Innovative Applications of O.R. A knowledge based approach to loss severity assessment in financial institutions using Bayesian networks and loss determinants David Häger a,b, * , Lasse B. Andersen a a University of Stavanger, Department of Industrial Economics, Risk Management and Planning, 4036 Stavanger, Norway b Bayes Risk Management AS, P.O. Box 6069, 4088 Stavanger, Norway article info Article history: Received 10 June 2009 Accepted 14 June 2010 Available online 19 June 2010 Keywords: Risk management OR in financial institutions Bayesian networks Loss determinants Advanced measurement approach abstract Modelling loss severity from rare operational risk events with potentially catastrophic consequences has proved a difficult task for practitioners in the finance industry. Efforts to develop loss severity models that comply with the BASEL II Capital Accord have resulted in two principal model directions where one is based on scenario generated data and the other on scaling of pooled external data. However, lack of rel- evant historical data and difficulties in constructing relevant scenarios frequently raise questions regard- ing the credibility of the resulting loss predictions. In this paper we suggest a knowledge based approach for establishing severity distributions based on loss determinants and their causal influence. Loss deter- minants are key elements affecting the actual size of potential losses, e.g. market volatility, exposure and equity capital. The loss severity distribution is conditional on the state of the identified loss determinants, thus linking loss severity to underlying causal drivers. We suggest Bayesian Networks as a powerful framework for quantitative analysis of the causal mechanisms determining loss severity. Leaning on available data and expert knowledge, the approach presented in this paper provides improved credibility of the loss predictions without being dependent on extensive data volumes. Ó 2010 Elsevier B.V. All rights reserved. 1. Introduction The Basel II Capital Accord (BCBS, 2006) has introduced the financial industry to formalised management of operational risk through regulatory capital requirements. Three, increasingly demanding, approaches are available for establishing the regula- tory capital: the Basic Indicator Approach (BIA), the Standardized Approach (SA) and Advanced Measurement Approaches (AMA). While the regulatory capital for the BIA and SA are directly linked to the revenues generated at firm or business unit level, the AMA allows a firm to develop internal models for calculation of regula- tory capital. The internally developed models are subject to super- visory approval and it is expected that (at least) internationally active banks will, in time, achieve AMA approval (BCBS, 2006). In the wake of the Basel II requirements being adopted by national regulations, both practitioners and researchers are searching for an optimal AMA solution. The emergent best practice for modelling operational losses un- der the AMA is the Loss Distribution Approach (LDA). The LDA is derived from the accumulated claims process originally developed for modelling insurance claims (Bühlmann, 1970), and implies that two distributions are established: one describing event frequency and the other describing loss severity. The established distribu- tions are subsequently combined to create a compound distribu- tion describing total losses. A significant challenge of the LDA has turned out to be the modelling of loss severity for events with potentially extreme losses, specifically establishing the tail of the loss severity distribution. Examples of such events are unauthor- ised trading, unethical acts and aggressive sales practices which historically have resulted in extreme losses illustrated by the loss events of Societe Generale (2008), Barings Bank (1995), AIB/Allfirst (2002), NAB (2004) and Caisse d’Epargne (2008). Furthermore, such events may also result in the loss of licence to provide finan- cial services in a region or as a company altogether demonstrated by the cases of Daiwa Bank (1995), Terra Securities (2007) and Car- negie (2008). Most financial institutions have experienced very few, if any, extreme losses and survived; hence, the available inter- nal data is insufficient for reliable statistical inference. Still, we find that highly data-driven models, such as actuary models, are sug- gested for modelling loss severity (see e.g. Embrechts et al., 2003; Chavez-Demoulin and Embrechts, 2004; Nešlehová et al., 0377-2217/$ - see front matter Ó 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.ejor.2010.06.020 * Corresponding author at: University of Stavanger, Department of Industrial Economics, Risk Management and Planning, 4036 Stavanger, Norway. Tel.: +47 51 83 14 48; fax: +47 51 83 17 50. E-mail addresses: david.hager@uis.no (D. Häger), lasse.b.andersen@uis.no (L.B. Andersen). European Journal of Operational Research 207 (2010) 1635–1644 Contents lists available at ScienceDirect European Journal of Operational Research journal homepage: www.elsevier.com/locate/ejor