Modelo de datos para predecir quiebras de pymes importadoras en Guayaquil : análisis con árboles de decisión y random forest.

Abstract

This research assesses the risk of operational bankruptcy linked to financial pressure in an import company in Guayaquil (FEGOCAR) through machine learning, using an alert tool for internal control. A supervised classifier was trained with 4,531 transactions between 2021 and 2025 using economic and operational variables such as price, transport, insurance, quantity, net weight, country of origin, type of goods, model, and supplier. Data cleaning, conversion of the monetary format to a standard, one-hot encoding of categorical variables, and the development of derived metrics (unit_price, unit_weight, and logistics_cost_ratio) were carried out. To classify operations as high or low risk, a Random Forest and a Decision Tree with class balance were trained. The operational risk KPI stood at 48.78% (2,210 high-risk transactions). Forest achieved an AUC of 0.824 and an accuracy of 0.827; the Decision Tree achieved an AUC of 0.817 and an accuracy of 0.819. The importance of variables yielded the following predictors with the greatest contribution: unit_weight, logistics_cost_ratio, and unit_price. Rankings were generated by supplier: risk rate and total expenditure, as well as readable tree rules and a calculator to estimate the risk in new transactions, useful for targeted auditing, logistics cost validation, and control of unusual transactions.