Predicting Corporate Failure Using a Neural Network Approach

This paper investigates the performance of Artificial Neural Networks for the classification and subsequent prediction of business entities into failed and non-failed classes. Two techniques, back-propagation and Optimal Estimation Theory (OET), are used to train the neural networks to predict bankruptcy filings. The data are drawn from Compustat data tapes representing a cross-section of industries. The results obtained with the neural networks are compared with other well-known bankruptcy prediction techniques such as discriminant analysis, probit and logit, as well as against benchmarks provided by directly applying the bankruptcy prediction models developed by Altman (1968) and Ohlson (1980) to our data set. We control the degree of ‘disproportionate sampling’ by creating ‘training’ and ‘testing’ populations with proportions of bankrupt firms ranging from 1% to 50%. For each population, we apply each technique 50 times to determine stable accuracy rates in terms of Type I, Type II and Total Error. We show that the performance of various classification techniques, in terms of their classification errors, depends on the proportions of bankrupt firms in the training and testing data sets, the variables used in the models, and assumptions about the relative costs of Type I and Type II errors. The neural network solutions do not achieve the ‘magical’ results that literature in this field often promises, although there are notable 'pockets' of superior performance by the neural networks, depending on particular combinations of proportions of bankrupt firms in training and testing data sets and assumptions about the relative costs of Type I and Type II errors. However, since we tested only one architecture for the neural network, it will be necessary to investigate potential improvements in neural network performance through systematic changes in neural network architecture.     

Feedforward versus recurrent neural networks for forecasting monthly japanese yen exchange rates

Neural networks are a relatively new computer artificial intelligence method which attempt to mimic the brain's problem solving process and can be used for predicting nonlinear economic time series. Neural networks are used to look for patterns in data, learn these patterns, and then classify new patterns and make forecasts. Feedforward neural networks pass the data forward from input to output, while recurrent networks have a feedback loop where data can be fed back into the input at some point before it is fed forward again for further processing and final output. Some have argued that since time series data may have autocorrelation or time dependence, the recurrent neural network models which take advantage of time dependence may be useful. Feedforward and recurrent neural networks are used for comparison in forecasting the Japanese yen/US dollar exchange rate. A traditional ARIMA model is used as a benchmark for comparison with the neural network models.Results for out of sample show that the feedforward model is relatively accurate in forecasting both price levels and price direction, despite being quite simple and easy to use. However, the recurrent network forecast performance was lower than that of the feedforward model. This may be because feed forward models must pass the data from back to forward as well as forward to back, and can sometimes become confused or unstable. Both the feedforward and recurrent models performed better than the ARIMA benchmark model.The author wish to thank the reviewers Drs. Kraft and Radford for their helpful comments.     

A horse race of models and estimation methods for predicting bankruptcy

We use a comprehensive set of performance metrics to analyze the improvement in the classification power and prediction accuracy of various bankruptcy prediction models after adding governance variables and/or varying the estimation method used. In a sample covering bankruptcies of U.S. public firms in the period 2000 to 2015, we find that the addition of governance variables significantly improves the performance of all bankruptcy prediction models. We also find that the additional explanatory power provided by governance measures improves the further the firm is from bankruptcy, which suggests that governance variables may provide earlier and more accurate warning of the firm's bankruptcy potential. Our findings show that the performance of any bankruptcy prediction model is significantly affected by the estimation method used. We find that regardless of the bankruptcy model, hazard analysis provides the best classification and out-of-sample forecast accuracy among the parametric methods. Furthermore, non-parametric methods such as neural networks, data envelopment analysis or classification and regression trees appear to provide comparable and sometimes superior classification accuracy to hazard analysis. Lastly, we use the dynamic panel generalized methods of moments model to address concerns raised in prior studies about the susceptibility of similar studies to endogeneity issues and find that our findings continue to hold.     

Bankruptcy prediction using fuzzy convolutional neural networks

We propose a combined method for bankruptcy prediction based on fuzzy set qualitative comparative analysis (fsQCA) and convolutional neural networks (CNN). Currently, CNNs are being applied to various fields, and in some areas are providing higher performance than traditional models. In our proposed method, a CNN uses calibrated variables from fuzzy sets to improve performance accuracy. In addition, there are no published studies on the effect of feature selection at the input level of convolutional neural networks. Therefore, this study compares four well-known feature selection methods used in financial distress prediction, (t-test, stepdisc discriminant analysis, stepwise logistic regression and partial least square discriminant analysis) to investigate their effect on classification performance. The results show that fuzzy convolutional neural networks (FCNN) lead to better performance than when using traditional methods.     

