Estratégias de modelagem e características influentes na composição do risco de rescisão em contratos públicos.

Abstract

Public contracts are essential instruments for formalizing agreements between public administrations and contracted companies. Federal, state, and municipal governments rely on these contracts to procure goods and services that address public needs. However, the growing volume of contracts presents significant challenges for public agencies and civil society in effectively monitoring and controlling their execution. In this context, machine learning has emerged as a promising strategy for prioritizing high-risk cases for inspection. This dissertation evaluates the practical effectiveness of a tool currently in use for identifying companies with risk-prone behaviors. Initial findings reveal that the tool's prioritization is ineffective, particularly due to the high number of companies incorrectly classified as high-risk. These results highlight the need to reframe the risk estimation problem at a more actionable level: the contract level. Using a novel dataset of contract-level labels, this study proposes and evaluates a predictive model for assessing municipal contract risk. The model incorporates features derived from the contract itself, the contracted company, and its supply history, while also exploring strategies to address class imbalance in risk categories. The results demonstrate the feasibility of contract-level risk estimation, with the model outperforming a random classifier on the dataset. However, the use of artificial class balancing techniques does not significantly enhance the model's predictive performance compared to weight balancing during training. Building on this baseline model, additional predictive features were introduced, including semantically latent features related to the contract’s subject matter and the GDP per capita of the contracting municipalities. The inclusion of latent features yielded model performance equivalent to or better than the baseline, while the use of GDP per capita provided a potential advantage. Based on the model with the highest AUC-PR, the study offers insights into the influence of predictive features on contract termination risk and their contribution to accurate contract classification. Finally, the study assesses the alignment between the model’s risk predictions and expert technical opinions on contract termination justifications. The findings indicate that the key risk factors identified by the model exhibit reasonable consistency with the real-world causes of contract terminations. This alignment underscores the model’s practical relevance and its potential to support decision-making in public contract management.