A FRAMEWORK FOR STATE TRANSITIONS ON THE SELF- ORGANIZING MAP: SOME TEMPORAL FINANCIAL APPLICATIONS PETER SARLIN, * ZHIYUAN YAO AND TOMAS EKLUND Turku Centre for Computer Science, Department of Information Technologies, Åbo Akademi University, Turku, Finland SUMMARY Self-organizing maps (SOMs) have commonly been used in temporal applications. This paper enhances the SOM paradigm for temporal data by presenting a framework for computing, summarizing and visualizing transition probabilities on the SOM. The framework includes computing matrices of node-to-node and node-to-cluster tran- sitions and summarizing maximum state transitions. The computations are linked to the SOM grid using transition- plane visualizations. We demonstrate the usefulness of the framework on two SOM models for temporal financial analysis: financial performance comparison of banks and monitoring indicators of currency crises. Copyright © 2012 John Wiley & Sons, Ltd. Keywords: state transitions; self-organizing map (SOM); temporal data; financial institutions; currency crises 1. INTRODUCTION Today’ s decision makers are often faced by enormous amounts of financial data available for decision- making purposes. Access to online financial databases, such as Thomson One, Amadeus and Bankscope, can provide nearly endless amounts of multivariate financial time-series data. However, owing to nonlinear relationships, high dimensionality and non-normality often inherent in financial data, utilizing these data can be a significant challenge for traditional statistical tools and spreadsheet programs. Instead, various data-mining and pattern recognition tools have been applied for this purpose. One potential tool is the self-organizing map (SOM; Kohonen, 1982, 2001), an unsupervised neural network-based projection and clustering method often used for exploratory data analysis. Although most of the early SOM applications have been in the area of medicine and engineering (Oja et al., 2002), the SOM has also been used in a large number of financial applications (Deboeck and Kohonen, 1998), including financial performance comparison (Back et al., 1998; Eklund et al., 2003), bankruptcy prediction (Martín-del-Brío and Serrano-Cinca, 1993; Kiviluoto, 1998), financial crisis monitoring (Sarlin and Marghescu, 2011; Sarlin and Peltonen, 2011), economic welfare analysis (Kaski and Kohonen, 1996), customer churn analysis and segmentation (Lingras et al., 2005; Kiang et al., 2006) and stock price forecasting (Afolabi and Olude, 2007; Hsu et al., 2009), just to name a few. The general SOM paradigm is an ideal tool for building visualization systems, as it reduces both dimensionality and data; however, manually identifying the positions and patterns in a SOM model is not necessarily a simple process. As the applications above illustrate, financial data typically belong * Correspondence to: Peter Sarlin, Department of Information Technologies, Åbo Akademi University, Turku Centre for Computer Science, Joukahaisenkatu 3–5, 20520 Turku, Finland. Email: psarlin@abo.fi Copyright © 2012 John Wiley & Sons, Ltd. INTELLIGENT SYSTEMS IN ACCOUNTING, FINANCE AND MANAGEMENT Intell. Sys. Acc. Fin. Mgmt. (2012) Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/isaf.1328