Artificial neural network application to business performance with economic value added.
An application of neural networks to classify the performance status of business firms is performed. An artificial neural network (ANN) model is developed using publicly available dataset as input and output variables. Several different neural network topologies are designed and applied to the datasets. A neural network model classifies both high and low business performance status. The ANN model can enhance strategic and managerial insights by providing meaningful financial information.
Since the development of an artificial neural network (ANN) approximately 50 years ago, it has received considerable attention from researchers. Neural network was once viewed as the theoretical foundation of building machine learning systems. It was proven to have many limitations. Recent neural network studies have overcome some previous limitations. ANN has been applied to a variety of problems for classification and prediction. It has been applied successfully for development of non-parametric statistical models. More robust outcome researches have been explored in the topics of pattern classification and pattern prediction. Neural network models are able to identify an existing pattern in data classification in different categories (Archer & Wang, 1993; Hsiao & Huang, 2002; Lotfi et al., 2000; Patuwo et al., 1993; Wang, 1995). Neural networks in business performance applications have been used for audit decision (Hansen et al., 1992; Lenard et al., 1995), initial public offerings (Jain & Nag, 1995), and multi-criteria decision-making (Malakooti & Zhou, 1994).
One of the most difficult issues in the ANN application discipline is to a proper utilization of a neural network model to implement a large-scale financial services research dataset. Moreover, if characteristics of the factors are complicated and complex, finding an optimal alternative is very difficult and may not get the best solution. Many researches have applied a neural network model to classify desired solutions or to improve methodological perspectives. Even though successful applications in ANN models have been made, little interests are given to the economic value added (EVA)-related business performance planning area. The utilization of neural network models in classification issues of management, finance, and other functional areas has been mostly limited to the usage of factors with continuous or ratio variables, rather than categorized values used in socioeconomic settings (Hart, 1992; Lee & Park, 2001; Sharda, 1994; Wilson, 1995).
The purpose of this study is to develop a neural network model to identify a proper classification between high and low business performance status. A neural network model is developed based on the publicly available financial dataset as input and output variables. The proposed model is expected to provide the factors affecting the classification of high/low EVA status, enhance business strategies to meet more appropriately business performance with financial measurement, and support decision-makers with more accurate information to implement.
Early studies have recognized that combining many simple nodes into neural network models is the source of increased computational power. The weights on the neural network are fixed so that the node performs a certain logic function with different nodes performing different functions. The nodes can be arranged into a network to generate any output that can be represented as an aggregation of logic functions. The learning mechanism has been designed for a neural network model. This mechanism is that the strength of the connection between them should be improved, if two nodes are active simultaneously. This idea is similar to the correlation matrix learning mechanism.
Many studies in neural networks have explored in business and financial areas during the last several decades. A study of Wong and Selvi (1998) provides an extensive literature analysis in finance application of neural network models. Neural network application to business performance and/or finance appeared in areas such as bankruptcy models (Chen, 1995; Fletcher & Goss, 1993; Markham & Ragsdale, 1995; Tam & Kiang, 1992), bonding models (Dutta et al., 1994; Quah et al., 1996), loan models (Bansal et al., 1993; Piramuthu et al., 1994), mutual fund models (Chiang et al., 1996; Hung et al., 1996), options models (Hutchison et al., 1994; Kaastra & Boyd, 1995; Swanson & White, 1995), real estate models (Borst, 1991; Collians & Evans, 1994; Worzala et al., 1995), stock prediction models (Wittkemper & Steiner, 1996; Wong & Long, 1995; Wood & Dasgupta, 1996).
A neural network is an imitating mechanism of human intelligence for the purpose of deriving certain performance characteristics. A neural network has been developed as generalization of a nonparametric methodology. A neural network model assumes that nodes have their own values for processing certain data, values are transmitted through nodes over connection links, each link has a weight that multiplies the values conveyed in a neural network, and each nodes applies an activation function to its input to identify the output value.
Assumption of information flow through the network is that a unit time step for data moves from one node to the next. This lead-time allows the network to model some perceptional processes. The most typical single layer network is consisted of an input layer connected by paths with fixed weights to associated nodes. The weights on the connection paths are adjustable. The learning rule of a single layer network uses an iterative weight adjustment. Learning rule can converge to the correct weights if weights allow the network to regenerate correctly all of the learning input and target output pairs. However, the theoretical proof on the convergence in iterative learning mechanism under suitable assumptions represents the limitations regarding what the single layer network type can learn.
