Artificial Intelligence Methods for Risk Assessment
As mentioned above, a dynamic approach for the development and applicability of a Risk Assessment Matrix would be better suited for a changing risk environment than a static matrix. Different AI methods can be used in order to get such a dynamic functionality for the Risk Assessment Matrix. In this paper we propose the use of Neural Networks (NN) to establish the assessment and the probability of this assessment using the risk criteria establish by the user. In the financial industry the use of NN is that the result of the net can't be explained, because the net works as a "black box". A better approach can be Genetic Programming (GP), since the results can be explain from the GP equations. The use of GP will be a future work in the establishment of the Risk Assessment Matrix process.
Neural Networks method
Neural networks (NNs) are computer-based learning systems that have demonstrated utility in prediction, classification, and decision-making , ,  . NNs identify patterns between input (predictor) and target (criterion) variables. These systems are typically described using the biological analogy of brain neurons. This analogy centers on the fact that neural networks do not operate on a set of programmed instructions, as do statistical packages. Rather, they pass data through a multiple parallel processing entities (nodes or neurodes) that "learn" and adapt to the patterns that are presented to them. Data are not stored in these entities, nor are particular answers stored at particular addresses . Processing functions assume a pattern throughout the system. This pattern, developed in the iterative learning process, comes to represent the relationships between the input variables and the target variable .
Potential Advantages of the NN Prediction Approach
Neural networks offer some advantages over traditional statistical prediction methods. First, instead of assuming a particular form of relationship between independent and dependent variables, then using a fitting procedure to adjust the size of parameters in the model, neural networks construct a unique mathematical relationship for a given data set based on observed patterns between explanatory variables and designated outcomes . Second, the nonparametric nature of neural networks may make them particularly suited to social science data, where normality and linearity cannot be assumed. NNs have demonstrated capacity in handling interactions and nonlinearities . Third, neural networks have been demonstrated to be fairly robust in regard to handling input corrupted by random error , , . Fourth, there is some evidence that suggests that ANNs may excel over linear discriminant models with increasingly stringent thresholds for class membership. A "threshold" refers to a minimum score that must be reached for an example to be classified into one of two classes . Fifth, NNs may perform well on problems with low base rates Gordon . Finally, it appears that there is no significant disadvantage, other than length of training time, in including a large number of predictor variables in neural network analysis . The network will disregard variables that are not associated with the output by not assigning weights to those variables, leaving them at their near zero initial values. Also, it appears that intercorrelation among predictor variables does not detract from goodness of fit . This would suggest that neural networks are suitable in classification problems in which the number of predictor variables is large, and the intercorrelation among those variables is high.
Neural networks are not panaceas; they do not substitute for wise variable selection and accurate measurement of data. However, it can be argued persuasively that neural network models are often superior to alternatives in terms of predictive power . Neural networks appear to have potential for making outcome predictions in areas where independent variables are likely to be intercorrelated and where data are affected by moderate levels of random error and missing data. This is likely to be the case in Risk assessment for online trust and security solution in the Financial Industry.
Applying NN for Risk assessment for Online trust and security solutions
For training the NN we have information about risks and their specific risk criteria. We have information about the assessment of the risks and risk groups in relation with the risk criteria and information of global risk if we join all individual risks.
The input for the NN is a vector that contains information about all risk criteria, and the output of the NN it could be a vector containing the assessments of the Risk Assessment Matrix and the probability of each assessment.
Another possible solution for applying NN to the Risk Assessment is that the output of the net is the assessment of the global risk with an associated probability.
The authors have explained the reasons for a Neural Networks based Risk Assessment Matrix for the selection process of Online Trust and Security Solutions. With the combination of multiple Risk Areas and intelligent assessment methods Organisations will be able to speed up the selection process and to reduce the assessment mistakes due to a large risk criteria base, selected over time. It is now necessary to define in detail the specific Risk Appearances, Risk Criteria and Risk measures for Online Trust and Security Solutions, followed by the definition of input variables for the NNs, to be able to finally proceed with the combined Matrix set-up.
Risk Criteria and measures have to be specified towards their correlation and applicability – how to measure specific risks and how to compare them with others. Although a scientific approach and methodology will be vital for the basic structure of such a Matrix it is indispensable to apply the Matrix to practical situations.
 Glasserman, Paul: The quest for precision through Value at Risk. In: Pickford, James (Executive Editor): Mastering Risk, Volume 1: Concepts. Pearson Education Limited (2001) 109-114.
 KonTraG, Bundesgesetzblatt I 1998, 30.04.1998, 786-794.
 Götze, U., Henselmann, K., Mikus, B. (Editors): Risikomanagement. Physica-Verlag (2001).
 Cross, S.S., Harrison, R.F., & Kennedy, R.L. (1995). Introduction to neural networks. The Lancet, 346, October 21, 1075-1079.
 Galletly, C.A., Clark, C.R., & McFarlane, A.C. (1996). Artificial neural networks: A prospective tool for the analysis of psychiatric disorders. Journal of Psychiatry and Neuroscience, 21 (4), 239-247.
 Patterson, D.A. & Cloud, R.N. (2000). The application of artificial neural networks for outcome prediction in a cohort of severely mentally ill outpatients. Forthcoming in The Journal of Technology for Human Services.
 Y., Shen, Y., Shu, L., Wang, Y., Feng, F., Xu, K., Qu, Y., Song, Y., Zhong, Y., Wang, M., & Liu, W. (1996). Artificial neural network to assist psychiatric diagnosis. British Journal of Psychiatry, 169 (1), 64-67.
 G.D. (1998). Neural Networks: An Introductory Guide for Social Scientists. Thousand Oaks, California. Sage.
 Marshall, D.B., & English, D.J. (2000). Neural network modeling of risk assessment in child protective services. Psychological Methods, 5(1) in press.
 Hartzberg, J., Stanley, J., & Lawrence, M. (1990). Brainmaker user's guide and reference manual (Computer program manual). Sierra Madre, CA: California Scientific Software.
 R.P. (1987). An introduction to computing with neural nets. IEEE ASSP Magazine, April, pp. 4-22.
 Weiss, S.M., & Kurlikowski, C.A. (1991). Computer systems that learn: Classification and prediction methods from statistics, neural nets, machine learning, and expert systems. San Mateo, CA: Morgan Kaufmann.
 Gordon, J.S. (1991). (Probability correct classification as a function of increasing decision thresholds). Unpublished raw data.
 Gordon, J.S. (1992). A neural network approach to the prediction of violence. Unpublished doctoral dissertation. Oklahoma State University.