Predicting direction shifts on Canadian–US exchange rates with artificial neural networks

The paper presents a variety of neural network models applied to Canadian–US exchange rate data. Networks such as backpropagation, modular, radial basis functions, linear vector quantization, fuzzy ARTMAP, and genetic reinforcement learning are examined. The purpose is to compare the performance of these networks for predicting direction (sign change) shifts in daily returns. For this classification problem, the neural nets proved superior to the naïve model, and most of the neural nets were slightly superior to the logistic model. Using multiple previous days' returns as inputs to train and test the backpropagation and logistic models resulted in no increased classification accuracy. The models were not able to detect a systematic affect of previous days' returns up to fifteen days prior to the prediction day that would increase model performance. Copyright © 2001 John Wiley & Sons, Ltd.     

Dynamic aspects in neural classification

When using neural networks for classification in an environment that is changing as time proceeds, methods for updating the parameters of the neural network should be considered in order to retain classification accuracy. Otherwise the neural network may lose performance due to structural changes. A classifier could be completely relearned from scratch at regular intervals. However, our experience from past credit scoring applications shows that users commonly prefer systems that change in as few cases as possible. Furthermore, this approach may be wasteful regarding the required computing time and that previously learned information will always be discarded. Therefore, we favor a methodology that attempts to detect changes and adapts a classifier only if inevitable. In this article, some methods for detecting and treating structural changes are applied to a credit scoring application. The results show that these methods may successfully be applied in a dynamic setting. Copyright © 1999 John Wiley & Sons, Ltd.     

PROBABILISTIC APPROACHES FOR CREDIT SCREENING AND BANKRUPTCY PREDICTION

Three probabilistic neural network approaches are used for credit screening and bankruptcy prediction: a logistic regression neural network (LRNN), a probabilistic neural network (PNN) and a semi‐supervised expectation maximization‐based neural network. Using real‐world bankruptcy prediction and credit screening datasets, we compare the three probabilistic approaches using various performance criteria of sensitivity, specificity, accuracy, decile lift and area under receiver operating characteristics (ROC) curves. The results of our experiments indicate that the PNN outperforms the other two techniques for decile lift and specificity performance metric. Using the area under ROC curve, we find that for bankruptcy prediction data the PNN outperforms the other two approaches when false positive rates (FPRs) are less than 40 %. LRNN outperforms the other two techniques for FPRs higher than 40 % for bankruptcy data. We observe that the LRNN results are very sensitive to the ratio of examples belonging to two classes in training data and there is a tendency to overfit training data. Copyright © 2012 John Wiley & Sons, Ltd.     

Effectiveness of Stochastic Neural Network for Prediction of Fall or Rise of TOPIX

Stochastic neural network is a hierarchical network of stochastic neurons which emit 0 or 1 with the probability determined by the values of inputs. We have developed an efficient training algorithm so as to maximize the likelihood of such a neural network. This algorithm enables us to apply the stochastic neural network to a practical problem like prediction of fall or rise of Tokyo Stock Price Index (TOPIX). We trained it with the data from 1994 to 1996 and predicted the fall or rise of 1 day ahead of TOPIX for the period from 1997 to 2000. The result is quite promising. The accuracy of the prediction of the stochastic network is the 60.28%, although those of non-stochastic neural network, autoregressive model and GARCH model are 50.02, 51.38 and 57.21%, respectively. However, the stochastic neural network is not so advantageous over other networks or models for prediction of the TOPIX used for training. This means that the stochastic neural network is less over fitting to the training data than others, and results in the best prediction. We will demonstrate how the stochastic neural network learns well non-linear structure behind of the data in comparison to other models or networks, including Generalized Linear model (GLM).JEL codes: D24, L60, 047     

A Neural Network Approach for Analyzing Small Business Lending Decisions

In this paper, we apply the neural network method to small business lending decisions. We use the neural network to classify the loan applications into the groups of acceptance or rejection, and compare the model results with the actual decisions made by loan officers. Data were collected from a leading bank in Central New York. The sample contains important financial statement and business information of borrowers and the loan officers' decisions. We conduct the network training on the data sample and find that the neural network has a stronger discriminating power for classifying the acceptance and rejection groups than traditional parametric and nonparametric classifiers. The results show that the neural network model has a high predictive ability. Our findings suggest that neural networks can be a very useful tool for enhancing small-business lending decisions and reducing loan processing time and costs.     

