Ignacio Olmeda

Hybrid Classifiers for Financial Multicriteria Decision Making: TheCase of Bankruptcy Prediction

This paper compares the accuracy of parametric and nonparametric classifiers on the problem of predicting Bankruptcy. Among the single classifiers, an artificial neural network is found to provide the best results. Two ways of combining classifiers are considered and an additive aggregation method is proposed. We show that both ways of combining produce classifiers whose forecasts are more accurate than the ones obtained with any single model. We suggest that an optimal system for risk rating should combine two or more different techniques.

Self-organizing maps could improve the classification of Spanish mutual funds

In this paper, we apply nonlinear techniques (Self-Organizing Maps, k-nearest neighbors and the k-means algorithm) to evaluate the official Spanish mutual funds classification. The methodology that we propose allows us to identify which mutual funds are misclassified in the sense that they have historical performances which do not conform to the investment objectives established in their official category. According to this, we conclude that, on average, over 40% of mutual funds could be misclassified. Then, we propose an alternative classification, based on a double-step methodology, and we find that it achieves a significantly lower rate of misclassifications. The portfolios obtained from this alternative classification also attain better performances in terms of return/risk and include a smaller number of assets.

Is the predictability of emerging and developed stock markets really exploitable

Abstract A number of recent papers have analyzed the degree of predictability of stock markets. In this paper, we firstly study whether this predictability is really exploitable and secondly, if the economic significance of predictability is higher or lower in the emerging stock markets than in the developed ones. We use a variety of linear and nonlinear – Artificial Neural Networks – models and perform a computationally demanding forecasting experiment to assess the predictability of returns. Since we are interested in comparing the predictability in economic terms we also propose a modification in the nets’ loss function for market trading purposes. In addition, we consider both explicit and implicit trading costs for emerging and developed stock markets. Our conclusions suggest that, in contrast to some previous studies, if we consider total trading costs both the emerging as well as the developed stock returns are clearly nonpredictable. Finally, we find that Artificial Neural Networks do not provide superior performance than the linear models.

Bankruptcy Prediction with Artificial Neural Networks

In this paper we compare the forecasting accuracy of feedforward neural networks against various competing models (C4.5, MARS, Discriminant Analysis and Logit) on the problem of predicting bankruptcy. The neural network model is found to provide generally better results, though the computational effort is several orders of magnitude higher. We also consider mixtures of the methods and show that many of these are always more accurate than any single method. We suggest that an optimal system for risk rating should include two or more of the models considered.

Risk forecasting models and optimal portfolio selection

This study analyses, from an investors perspective, the performance of several risk forecasting models in obtaining optimal portfolios. The plausibility of the homoscedastic hypothesis implied in the classical Markowitz model is dicussed and more general models which take into account assymetry and time varying risk are analysed. Specifically, it studies whether ARCH-type based models obtain portfolios whose risk-adjusted returns exceed those of the classical Markowitz model. The same analysis is performed with models based on the Lower Partial Moment (LPM) which take into account the assymetry in the distribution of returns. The results suggest that none of the models achieve a clearly superior average performance. It is also found that models based on semivariance perform as well as those based on the variance, but not better than, even if the evaluation criterion is based on the Reward-to-Semivariance ratio. When attention turns to the analysis of worst case performance, the results are clearly different. Models which employ LPM with a high degree of risk aversion (n>2) as the risk measure are consistently superior to those which employ a symmetric measure, either homoscedastic or heteroscedastic.

Modelos Paramétricos y no Paramétricos en Problemas de Credit Scoring

RESUMEN Dada la importancia creciente que esta cobrándose la actividad crediticia en la gestión diaria de los bancos, comienza a ser imprescindible la utilización de modelos de clasificación automáticos que faciliten la concesión o no del crédito solicitado con alto grado de exactitud, de manera que permita reducir la morosidad. En el trabajo que presentamos se realiza un exhaustivo estudio de la capacidad predictiva de dos modelos paramétricos (Análisis Discriminante y Logit) y cinco no paramétricos (Árboles de regresión, Redes Neuronales Artificiales, Algoritmo C4.5, Splines de Regresión Adaptativa Multivariante y Regresión Localmente Ponderada) en un problema de concesión de tarjetas de crédito.

Forecasting Exchange Rates Volatilities Using Artificial Neural Networks

This paper employs Artificial Neural Networks to forecast volatilities of the exchange rates of six currencies against the Spanish peseta. First, we propose to use ANN as an alternative to parametric volatility models, then, we employ them as an aggregation procedure to build hybrid models. Though we do not find a systematic superiority of ANN, our results suggest that they are an interesting alternative to classical parametric volatility models.

Complex Permittivity Estimation by Bio-inspired Algorithms for Target Identification Improvement

Identification of aircrafts by means of radar when no cooperation exists (Non-Cooperative Target Identification, NCTI) tends to be based on simulations. To improve them, and hence the probability of correct identification, right values of permittivity and permeability need to be used. This paper describes a method for the estimation of the electromagnetic properties of materials as a part of the NCTI problem. Different heuristic optimization algorithms such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO), as well as other approaches like Artificial Neural Networks (ANN), are applied to the reflection coefficient obtained via free-space measurements in an anechoic chamber. Prior to the comparison with real samples, artificial synthetic materials are generated to test the performance of these bio-inspired algorithms.

Some remarks on the application of artificial neural networks to optical character recognition

In this paper we analyze one important aspect related to handwritten Optical Character Recognition, specifically, we demonstrate that the standard procedure of minimizing the number of misclassified characters could be inconsistent in some applications. To do so, we consider the problem of automatic reading of amounts written in bank cheques and show that the widely used confusion matrix does not provide an appropriate measure of the performance of a particular classifier. We motivate our approach with some examples and suggest a novel procedure, using real data, to improve the performance by considering the true economic costs of the expected misclassification errors.

An Application of Hybrid Models in Credit Scoring

The predictive capability of parametric and non-parametric models in solving problems related to financial classification has been widely proved in empirical research carried out in the financial field, particulary in problems like bond rating, bankruptcy prediction and credit scoring. However, recently, it has been shown that a combination of different models generally reduces the prediction error, so that the best alternative to consider may not be a specific model but a combination of them. In this paper, we study hybrid systems based on the aggregation of individual (parametric and nonparametric) models. Our hybrids are built by using both parametric and non parametric models as the system aggregation. We present an example of this procedure on the problem of classifying credit card applicants.