Machine learning applied to detection and prevention of unpaid transactions
Tutor / director / evaluatorBarrado Muxí, Cristina
Document typeBachelor thesis
Rights accessOpen Access
Financial entities are increasingly looking for new technologies in order to improve their services and increase their profits. One of the challenges raised by a well-known financial entity is to reduce doubtful credit from the customers' transactions (face-to-face purchases). When transactions are unpaid by customers, the financial entity has to afford it and the corresponding amount of money is added to the doubtful credit. These transactions are accepted due to incorrect validation during the authorization. When the financial entity has connectivity problems and it is not able to validate a transaction, it allows a payment processor entity to validate the transaction. The payment processor entity does not have the legal permission to check if the customer has enough credit to assume the transaction. Therefore, the payment processor entity sometimes validates transactions that the customer will not pay. This project proposes to implement an algorithm able to predict which transactions will be paid by the customer and which not. This algorithm has to be executed by the payment processor entity. This project proposes to implement an algorithm able to predict which transactions will be paid by the customer and which not. Predictions are made without information about the available credit that the customer has. The algorithm has to be executed by the payment processor entity. After testing several models, it is concluded that the Random Forest algorithm trained with balancing techniques can predict 57.6% of all the transactions that will not be paid by customers. All with a certainty of 97.3%.