An algorithm sees a company’s trouble a year before the court does. Statistics Poland President’s Award for a UEW doctorate

Early warning systems for the deteriorating financial condition of enterprises in Poland have typically been built on small research samples and focused mainly on predicting bankruptcy in the legal sense. Dr Aleksandra Szymura, a graduate of the Doctoral School of the Wroclaw University of Economics and Business, developed machine learning models in her dissertation based on data from hundreds of thousands of Polish companies and demonstrated that the risk of loss of profitability, loss of liquidity or excessive debt can be effectively predicted a year in advance. Her doctoral dissertation took second place in the competition of the President of Statistics Poland for the best doctoral dissertation in the field of statistics defended in 2025.

Statistics Poland President’s Award for a Wrocław researcher

The competition of the President of Statistics Poland for the best master’s and doctoral theses in the field of statistics annually selects works that combine statistical craft with answers to concrete problems of the economy. This year’s edition was open to dissertations defended in 2025 at Polish universities (source: konkursnaprace.stat.gov.pl). The laureates received their diplomas during the 6th Congress of Polish Statistics, organised by Statistics Poland and the Polish Statistical Association, held in Warsaw on 1-2 July 2026 (source: kongres.stat.gov.pl; award ceremony: own information). 

The awarded dissertation is entitled “Zastosowanie metod uczenia maszynowego w procesie oceny spółek kapitałowych”.  It was written under the supervision of prof. dr hab. Marek Kośny, Vice-Rector for Science at WUEB, with dr Alicja Grześkowiak as auxiliary supervisor (source: dissertation title page, WIR Repository UEW). 

Early identification of deteriorating financial condition

Most existing early warning systems have concentrated on predicting bankruptcy in the legal sense, that is, the moment when insolvency is confirmed by a court. From the perspective of banks, investors and business partners, this is a stage at which the value of the enterprise and the prospects of recovering receivables may already be severely limited. 

Dr Szymura changes this perspective. Rather than waiting for a formal declaration of bankruptcy, the system she developed picks up earlier symptoms of a worsening financial situation and analyses the risk of three negative phenomena: loss of profitability, loss of liquidity and excessive debt, a year in advance (source: dissertation, Introduction p. 10 and Conclusion p. 218). 

The application of machine learning methods

The novelty of the dissertation lies in its simultaneously multi-stage and multi-aspect approach, in which machine learning methods are applied at every stage of the research procedure: from defining the scope of entities, through the selection of financial variables, to the construction of risk assessment models. The study covered Polish capital companies, that is, limited liability companies and joint-stock companies, and the final result takes the form of a synthetic measure. A company’s assessment is determined by the highest of the three types of risk analysed (source: dissertation, subchapter 5.6, formula 35). 

The scale of the study sets the work apart from the literature. Existing early warning systems were usually built on samples of fewer than one hundred observations. Dr Szymura initially retrieved from the Orbis database data on around a quarter of a million limited liability companies and more than six thousand joint-stock companies per year for the period 2017 to 2020, which she then subjected to preprocessing and analysis (source: dissertation, table 5, p. 127). In the modelling she used algorithms including XGBoost, LightGBM and CatBoost (source: dissertation, p. 110). 

The significance of the dynamics of financial change

The work also delivers findings whose relevance extends beyond the methodological approach applied. Year-on-year rate-of-change indicators, describing the dynamics of financial statement items, proved just as important in predicting negative events as classic ratios calculated as at the balance sheet date (source: dissertation, Conclusion, p. 219). Existing systems have rarely included variables of this kind. This suggests that approaching financial trouble may be signalled not only by the current state of a company’s finances, but also by the direction and pace of change.

Transparency of the models

The second distinguishing feature of the work is its treatment of model interpretability as a condition of practical application. The analytical procedure developed does not operate as a black box: the system indicates which variables influence the risk assessment and in which direction. The SHAP technique was used to explain the predictions (source: dissertation, p. 220). This matters in two ways. The analyst understands where a given assessment comes from, and a financial institution meets the growing regulatory requirements concerning the explainability of artificial intelligence models. The author stresses that the purpose of the models is to support the analyst in the decision-making process, not to replace them (source: dissertation, p. 222).

Potential applications 

The tool may serve capital market analysts, banks assessing credit risk, investors, auditors and public institutions, including the tax administration (source: dissertation, chapter 1.2). Early and interpretable assessment of companies’ financial condition translates into better credit, investment and supervisory decisions. 

Who will benefit from this research

The users of such tools are concrete: capital market analysts, banks assessing credit risk, investors, auditors and public institutions, including the tax administration (source: dissertation, chapter 1.2). Early and explainable assessment of companies’ financial health translates into better lending, investment and supervisory decisions.

Not the researcher’s first success 

This is another distinction for Dr Aleksandra Szymura. In 2024, while still a doctoral student, she was nominated by Perspektywy Women in Tech to the list of Top 100 Women in Data Science in Poland (source: uew.pl/sukces-naszej-doktorantki). 

The dissertation is available in open access in the WIR Repository of the Wroclaw University of Economics and Business: https://doi.org/10.48812/uew.wir/xsrl1624 

Author of text: Justyna Morawska-Płoskonka