The need to identify credit card fraud is one of the biggest challenges the world is facing nowadays. The number of electronic transactions is increasing year by year, even so much that it is basically impossible to ’manually’ check all of them. For this experiment we used data set that consist of about 285000 transactions that are collected in just 2 days in Europe. Using that information we can only imagine how many transactions take place every day in the world using various platforms. Banks use many different machine learning algorithms to detect fraudulent transactions among non fraudulent ones. The goal is to ’build’ a model that is precise and sensitive as possible, meaning that we don’t want a model that is excellent at detecting frauds, but also recognizes many valid transactions as fraudulent. The major problem we are dealing with is the class imbalance, ie. there are many more valid transactions than fraudulent ones (in credit card fraud detection data set for each fraudulent transaction there is 578 valid ones). In section 1.2 we explained why accuracy is not good performance metric when it comes to imbalanced data sets such as credit card fraud detection data set. Then we mentioned which measures should be used instead and why are we using those other metrics. In chapter 2 we described challenges that we encounter when dealing with class imbalance and how to deal with problems that arise from those challenges during training phase. Chapter 3 is dedicated to exploring resampling methods. We present in detail description of several different undersampling and oversampling methods and which advantages or disadvantages those methods bring. For the experiment we used many different classifiers which we describe in chapter 4. The results obtained by different combinations of resampling methods and classifiers are presented in the chapter 5. Based on the obtained results, we drew conclusions (chapter 6). The credit card fraud detection system should be a part of every bank’s security system and improving that system requires constant research and development (because new fraud techniques are also being developed). In other words, fraudsters realize when the bank have good anti-fraud system in place so they alter their techniques meaning that bank will have to update their system to adapt. Taking all that into consideration, resampling methods may not be good enough to stop fraudsters. We mentioned that resampling methods are not the only way to ’fight’ class imbalance. We can also use cost-sensitive approaches which don’t require adjusting imbalance ratio, but they demand better model and data understanding.