Transformation of Local Finance Management in Ukraine Using Data Science Technologies and Predictive Analytics

Abstract

This paper presents a framework for transforming local finance management in Ukraine through Data Science and machine learning-based predictive analytics, and demonstrates one component of this framework - shortterm revenue forecasting - on real fiscal data. We identify three systemic data gaps constraining communities' financial autonomy: personal income tax (PIT) allocation by employer registration, incomplete property registries, and informal employment. As a methodological backbone we adopt the fourlevel descriptive-diagnostic-predictive-prescriptive analytics hierarchy [1], adapted to the public finance domain. The experimental component evaluates four machine learning models (Ridge, Random Forest, Gradient Boosting, XGBoost), classical SARIMA, and a seasonal naive baseline for monthly revenue forecasting of Lviv region local budgets using BOOST data (openbudget.gov.ua), 2018-2026. On the main holdout (test 2024, train 2019-2023, 5-fold TimeSeriesSplit), Ridge achieves $\text{MAPE}=7.365 \%, \mathrm{R}^{2}=0.464, \text{MAE}=311.4$ million UAH - outperforming SARIMA (MAPE 18.643%) by more than ten percentage points and all tree-based methods. A 26-oblast crossregion transfer experiment further reduces Ridge MAPE on Lviv 2024 to $5.186 \%(\mathrm{R}^{2}=0.764)$. On a 2025 out-of-distribution stress test (+19% YoY growth) tree models suffer a dramatic $\mathrm{R}^{2}$ collapse, empirically confirming the well-known extrapolation limitation of tree-based regressors [2]. The results show that algorithm choice is critical for fiscal time series with structural level shifts and that simple regularized linear models, combined with cross-region training, provide deployable forecasts for Ukrainian territorial communities.