A neural network model consists of a number of data processing layers interconnected in a network. Each layer has a mathematical mechanism with computational functions. A layer receives input values from other layers, and synthesizes the values by an input function. The layer creates output values by an output function. Then, the output values transferred to other layers as designed by the neural network architecture.
Characteristics of neural network have the architecture (the pattern of connections between the nodes), training or learning (the method of determining the weights on the connections), and the activation function. Each node is connected to other nodes by direct interaction links, each with an associated weight. This weight represents information used by the network to find a solution. Each node has an activation level that is a function of the inputs. Because a node sends its activation as a value to several other nodes, it can transmit one value at a time, while the value is transmitted to several other nodes.
An alternative learning rule can be used to adjust the connection weights to a unit whenever the unit response is incorrect. The unit response represents a classification of the input pattern. The least-mean-square or delta rule as an alternative learning rule adjusts the weights to reduce the difference between the input and the target output. This results in the smallest mean-squared error (MSE). The learning rule for a single-layer or multi-layers networks is interpreted as an adaptive linear system. The difference in learning rule leads to an improved capability of the neural network model to generalize. Some elaborated studies deal with associative memory neural networks with the development of self-organizing feature maps that use a topology for the cluster units. This feature truncates the linear output to prevent the output from being too large to get an optimal solution as the network iterates.
One of the advanced features is development of a back propagation algorithm (BPA) in a learning mechanism to train multi-layer networks. The BPA has a linear approximation function for the input layer and a logistic function for the output layer. The BPA using the hidden layer allows the data to be classified. A BPA has been developed to overcome the limitation of single-layer networks. It can solve complex problems, lacking a general methodology of learning a multi-layer neural network model. An improved model is derived from a number of neural network model based on fixed weights and adaptive activation. This neural network can use as associative memory networks and can be used to solve constraint satisfaction problems. Stochastic neural networks that weights or activations are changed on the basis of a probability density function incorporate such ideas as simulated annealing and Bayesian decision theory.
This study utilizes the financial dataset, which is a longitudinal study dataset of firms with high or low business performance. Information has been gathered on 764 high/low business firms. Pattern characteristics of financial services and changes in these factors over the course of the business can be analyzed for high/low EVA firms. Dozens of variables are selected initially from the original dataset. After controlling these variables, one output variable (dependent variable) and eight input variables (independent variable) are selected after diagnostic controlling them by the variance inflation factor (VIF) method for detecting multicollinearity among input variables. VIF values are less than 1.700 for multicollonearity diagnosis, so that multicollinearity problem is not significant among input variables. Out of 764 cases, valid cases are selected, implying a firm has either on high EVA status or on low EVA status. Selected cases provide valid information based on the responses of specific firms. Thus, cases have completed information on each firm, based on the selected input and output variables. The descriptive statistics for input and output variables are presented in Table 1.
The neural network model developed to classify a current EVA status of high/low business performance firms is the three-layer back-propagation algorithm neural network model. An input layer is used to represent a set of input variables. An output layer is used to represent an output variable. A hidden layer has an arbitrary number of the hidden nodes. Thus, numbers of hidden nodes are chosen arbitrarily and they derive different outcomes. Categorical variables have been recorded. For the first node in the output layer, it is given 0 if the current EVA status is low, and 1 if the current EVA status is high. For the second node in the output layer, the code is assigned to the reversed way. Table 2 presents descriptions about input variables and output variables in the neural network model.
Since no prior information is available as to how layers should be connected in the three-layer network, all nodes in the two adjacent layers are fully connected each other. The neural network model is for classifying the current EVA status of high/low business performance firms. Input layer has 8 nodes: RDI93, RDI94, RDI95, RDI96, INV96, ROE96, CAP96, SIZ96. Each of these variables is entered into the corresponding input layer of the neural network. These values are multiplied by model-generated random numbers so that they become the input values of the hidden layer. Each value is placed in a logistic function that computes the network activation of the hidden layer, becoming input values of the output layer. This value is entered into the same logistic function that computes the activation of the output layer, resulting in the output values: high EVA status or low EVA status. Thus, in actual practice, the output values could be considered as representing the likelihood of high EVA status or low EVA status in the current business performance of each firm.
