Intelligent systems in accounting, finance and managementassessing predictive performance of ann-based classifiers, страница 5

3.  FINANCIAL PERFORMANCE CLASSIFICATION MODELS

In this section we present our two approaches for financial performance classifications. We describe two financial ANN classification models which differ based on the training mechanism that they use. The first model is based on gradient-descent-like training mechanisms (RT-based ANN classification model), and the other is based on the principles of natural evolution (GA-based ANN classification model). In Section 6 we apply these two models on 15 different datasets (one for each data distribution–preprocessing method combination). In Section 5 we describe how the different datasets with different distributions and preprocessing methods were obtained. In this section, we first present for each of the datasets obtained the preliminary steps necessary to build the class variable and to obtain the training and test datasets. Then, we present the empirical procedure for determining the ANN architecture. Finally, we describe the two ANN training mechanisms.

The generic classification model based on neural approaches is depicted in Figure 1.

Usually, when constructing classification models, the first step is to separate the data into training (TR) and test (TS) sets. In case the class variable is missing, as in our case, a clustering method could be applied to build this variable (Section 3.1). The second step consists of selecting the proper ANN architecture (Section 3.2). This step is concerned with determining the proper number of hidden layers and the right number of neurons in each hidden layer. Here, we also decide how the class variable should be codified. In other words, we ask how many neurons are necessary on the output layer to represent the class variable. The last step, ANN training, consists of specific tasks depending on the training mechanism used (Sections 3.3 and 3.4).

3.1.  Preliminary Steps

Next, we present the steps undertaken to create the training and test sets for the classification models, which we generically call preliminary steps:

1.  A clustering technique was applied to build the class variable for each dataset. We have usedthe fuzzy C-means (FCM) clustering algorithm (Bezdek, 1981) to build the clusters and, consequently, the class variable. The number of clusters is a parameter of our models. This was set to seven, as this was the proper number of classes reported in our previous studies (Costea and Eklund, 2003).

2.  In order to allow the ANN to learn the patterns within each cluster equally, we chose an equalnumber of observations for each cluster.

3.  Finally, we split the data into approximately 90% for training and the remainder for testing.

As was described above, we reduced as much as possible the subjectivity in determining the class variable by applying the FCM clustering algorithm directly. When fuzzy clustering algorithms are applied, every observation gets a vector representing its membership degree in every cluster, which indicates that observations may contain, with different strengths, characteristics of more than one

Table II. Characterization of clusters (Alcaraz and Costa (2004))