Since 2019 in Cuba, the need for more effective bank fraud detection systems has become apparent due to the increase in transactions and fraud. The problem is exacerbated by data imbalance, with the fraud class accounting for less than 1%. Banking data is inherently complex, making it difficult to separate effectively using traditional classification methods. Concept drift, where behavioural patterns change over time, further complicates accurate fraud detection and requires innovative solutions. Misclassification of data is due to the human factor, as some frauds go unreported to hide illegalities such as currency trafficking or tax evasion, while others are fabricated to obtain a refund from the bank. This study proposes an Autoencoder-based deep learning model for anomaly detection and feature selection. It is experimentally trained on normal transactions only, and the decision threshold is determined by the overall maximum value of the F-score. After identifying the best hyperparameters, feature selection is performed based on the sum of the individual reconstruction error of each feature with anomalies only, setting a new decision threshold for feature selection. Based on these results, a model with normal operations is trained with an F-score of 83.04% on the Kaggle credit card database. This approach provides an efficient solution to the problems.