Ordinal Pairwise Partitioning (OPP) Approach to Neural Networks Training in Bond rating

Statistical classification methods such as multivariate discriminant analysis have been widely used in bond rating classification in spite of the limitations of the methodology. Recently, neural networks have emerged as new methods for business classification. This approach to neural networks training is to categorize a new instance as one of the predefined bond classes. Such a conventional approach has limitations in dealing with the ordinal nature of bond rating. In addition, most of the prior studies have used sample data which are evenly divided among the classes. However, the natural population in real application is usually unevenly divided among the classes. Under such circumstances, it is hard to achieve good predictive performance. As the number of classes to be recognized increases, the predictive performance decreases. In this article, to increase the predictive performance in real-world bond rating, we propose the ordinal pairwise partitioning (OPP) approach to backpropagation neural networks training. The main idea of the OPP approach is to partition the data set in the ordinal and pairwise manner according to the output classes. Then, each backpropagation neural networks model is trained by using each partitioned data set and is separately used for classification. Experimental results show that the predictive performance of the proposed OPP approach can be significantly enhanced, when compared to the conventional neural networks modeling approach as well as multivariate discriminant analysis. The OPP approach has two computation methods, and we discuss under which circumstances one method performs better than the other. We also show the generalizability of the OPP approach. © 1997 by John Wiley & Sons, Ltd.     

Stock Price Prediction Using Prior Knowledge and Neural Networks

In this paper we investigate ways to use prior knowledge and neural networks to improve multivariate prediction ability. Daily stock prices are predicted as a complicated real-world problem, taking non-numerical factors such as political and international events are into account. We have studied types of prior knowledge which are difficult to insert into initial network structures or to represent in the form of error measurements. We make use of prior knowledge of stock price predictions and newspaper information on domestic and foreign events. Event-knowledge is extracted from newspaper headlines according to prior knowledge. We choose several economic indicators, also according to prior knowledge, and input them together with event-knowledge into neural networks. The use of event-knowledge and neural networks is shown to be effective experimentally: the prediction error of our approach is smaller than that of multiple regression analysis on the 5% level of significance. © 1997 by John Wiley & Sons, Ltd.     

Bankruptcy prediction of financially stressed firms: an examination of the predictive accuracy of artificial neural networks

This is an extension of prior studies that have used artificial neural networks to predict bankruptcy. The incremental contribution of this study is threefold. First, we use only financially stressed firms in our control sample. This enables the models to more closely approximate the actual decision processes of auditors and other interested parties. Second, we develop a more parsimonious model using qualitative ‘bad news’ variables that prior research indicates measure financial distress. Past research has focused on the ‘usefulness’ of accounting numbers and therefore often ignored non‐accounting variables that may contribute to the classification accuracy of the distress prediction models. In addition, rather than use multiple financial ratios, we include a single variable of financial distress using the Zmijewski distress score that incorporates ratios measuring profitability, liquidity, and solvency. Finally, we develop and test a genetic algorithm neural network model. We examine its predictive ability to that of a backpropagation neural network and a model using multiple discriminant analysis. The results indicate that the misclassification cost of the genetic algorithm‐based neural network was the lowest among the models. Copyright © 2001 John Wiley & Sons, Ltd.     

A two‐step system for direct bank telemarketing outcome classification

A two‐step system is presented to improve prediction of telemarketing outcomes and to help the marketing management team effectively manage customer relationships in the banking industry. In the first step, several neural networks are trained with different categories of information to make initial predictions. In the second step, all initial predictions are combined by a single neural network to make a final prediction. Particle swarm optimization is employed to optimize the initial weights of each neural network in the ensemble system. Empirical results indicate that the two‐step system presented performs better than all its individual components. In addition, the two‐step system outperforms a baseline one where all categories of marketing information are used to train a single neural network. As a neural networks ensemble model, the proposed two‐step system is robust to noisy and nonlinear data, easy to interpret, suitable for large and heterogeneous marketing databases, fast and easy to implement.     

Modelling non-linearity in economic classification with neural networks

In this paper results are presented of a study on economic classification with neural networks. Comparison is made between neural networks and linear modelling techniques and, in particular, comments are made on the problem of overfitting and the estimation of prediction errors in cases where the available data sets are relatively small. It is shown that selecting network parameters by k-fold cross-validation and weight decay training are effective remedies for these phenomena. The conclusions are illustrated in two cases: predicting the volume of the mortgage market in the Netherlands and the classification of bond ratings. © 1997 John Wiley & Sons, Ltd.     

