DYNAMIC LOGISTIC REGRESSION MODELING FOR BANKRUPTCY RISK PREDICTION IN UKRAINIAN BUILDING SECTOR

7 СИСТЕМНИЙ АНАЛІЗ ТА НАУКА ПРО ДАНІ DOI: https://doi.org/10.20535/kpisn.2026.1.350247 UDC 519.237.5+336.201.2 Vadim V. Romanuke*, https://orcid.org/0000-0001-9638-9572 Vinnytsia Institute of Trade and Economics of State University of Trade and Economics, Vinnytsia, Ukraine, https://ror.org/03r6hpz93 *Corresponding author: DYNAMIC LOGISTIC REGRESSION MODELING FOR BANKRUPTCY RISK PREDICTION IN UKRAINIAN BUILDING SECTOR Background. Financial distress and bankruptcy forecasting has gained significant importance in the context of post-war economic recovery and restructuring of Ukrainian industries. Firms in the buildingand-construction materials sector operate under high uncertainty, where early detection of insolvency risk is crucial for maintaining financial stability. Logistic regression models, widely used in environmental and risk analytics, can be adapted to represent the nonlinear transition from solvency to bankruptcy as a probabilistic process. Objective. The paper aims to develop and evaluate both static and dynamic logistic regression mo dels for predicting the potential bankruptcy of a representative Ukrainian building-materials manufacturer. The dynamic extension seeks to capture the temporal persistence in financial performance through lagged predictors. Methods. A synthetic monthly dataset (5 years, 60 observations) is generated to simulate realistic financial ratios, including liquidity, leverage, profitability, efficiency, and interest coverage (solvency). The models are estimated in MATLAB using maximum-likelihood logistic regression with L2 regularisation (ridge penalty) to retain correlated predictors. The dynamic model incorporated one-period lags of all financial ratios and the one-period-lagged response. Predictive performance is assessed by accuracy, precision, recall, F1-score, and the confusion matrix. Results. The static logistic model achieved an average accuracy of around 89 %, yet it failed to predict two bankruptcy-risky months out of six ones. The dynamic model improved performance to 94 % accuracy, without missing a bankruptcy-risky month, but falsely labelling a non-risky month as bankruptcy-risky one. The signs of estimated coefficients are consistent with economic logic: higher leverage increases bankruptcy probability, whereas greater liquidity, profitability, efficiency, and solvency reduce it. Conclusions. Dynamic L2-regularised logistic regression provides an interpretable and computatio nally efficient framework for early bankruptcy prediction in Ukrainian industrial firms. The inclusion of lagged financial indicators enhances predictive stability and timeliness, enabling practical early-warning applications. Keywords: bankruptcy prediction; logistic regression; dynamic modelling; financial ratios; L2 regula risation; early-warning system; Ukrainian building sector. Introduction The probability of firm bankruptcy remains one of the central concerns in modern financial analytics, especially in economies exposed to structural trans- formations and unstable market conditions [1, 2]. In Ukraine, the construction and building-materials sectors have faced recurrent disruptions caused by macroeconomic turbulence, exchange rate volatility, and evolving regulatory frameworks. These circumstances Пропозиція для цитування цієї статті: В.В. Романюк, «Моделі динамічної логістичної регресії для прогнозування ризику банкрутства у будівельній галузі України», Наукові вісті КПІ, № 1, с. 7–18, 2026. doi: https://doi. org/10.20535/kpisn.2026.1.350247 Offer a citation for this article: Vadim V. Romanuke, “Dynamic logistic regression modeling for bankruptcy risk prediction in Ukrainian building sector”, KPI Science News, No. 1, pp. 7–18, 2026. doi: https://doi.org/10.20535/ kpisn.2026.1.350247 © The Autor(s). The article is distributed under the terms of the license CC BY 4.0 8 KPI Science News 2026 / 1 ISSN 2617-5509 (print), ISSN 2663-7472 (online) increase the risk of financial distress among manufacturing enterprises and, consequently, create the need for quantitative models capable of providing early warning signals about potential bankruptcy. Reliable bankruptcy prediction models not only help firms assess their own financial sustainability but also support creditors, investors, and policymakers in managing credit and investment risks more effectively [3, 4]. A considerable body of research on bankruptcy prediction has emerged since the mid-20th century, ranging from discriminant analysis [5] and logit models [6] to more recent approaches employing machine learning and hybrid ensemble techniques [7, 8]. Among these, logistic regression continues to occupy a prominent position because of its interpretability, robustness, and statistical groun ding [9, 10]. Logistic models explicitly connect the probability of bankruptcy to a vector of financial indicators (such as liquidity, leverage, profitability, and efficiency), allowing the analyst to quantify the marginal impact of each ratio on the likelihood of financial failure. Moreover, unlike linear discriminant methods, logistic regression does not impose the assumption of normally distributed predictors, which makes it particularly suitable for financial data often characterised by skewness, outliers, and bounded ratios [11]. While static logistic models have proven useful, they fail to account for the temporal dynamics inherent in financial distress processes. A firm’s transition toward insolvency rarely occurs as a sudden event. Instead, it develops gradually as liquidity deteriorates, leverage increases, or profitability declines over successive periods. Therefore, incorporating lagged predictors into the logistic framework enables one to capture persistence and delayed effects of financial indicators on bankruptcy risk. This leads to a dynamic logistic regression model, where the current probability of bankruptcy depends not only on present-period ratios but also on their historical trajectories. However, the inclusion of multiple correlated and lagged predictors increases the risk of multicollinearity and model instability. To mitigate this, L2 regularisation (ridge penalty) can be introduced into the likelihood function [12]. Regularisation shrinks coefficient magnitudes toward zero without eliminating predictors entirely, thereby preserving all available financial information while controlling overfitting [13, 14]. This makes the model more stable and generalizable, particularly when the number of predictors approaches or exceeds the number of observed periods, which is a typical limitation in firm-level bankruptcy datasets. Problem statement The objective of this study is to construct and analyse a regularised dynamic logistic regression model for bankruptcy prediction in the context of a Ukrainian building-materials manufacturer. A synthetic dataset will be generated to emulate realistic financial ratios and their temporal dependencies, reflecting the operational specifics of a (...truncated)