The limitations of business failure prediction models and possible avenues for progress

Éric Séverin*

Professor of Management Sciences

University of Lille

*Member of the faculty of the Business Science Institute

---

Introduction

Assessing the risk of bankruptcy remains a complex and imperfect exercise, despite the abundance of tools developed for this purpose. The COVID-19 health crisis has highlighted the fragility of many companies and, consequently, the weaknesses of existing forecasting mechanisms. This situation calls into question the real ability of current models to anticipate failures and effectively guide economic decisions.

The predominance of models based on financial statements

For several decades, financial institutions have relied primarily on rating systems based on the analysis of financial statements. These methods aim to synthesize a company's economic situation through numerical indicators in order to estimate the probability of short-term bankruptcy.

The gradual integration of artificial intelligence techniques has improved predictive performance by taking into account more complex relationships between variables. However, despite these advances, the models plateau at a success rate of around 90%, which remains insufficient in view of the financial stakes. These forecasting errors can lead to misallocation of resources and suboptimal decisions. The central question then becomes the reliability of these “statistical experts” on which economic actors base their choices.

The limits of excessive reliance on accounting data

The scope for improvement lies not only in algorithmic refinement, but also in the very nature of the information used. First, the almost exclusive reliance on accounting data is problematic. Although standardized and easily accessible, this data is not always completely reliable.

Earnings management practices, or even accounting manipulation, alter the quality of the information and distort the indicators used by the models. When the input data is biased, the relevance of the results is automatically reduced. Several studies show that incorporating accounting manipulation signals significantly improves the accuracy of bankruptcy predictions.

The decisive contribution of qualitative factors

Secondly, figures alone are not enough to reflect the economic reality of a company. Qualitative factors such as governance, shareholder structure, the health of the CEO, and the quality of relationships with partners play a decisive role in the survival of organizations, particularly small businesses.

Although more difficult to collect and quantify, this information greatly enriches the analysis and strengthens the predictive power of the models.

Thinking of failure as a process rather than an event

Finally, the failure of a company is more often part of a gradual process than a sudden event. Studying how data evolves over time provides a better understanding of the trajectories of financial deterioration. The history and dynamics of accounting indicators thus provide valuable signals about the accumulation of difficulties, beyond a snapshot of the situation at a given point in time.

Conclusion

If these avenues are still under-exploited by banks, it is mainly due to the high costs of collecting, processing, and interpreting information, as well as reluctance to use models that are perceived as opaque. Financial institutions often favor simple tools that are considered sufficiently effective in relation to their cost. It is now up to researchers to develop more efficient approaches that combine accuracy, transparency, and economic viability in order to improve bankruptcy prediction and, more broadly, the allocation of resources in the economy.

---

Discover Eric Séverin's publications on CAIRN.Info