Machine learning as an early warning system to predict financial crisis

This paper studies on “Early Warning Systems” (EWS) by investigating possible contagion risks, based on structured financial networks. Early warning indicators improve standard crisis prediction models performance. Using network analysis and machine learning algorithms we find evidence of contagion risk on the dates where we observe significant increase in correlations and centralities. The effectiveness of machine learning reached 98.8%, making the predictions extremely accurate. The model provides significant information to policymakers and investors about employing the financial network as a useful tool to improve portfolio selection by targeting assets based on centrality.     

Probability distributions and leveraged trading strategies: an application of Gaussian mixture models to the Morgan Stanley Technology Index Tracking Fund

The purpose of this paper is twofold. Firstly, to assess the merit of estimating probability density functions rather than level or direction forecasts for one-day-ahead forecasts of the Morgan Stanley Technology Index Tracking Fund (MTK). This is implemented using a Gaussian mixture model neural network, benchmarking the results against standard forecasting models, namely a naïve model, a moving average convergence divergence technical model (MACD), an autoregressive moving average model (ARMA), a logistic regression model (LOGIT) and a multi-layer perceptron network (MLP). Secondly, we examine the possibilities of improving the trading performance of those models with confirmation filters and leverage. While the two network models outperform all of the benchmark models, the Gaussian mixture model does best: it is worth noting that it does well on a time series where the training period is showing a strong uptrend while the out-of-sample period is characterized by a downtrend.     

Data mining using a genetic algorithm-trained neural network

Neural networks have been shown to perform well for mapping unknown functions from historical data in many business areas, such as accounting, finance, and management. Although there have been many successful applications of neural networks in business, additional information about the networks is still lacking, specifically, determination of inputs that are relevant to the neural network model. It is apparent that by knowing which inputs are actually contributing to model prediction a researcher has gained additional knowledge about the problem itself. This can lead to a parsimonious neural network architecture, better generalization for out-of-sample prediction, and, probably the most important, a better understanding of the problem. It is shown in this paper that by using a modified genetic algorithm for neural network training, relevant inputs can be determined while simultaneously searching for a global solution. Copyright © 2001 John Wiley & Sons, Ltd.     

A neural network enhanced volatility component model

Volatility prediction, a central issue in financial econometrics, attracts increasing attention in the data science literature as advances in computational methods enable us to develop models with great forecasting precision. In this paper, we draw upon both strands of the literature and develop a novel two-component volatility model. The realized volatility is decomposed by a nonparametric filter into long- and short-run components, which are modeled by an artificial neural network and an ARMA process, respectively. We use intraday data on four major exchange rates and a Chinese stock index to construct daily realized volatility and perform out-of-sample evaluation of volatility forecasts generated by our model and well-established alternatives. Empirical results show that our model outperforms alternative models across all statistical metrics and over different forecasting horizons. Furthermore, volatility forecasts from our model offer economic gain to a mean-variance utility investor with higher portfolio returns and Sharpe ratio.     

Assessing the predictive performance of artifIcial neural network‐based classifiers based on different data preprocessing methods,distributions and training mechanisms

We analyse the implications of three different factors (preprocessing method, data distribution and training mechanism) on the classification performance of artificial neural networks (ANNs). We use three preprocessing approaches: no preprocessing, division by the maximum absolute values and normalization. We study the implications of input data distributions by using five datasets with different distributions: the real data, uniform, normal, logistic and Laplace distributions. We test two training mechanisms: one belonging to the gradient‐descent techniques, improved by a retraining procedure, and the other is a genetic algorithm (GA), which is based on the principles of natural evolution. The results show statistically significant influences of all individual and combined factors on both training and testing performances. A major difference with other related studies is the fact that for both training mechanisms we train the network using as starting solution the one obtained when constructing the network architecture. In other words we use a hybrid approach by refining a previously obtained solution. We found that when the starting solution has relatively low accuracy rates (80–90%) the GA clearly outperformed the retraining procedure, whereas the difference was smaller to non‐existent when the starting solution had relatively high accuracy rates (95–98%). As reported in other studies, we found little to no evidence of crossover operator influence on the GA performance. Copyright © 2005 John Wiley & Sons, Ltd.