changes in capital flows, which can destabilize their economies. This study introduces a novel machine learning approach, the isolation forest algorithm, to analyze sudden stops and surges in equity and bond fund flows. By addressing the limitations of traditional methods, the authors provide a more nuanced understanding of these anomalies, particularly during periods of financial stress, such as the COVID-19 pandemic and recent interest rate reversals in advanced economies.

Emerging markets have become increasingly integrated into global financial systems, making them more susceptible to sudden stops – sharp reversals in capital inflows – and surges, which can destabilize their economies. Traditional methods of analyzing these phenomena rely on arbitrary thresholds and univariate approaches, which fail to capture the complexity of portfolio flows. This paper seeks to overcome these limitations by applying the isolation forest algorithm, a machine learning technique, to detect anomalies in equity and bond fund flows to emerging markets. The study not only provides a continuous measure of anomaly intensity but also introduces a bivariate approach to analyze equity and bond flows simultaneously, offering a more comprehensive perspective on capital flow dynamics.

A Machine Learning Approach to Sudden Stops
The authors employ the isolation forest algorithm, an unsupervised machine learning method designed to detect anomalies by isolating data points that deviate significantly from the norm. Unlike traditional methods, which rely on fixed thresholds based on standard deviations, the isolation forest identifies anomalies by measuring the depth of decision trees required to isolate each data point. Anomalies, being rare and distinct, are isolated with fewer partitions, resulting in shorter path lengths. This approach allows for the creation of a continuous anomaly score, enabling a more detailed analysis of the intensity and frequency of sudden stops and surges. The study uses weekly data from the Emerging Portfolio Fund Research (EPFR) Global Database, covering equity and bond fund flows to emerging markets from 2004 to 2024.

Why Traditional Methods Fall Short
Traditional sudden stop analysis has three key limitations. First, it is univariate, analyzing equity and bond flows separately, which overlooks the interconnectedness of these markets. Second, it categorizes flows into only three states – sudden stops, normal flows, and surges – based on arbitrary thresholds, limiting the granularity of the analysis. Third, these thresholds are static and do not adapt to the evolving nature of financial markets. By contrast, the isolation forest algorithm addresses these shortcomings by providing a continuous anomaly score and enabling bivariate analysis, which captures the simultaneous dynamics of equity and bond flows.

Key Findings: A New Perspective on Anomalies
The study reveals several important insights. First, the isolation forest algorithm identifies a significant increase in both the frequency and intensity of anomalies in recent years, particularly during the COVID-19 pandemic and the subsequent interest rate reversal in advanced economies. These anomalies are often linked to simultaneous surges or stops in both equity and bond flows, highlighting the interconnectedness of these markets. For instance, the bivariate analysis shows that most anomalies occur during periods of simultaneous inflows or outflows in equity and bond markets, which the authors term "fund flow surges" and "fund flow stops." A smaller number of anomalies represent mixed patterns, such as equity inflows coupled with bond outflows, which may indicate shifts in investor risk preferences.

Second, the bivariate approach provides a more accurate classification of anomalies compared to traditional methods. While traditional methods often misclassify normal data points as anomalies, the isolation forest algorithm identifies anomalies that are clearly on the edges of the data space, indicating their distinctiveness. This is particularly evident during the financial turmoil of 2021–2022, when the isolation forest detected prolonged periods of anomalous fund flows, whereas traditional methods identified only brief episodes.

Third, the isolation forest's continuous anomaly score allows for a more nuanced analysis of the magnitude of anomalies. For example, the study finds that the surge in fund flows during 2021 was likely driven by a "search for yield" in a low-interest-rate environment, while the sudden stop in 2022 coincided with monetary tightening in advanced economies. These findings underscore the importance of considering both the intensity and the context of anomalies in capital flow analysis.

Implications for Policy and Future Research
The findings have significant implications for policymakers and researchers. The ability to detect and analyze anomalies in real time can enhance the monitoring of capital flow volatility, enabling more proactive policy responses to mitigate financial instability. For instance, the bivariate approach can help identify periods of heightened systemic risk, where simultaneous disruptions in equity and bond markets may amplify economic vulnerabilities. Moreover, the continuous anomaly score provides a valuable tool for stress testing and scenario analysis, offering insights into the potential impact of future shocks on emerging markets.

The study also opens new avenues for research. The authors suggest extending the bivariate approach to a multivariate framework, incorporating additional asset classes or country-specific data to capture the full complexity of global capital flows. Furthermore, aligning traditional methods with modern machine learning techniques could provide a more comprehensive understanding of sudden stops and surges, bridging the gap between established economic theories and data-driven approaches.

Conclusion: A Step Forward in Understanding Capital Flow Dynamics
This paper demonstrates the potential of machine learning techniques, particularly the isolation forest algorithm, to revolutionize the analysis of sudden stops and surges in capital flows. By addressing the limitations of traditional methods, the study provides a more detailed and accurate understanding of portfolio flow anomalies, with important implications for financial stability in emerging markets. As global financial markets continue to evolve, the integration of advanced analytical tools will be crucial for navigating the challenges and opportunities of an increasingly interconnected world.

Griebsch S., A. Röthig, (2026), Bivariate sudden stop analysis of equity and bond fund flows to emerging markets using isolation forest, Bundesbank Discussion Paper, No 15/2026

https://www.bundesbank.de/en/publications/research/discussion-papers/bivariate-sudden-stop-analysis-of-equity-and-bond-fund-flows-to-emerging-markets-using-isolation-forest-998888