Inteligencia artificial y evasión fiscal: un estudio exploratorio con ciencia de datos en América Latina

Abstract

Abstract: This thesis developed and evaluated a predictive model for estimating the probability of tax evasion using machine learning techniques applied to a synthetic database designed considering plausible economic assumptions and regional empirical evidence. The research includes a descriptive análisis of the data set, a comparative estimation of classification models, including logistic regression and nonlinear methods, and a performance evaluation using standard metrics and confusion matrices. Interpretability techniques such as SHAP were also incorporated to identify the variables whit the greatest impacto on the target variable. The results reveal that the Random Forest model outperforms the other models, supporting the usefulness of these tools in tax management. The main contribution of the research lies in the replicable methodological proposal that articulates economic theory, predictive modeling, and public policy implications.