Application of Neural Networks for Predicting the Processing Time of Judicial Cases at the Superior Tribunal de Justiça (STJ)
Artificial Neural Networks; Judicial Prediction; SHAP Values; Machine Learning; Artificial Intelligence; STJ.
The slowness in judicial proceedings represents a critical challenge for the Brazilian Judiciary, negatively impacting operational efficiency and public perception of justice. This study aimed to develop and validate a predictive model using artificial neural networks, specifically the Multi-Layer Perceptron (MLP), to enhance the accuracy in predicting the overall processing duration of judicial cases at the Superior Tribunal de Justiça (STJ). A quantitative exploratory and explanatory methodology was applied, emphasizing advanced machine learning techniques, including RandomizedSearchCV for hyperparameter optimization. The dataset comprised over 600,000 judicial cases completed by the STJ from January 2023 to June 2024, extracted from the National Justice Council's (CNJ) "Justice in Numbers" panel. The developed algorithm demonstrated high predictive performance, achieving a determination coefficient (R²) of 0.9243 and a Mean Absolute Error (MAE) of 71 days. Additionally, the SHapley Additive exPlanations (SHAP Values) analysis identified "Judging Body" (Órgão Julgador) as the most influential variable affecting case duration predictions. Practically, this model can be integrated into the STJ's case management system, enabling strategic resource optimization and proactive management of lengthy judicial processes. Academically, this research contributes to expanding knowledge on artificial intelligence applications in the Judiciary.