The network architecture is designed to be the three-layer BPA networks. After configuring the neural network model, a learning rate, initial weight, and momentum learning epoch are assigned to the model to initiate the learning. Since assigning a learning rate, momentum, and number of epoch is arbitrary, a default value for each of them is assigned to the model. Once the model is designed, a certain percent over total pattern is extracted for the learning set and the test set. Approximation of 10 percent has been randomly extracted from the total pattern. An epoch is considered completed after the network examines all of the input and output patterns for all training sets. Epochs for training set are repeated for 200 times with a learning rate. In order to avoid the overfitting the network, the learning process was stopped when total number of epoch repeated reached at 20,000. A software NeuroShell[R]2 was used to implement this model (NeuroShell, 1993). Table 3 indicates a model statistics about the BPA neural network model.
This study proposed a neural network model to classify an EVA status of either high or low business performance. Neural network modeling of high/low EVA classification involves the interaction of many diverse factors. The relationships are often complex and complicated so that it makes the classification of outcomes very difficult and contingent.
This study presents an application of neural networks to classify by firms with the EVA status of business performance. A neural network model in classifying both the high and low status of EVA-related business performance is developed. Financial dataset is used in order to demonstrate the neural network's capability.
Neural networks are known to be able to identify relationships even when some of the input data are very complex and complicated. One of the advantages in ANN is to discriminate the linearly inseparable data. Even though neural network have been applied to a variety of areas such as management, finance, and other functional areas, a specific neural network model will be difficult to generalize to apply for a certain business environment with an appropriate interpretation. If the appropriate methodologies in various ANN design models are different, it would be interesting to see what impact would have on classification of high/low EVA firms.
Archer, N. P. & S. Wang. (1993). Application of the back propagation neural network algorithm with monotonicity constraints for two-group classification problems, Decision Sciences, 24, 60-75.
Bansal, A., R.J. Kauffman & R.R. Weitz. (1993). Comparing the modeling performance of regression and neural networks as data quality varies: A business value approach, Journal of Management Information Systems,10(1), 11-32.
Borst, R.A. (1991). Artificial neural networks: The next modeling/calibration technology for the assessment community, Property of Tax Journal,10(1), 69-94.
Chen, S., P. Mangiameli & D. West (1995). The comparative ability of self-organizing neural networks to define cluster structure, OMEGA, 23(3), 271-279.
Chiang, W.-C., T.L. Urban & G.W. Baldridge. (1996). A neural network approach to mutual fund net asset value forecasting, OMEGA, 24(2), 205-215.
Collins, A. & A. Evans. (1994). Aircraft noise and residential property values: an artificial neural network approach, Journal of Transport Economics and Policy, 28(2), 175-197.
Dutta, S., S. Shekhar & W.Y. Wong. (1994). Decision support in non-conservative domains: generalization with neural networks, Decision Support Systems, 11(5), 527-544.
Fletcher, D. & E. Goss. (1993). Forecasting with neural networks: an application using bankruptcy data, Information and Management, 24(3),159-167.
Hansen, J.V., J.B. McDonald & J.D. Stice. (1992). Artificial intelligence and generalized qualitative-response models: an empirical test on two audit decision-making domains, Decision Sciences, 23, 708-723.
Hart, A. (1992). Using neural networks for classification tasks-some experiments on datasets and practical advice, Journal of the Operational Research Society, 44, 1129-1145.
Hsiao, S.W. & H.C. Huang. (2002) & A neural network based approach for product form design, Design Studies, 23(1), 67-84.
Hung, S.-Y., T.-P. Liang & W.-L. Liu. (1996). Integrating arbitrage pricing theory and artificial neural networks to support portfolio management, Decision Support System, 18, 301-316.
Hutchison, A., A.W. Lo & T. Poggio. (1994). A nonparametric approach to pricing and hedging derivative securities via learning networks, The Journal of Finance, 49(3), 851-889.
Jain, B.A. & B.N. Nag. (1995). Artificial neural network models for pricing initial public offerings, Decision Sciences, 26, 283-302.
Kaastra, I. & M.S. Boyd (1995). Forecasting futures trading volume using neural networks, The Journal of Futures Markets, 15(8), 953-970.
Lee, C.W. & J. A. Park (2001). Assessment of HIV/AIDS-related health performance using an artificial neural network, Information and Management, 38(4), 231-238.
Lenard, M.J., P. Alam, G.R. Madey (1995). The application of neural networks and a qualitative response model to the auditor's going concern uncertainty decision, Decision Sciences, 26, 209-227.
