Can a universal bankruptcy model adapt to economic ups and downs? New evidence says it might

Machine learning models delivered the strongest performance across nearly all evaluation metrics. CHAID and CART provided the highest and most stable sensitivity, accuracy and discriminatory power, while artificial neural networks performed well but required more complex calibration. Discriminant analysis consistently ranked as the weakest method, reinforcing concerns about relying on older statistical techniques in unpredictable economic environments.

A new economic analysis suggests that organizations may not need to rebuild or recalibrate their bankruptcy prediction systems every time the economy enters a new phase. The findings point to a significant breakthrough in how financial risk can be monitored at scale.

The study, titled One Model Fits All? Evaluating Bankruptcy Prediction Across Different Economic Periods and published in Economies, examines whether a single predictive model can effectively signal insolvency risk through periods of economic calm, disruption, and recovery. The research evaluates more than 320,000 firm-year observations from Slovak small and medium-sized enterprises between 2018 and 2023, covering the pre-pandemic period, the global COVID-19 shock and the post-pandemic adjustment years.

Predictive models tested against three distinct economic climates

The authors structured the research around two opposing approaches. The first involved building separate bankruptcy prediction models for each economic phase: pre-pandemic stability, the highly disrupted pandemic years and the recovery period that followed. The second approach used a comprehensive unified model trained on all observations combined, with an added economic-period indicator variable to capture structural differences.

Six modeling techniques were compared: CART, CHAID, C5.0, logistic regression, discriminant analysis and artificial neural networks. This mix allowed the researchers to test classical statistical methods against modern machine learning techniques, offering insight into which systems remain resilient when firms experience destabilizing shocks.

Across all models and economic periods, two financial indicators emerged as dominant predictors of distress: liquidity and indebtedness. Liquidity 1, a short-term liquidity measure, and the Total Debt Ratio consistently appeared as the most influential variables. Profitability measures played supporting roles but did not match the predictive strength of liquidity and debt-based ratios. This consistency across volatile time periods highlighted the enduring diagnostic value of these two financial signals.

Machine learning models delivered the strongest performance across nearly all evaluation metrics. CHAID and CART provided the highest and most stable sensitivity, accuracy and discriminatory power, while artificial neural networks performed well but required more complex calibration. Discriminant analysis consistently ranked as the weakest method, reinforcing concerns about relying on older statistical techniques in unpredictable economic environments.

Notably, economic shocks did not fundamentally undermine model performance. Even during pandemic-year instability, the patterns that distinguish healthy firms from distressed ones remained detectable when advanced modeling approaches were used. This result challenged assumptions that economic turmoil fundamentally distorts firm-level financial behavior in ways that defeat traditional predictive tools.

A single comprehensive model outperforms period-specific alternatives

The study was aimed at testing whether a unified model could match or outperform separate period-specific models. The expectation among many practitioners is that each economic environment, particularly one as extraordinary as the pandemic, requires a tailored predictive system. The researchers found the opposite.

The comprehensive CART model achieved sensitivity above 86 percent and accuracy near the same level. Its overall discriminatory ability effectively mirrored the best period-specific models and, in some cases, exceeded them. This stability held across all tested machine learning techniques, suggesting that well-structured unified models can adapt to structural differences without losing predictive power.

The period-indicator variable played an important role in enabling this flexibility. Rather than forcing the model to assume that all macroeconomic conditions remain constant, the indicator allowed the system to learn how firms behave differently during crisis periods. By capturing these contextual differences within a single architecture, the unified model eliminated the need for repeated redevelopment with each economic shift.

This finding has major consequences for financial institutions, credit rating agencies, regulators and lenders managing large portfolios. Maintaining multiple models is costly, time-intensive and prone to calibration errors. A single model that remains reliable through economic shocks reduces operational burden and allows for continuous risk monitoring without frequent retraining.

The results also challenge the common assumption that economic extremes break the statistical relationships on which bankruptcy prediction relies. Instead, the authors demonstrate that liquidity shortages, rising debt pressure and impaired profitability remain recognizable signals even when macroeconomic conditions deteriorate sharply.

The research underscores that small and medium-sized enterprises respond to economic shocks in ways that still reflect underlying balance-sheet strength. Firms entering a crisis with weaker liquidity or higher leverage tended to exhibit distress patterns that prediction models could still distinguish, regardless of whether the broader economy was stable or under strain.

Implications for Risk Management, Future Modeling and Economic Policy

The study provides a roadmap for how financial institutions and policymakers can improve risk assessment during volatile periods.

For lenders and credit managers, the findings suggest that investing in robust machine learning models may provide far greater long-term value than maintaining multiple versions of older statistical models. A single, flexible model simplifies implementation and ensures consistency in risk scoring, which is essential for evaluating SME creditworthiness in uncertain economic cycles.

The authors highlights the importance of handling extreme financial ratios with care. SME datasets often include unusually high or low values that reflect real distress signals rather than noise. Removing these values during preprocessing may weaken predictive accuracy. As a result, thoughtful data treatment becomes a core component of effective risk modeling.

The study also calls for future research to integrate additional macroeconomic variables. While the period-indicator variable improved model adaptability, incorporating metrics such as inflation, interest rate shifts, government support measures or sector-specific shocks could further strengthen predictions. Cross-validation techniques and improved missing-data handling are also recommended as next steps.

For policymakers, the findings offer insight into how early-warning systems could be standardized across agencies or across countries. A well-constructed unified model may be capable of supporting nationwide SME monitoring, particularly during recessions or crisis periods when business failures spike. This has implications for designing support schemes, identifying at-risk industries and managing credit guarantees.