Intelligent systems in accounting, finance and managementassessing predictive performance of ann-based classifiers, страница 3

ANNs in the form of SOMs have been used extensively in financial applications. Martín-del-Brío and Serrano Cinca (1993) propose the SOM as a tool for financial analysis. Among the problems associated with the use of traditional statistical models in financial analysis, Serrano Cinca (1996) mentions: ‘the difficulty of working with complex statistical models, the restrictive hypotheses that need to be satisfied and the difficulty of drawing conclusions by non-specialists in the matter’. The author proposes the SOM for predicting corporate failure and compares the SOM with LDA and a multilayer perceptron (MLP) trained with the backpropagation algorithm (BP). The database contains five financial ratios taken from Moody’s Industrial Manual from 1975 through to 1985 for a total of 129 firms, of which 65 are bankrupt and the rest are solvent. Serrano Cinca (1998a,b) extended the scope of the decision support system proposed in the earlier studies by addressing, in addition to corporate failure prediction, problems such as bond rating, the strategy followed by the company in relation to the sector in which it operates based on its published accounting information, and the comparison of the financial and economic indicators of various countries. Deboek (1998) outlines 12 financial, 4 economic and 5 marketing applications of the SOM. Another major SOM financial application is that of Back et al. (1998), which is an extended version of Back et al. (1996b) where the authors analyse and compare 120 pulp-and-paper companies between 1985 and 1989 based on their annual financial statements. Eklund et al. (2003) proposed the SOM as an alternative data mining technique for financial benchmarking of worldwide pulp-and-paper companies. Karlsson (2002) used the SOM to analyse and compare the companies from the telecommunication sector.

Koskivaara (2004) summarizes the ANN literature relevant to auditing problems. She concludes that the main auditing application areas of ANNs are as follows: material error, going concern, financial distress, control risk assessment, management fraud, and audit fee, which are all, in our opinion, particular cases of classification problems. In other words, in these applications ANNs were used, mainly, as classifiers. Going concern and financial distress can be considered to be particular cases of bankruptcy prediction.

Costea and Eklund (2004) compared three classifiers for financial performance classification of telecom companies and found out that the ANN performed similarly to statistical and induction techniques in terms of accuracy rates.

Coakley and Brown (2000) classified ANN applications in finance by the parametric model used, the output type of the model and the research questions.

Another technique to learn the connection weights for an ANN corresponds to the evolutionary approach and is represented by GAs. The literature in this area is relatively rich: Schaffer et al. (1992) listed 250 references that combined ANNs and GAs. GAs are used in the majority of these papers for solving the following problems: to find the proper architecture for the ANN, to reduce the input space to the relevant variables, and as an alternative way of learning the connection weights. One paper that uses GAs to solve the last two aforementioned problems is that of Sexton and Sikander (2001). The GA was found to be an appropriate alternative to gradient-descent-like algorithms for training neural networks and, at the same time, the GA could identify relevant input variables in the dataset.

Yao (1999) explores the possible benefits of combining ANNs and evolutionary algorithms (EAs). EAs refer to a class of population-based stochastic search algorithms such as evolution strategies, evolutionary programming and GAs that are based on principles of natural evolution (Yao, 1999: 1424). Yao presents different combinations between ANNs and EAs, such as evolution of ANN connection weights, evolution of ANN architectures and evolution of ANN learning rules. Through a large literature review, the author shows that the combinations of ANNs and EAs can lead to better models and systems than relying on ANNs or EAs alone. Yao (1999: 1427) presents tens of papers where one of the two training mechanisms (EAs and gradient-descent-like algorithms) was found to achieve better results than the other, and attributes these contradictory results to ‘whether the comparison is between a classical binary GA and a fast BP algorithm, or between a fast EA and a classical BP algorithm. [ . . . ] The best one is always problem dependent’. Yao and Liu (1997) proposed a new evolutionary system, i.e. EPNet, for evolving ANNs; they used evolutionary programming for evolving simultaneously ANN architecture and connection weights. The negative effect of the permutation problem^[2] (Hancock, 1992) was avoided simply by not using crossover operators. EPNet uses five different mutations: hybrid training, node deletion, connection deletion, node addition and connection addition. The goal of each mutation is to obtain better offspring. First, EPNet uses BP to train the network; the simulated annealing algorithm is then used in training and the mutation stops if the network is improved above some threshold. Otherwise, other mutations are applied gradually (Yao and Liu, 1997: 6, fig. 5). EPnet was applied in a number of experiments (N-parity problem, the two-spiral problem, four medical diagnosis problems, the Australian credit card assessment problem and the Mackey–Glass time-series prediction problem) which show that EPNet can discover ANNs that would be difficult to design by human beings. However, EPNet is suitable for applications where the time factor is not crucial, since ‘. . . it searches a much larger space than that searched by most other constructive or pruning algorithms. . . .’ (Yao and Liu, 1997: 20).