Lotfi, K.G. & M.H. Choueiki. (2000) A neural network model for solving the lot-sizing problem, OMEGA, 28(2), 175-184.
Malakooti, B. & Y.Q. Zhou. (1994). Feedforward artificial neural networks for solving discrete multiple criteria decision making problems, Management Science, 40, 1542-1561.
Markham, I.S. & C.T. Ragsdale. (1995). Combining neural networks and statistical predictions to solve the classification problem in discriminant analysis, Decision Sciences, 26(2), 229-242.
NeuroShell[R]2 (1993) Frederick, MD: Ward Systems Groups.
Patuwo, E., M.H. Hu & M.S. Hung. (1993). Two-group Classification using neural networks, Decision Sciences, 24, 825-845.
Piramuthu, S., M.J. Shaw & J.A. Gentry. (1994). A classification approach using multi-layered neural networks, Decision Support Systems, 11(5), 509-525.
Quah, T.-S., C.-L. Tan, K.S. Raman & B. Srinivasan. (1996). Towards integrating rule-based expert systems and neural networks, Decision Support Systems, 17, 99-118.
Sharda, R. (1994). Neural networks for the MS/OR analyst: an application bibliography, Interfaces, 24, 116-130.
Swanson, N.R. & H. White (1995). A model-selection approach to assessing the information in the term structure using linear models and artificial neural networks, Journal of Business and Economic Statistics, 13(3), 265-275.
Tam, K.Y. & M. Kiang (1992). Managerial applications of neural networks: The case of bank failure predictions, Management Science, 38(7), 926-947.
Wang, S. (1995). The unpredictability of standard back propagation neural networks in classification applications, Management Science, 41, 555-559.
Wilson, R.L. (1995). Ranking college football teams: a neural Network approach, Interfaces, 25, 44-59.
Wittkemper, H.-G. & M. Steiner. (1996). Using neural networks to forecast the systematic risk of stock, European Journal of Operational Research, 90, 577-588.
Wong, B.K. & Y. Selvi. (1998). Neural network application in finance: a review and analysis of literature (1990-1996), Information and Management, 34, 129-139.
Wong, S.Q. & J.A. Long. (1995). A neural network approach to stock market holding period returns, American Business Review, 13(2), 61-64.
Wood, D. & B. Dasgupta (1996). Classifying trend movements in the MSCI USA capital market index--A comparison of regression, ARIMA and neural network methods, Computers and Operations Research, 23(6), 611-622.
Worzala, E., M. Lenk & A. Silva. (1995). An exploration of neural networks and its application to real estate valuation, The Journal of Real Estate Research, 10(2), 185-201.
Chang W. Lee, Jinju National University
Table 1: Descriptive Statistics of Input and Output Variables Variables X-bar SD MIN MAX N EVA96 -1216 3764 -1354 13364 73 RDI93 1.420 1.566 0 6.580 73 RDI94 1.449 1.534 0 6.734 72 RDI95 1.549 1.679 0 7.073 68 RDI96 1.505 1.563 0 5.719 48 INV96 -1.433 19.601 -164.65 16.211 73 ROE96 0.030 0.102 -0.538 0.276 73 CAP96 3.048 2.140 0.547 12.008 73 SIZ96 8.676 1.184 6.599 11.975 73 Table 2: Summary of Input and Output Variables Variables Description Input Variables RDIt Intensity of R&D(%) where t = 93, 94, 95, and 96 INVt Tangible Assets Investment(%) at time t ROEt Ratio of Return of Equity(%)at time t CAPt Intensity of Capital(Korean Won) at time t SIZt Firm Size (Korean Won) at time t Output Variable Current EVA status (either high status or low status) Table 3: Model Architecture for the BPA Neural Network Neural Network Modeling Parameters Total pattern 74 Learning set 61 Test set 13 with 20000 events Pattern selection Random Weight updates Momentum with 0.1 Learning epoch 200 with learning rate 0.1 Initial weight 0.3 Hidden Nodes H1=3, H2=7, H3=10, and H4=16
|Printer friendly Cite/link Email Feedback|
|Author:||Lee, Chang W.|
|Publication:||Academy of Information and Management Sciences Journal|
|Date:||Jan 1, 2003|
|Previous Article:||Toward an understanding of MIS survey research methodology: current practices, trends, and implications for future research.|
|Next Article:||Employee performance evaluation using the analytic hierarchy process.|