Detecção de anomalias no transporte de mercadorias para o Ceará: uma análise com dados do MDF-e e técnicas de machine learning

Abstract

This work investigates the use of data from the Electronic Manifest of Tax Documents (MDF-e) in the road transportation of cargo destined for the state of Ceará, focusing on detecting anomalies that may indicate tax fraud. More than 580,000 interstate trips made in 2023 were analyzed, using variables such as geodesic distance, value and weight of the cargo, as well as the Federation Units traveled. An anomaly detection architecture composed of three complementary methods was applied: Robust Linear Regression (RLM), Isolation Forest (ISO) and Local Outlier Factor (LOF). The results revealed spatial and seasonal patterns of anomalies, with emphasis on states in the North and Central-West. The identified anomalies did not present a significant relationship with the value or weight of the cargo, suggesting the influence of logistical and operational factors. It is concluded that the analysis of electronic tax data, combined with machine learning techniques, can help prioritize suspicious routes and taxpayers, promoting greater efficiency in tax inspection and combating tax evasion.