Contabilidad forense y blanqueo de capitalesaplicación del aprendizaje automático en un proceso judicial español

Resumen

This PhD Dissertation adds two new results in detecting signs of financial fraud: (1) the application of machine learning techniques to internal accounting databases of companies to detect money laundering, and (2) the offer of information to the investigating authorities on how the money laundering network is organized, with the objective of orientating the judicial investigation towards those companies or physical persons who present signs of suspicious patterns. Thus, in the context of a real macro-case on money laundering in which the author has collaborated as forensic accountant, this study analyses the database available of the operations carried out between a core company and a set of 643 supplier companies, 26 of which had already been identified a priori by the Judicial Police as fraudulent. Faced with a well-founded suspicion that other suppliers within the network might have committed criminal acts, and in order to better manage the scarce police resources available, machine learning techniques are proposed with two different approaches to detect patterns of fraud. The first proposed approach is the implementation of Neural Network models to the expert-assisted work for the detection of fraud operations. For this purpose, based on machine learning techniques, the network structure used is that proposed by Hastie et al. (2008): The Back-Propagation Network. In the second approach, it is proposed a more ambitious procedure to pattern detection than the previous one, in which Benford's Law (Nigrini and Mittermaider, 1997), a tool to characterize accounting records of the commercial operations between the main company and its supplier, is combined with four models of classification: Ridge Logistic Regression (LG) (Le Cessie and van Houwelingen, 1992), Artificial Neural Networks (NN) (Hastie et al., 2008), Decision Tree C4.5 (DT) (Quinlan, 1993 and 1996) and Random Forest (RF) (Breiman, 2001). Overall, the Random Forest showed the best results with the SMOTE transformation, obtaining 96.15% of true negatives (TN Rate) and 94.98% of true positives (TP Rate). The classification capacity of this methodology is undoubtedly very high.Thus, the machine learning techniques proposed in this paper represent an efficient and objective new tool for detecting fraudulent patterns of behaviour for the investigation of money laundering offences, allowing police investigators to focus the limited economic and human resources available in the judicial processes on those companies under suspicion who present a pattern of behaviour similar to that of previously recognized fraudulent companies. This PhD Dissertation is structured in two parts. On the first part, composed of three Chapters, establishes the theoretical framework on which the research is based. The first Chapter outlines the concept of money laundering and studies the tendency of this crime in Spain. Chapter II describes the process of management and access to information prior to the application of the proposed techniques (Data Pre-processing). Next, Chapter III specifies the methodology applied based on machine learning techniques for the detection of money laundering pattern. The second part is devoted to the presentation of the judicial process and the analysis of the results. After the presentation of the judicial process and the description of the sample, on the Chapter IV are presented the results obtained in the application of the machine learning techniques proposed in the two approaches. The PhD Dissertation ends with the conclusions and with proposals for further research.