| dc.description.abstract | With all the technological progress we have witnessed during the last couple of decades, it comes as no surprise that modern fraudsters use highly complex methods to disguise their suspicious activities. Traditional knowledge-based systems have proven their value in detecting fraud, but with fraudsters constantly adapting their behavior, they soon become outdated. Therefore, there is a strong need for modern data-driven approaches that could help us to unveil complex fraudulent structures. In this master thesis we focus on a specific type of financial fraud, namely money laundering. A typical money laundering scenario consists of series of seemingly legitimate transactions that are intended to transform illegally-gained proceeds into legitimate money. Most of the research throughout the last two decades focuses on other types of fraud such as credit card fraud or social security fraud. The traditional approach utilizes the local, entity-based data to predict the probability of committing fraud. Since fraud is a social phenomenon, studying entities as if they were acting independently of each other might not be enough. Recent techniques therefore try to incorporate inherent network structure into their models. In this work we try to apply some of the recent network-based techniques on transaction data provided by DNB, the largest Norwegian bank. The main aim of this thesis is to construct covariates based on network data and analyze whether these additional covariates can help us to distinguish reported and non-reported cases of money laundering. In order to do so, we first construct a network from the transaction data and extract two types of features, namely so called RFM features (transaction summaries) and network features. These features range from the proportion of transactions made in foreign currency in different time windows to various network features such as the PageRank scores or the clustering coefficient based on a local network structure. Then we build supervised learning models utilizing these additional sources of information. We have opted for a regularized version of logistic regression and different tree-based ensemble methods. Our results show that adding these additional covariates brings only marginal increase in model performance, which varies from model to model. However, we strongly believe, that this lack of success can partly be ascribed to the type of response variable. In other words, having a target variable that distinguishes actual money laundering cases from legitimate behavior, could result in a very different conclusion. | eng |