George D. Dounias

Hierarchical classification trees using type-constrained genetic programming

We investigate the capability of the genetic programming approach for producing hierarchical, rule-based, classification trees. These trees can be seen as an extension to the machine learning decision trees concept, where the predicates here can be complex expressions rather than just simple attribute-value comparisons. In order to improve the search ability and to produce meaningful results, type-constraints are applied to the genetic programming procedure, expressed in a BNF grammar. The model is tested in two well-known domains. In the Balance-Scale data, the system achieves in revealing the data creation rule. In the E-Coli Protein Localization Sites data, the system realizes a competitor to the literature classification score, retaining the solution comprehensibility. The training procedure is guided by an adaptive fitness measure. The overall performance of this system denotes its competitiveness to standard computational intelligent procedures.

An intelligent statistical arbitrage trading system

This paper proposes an intelligent combination of neural network theory and financial statistical models for the detection of arbitrage opportunities in a group of stocks. The proposed intelligent methodology is based on a class of neural network-GARCH autoregressive models for the effective handling of the dynamics related to the statistical mispricing between relative stock prices. The performance of the proposed intelligent trading system is properly measured with the aid of profit & loss diagrams.

Improving neural network based option price forecasting

As is widely known, the popular Black & Scholes model for option pricing suffers from systematic biases, as it relies on several highly questionable assumptions. In this paper we study the ability of neural networks (MLPs) in pricing call options on the SP in particular we investigate the effect of the hidden neurons in the in- and out-of-sample pricing. We modify the Black & Scholes model given the price of an option based on the no-arbitrage value of a forward contract, written on the same underlying asset, and we derive a modified formula that can be used for our purpose. Instead of using the standard backpropagation training algorithm we replace it with the Levenberg-Marquardt approach. By modifying the objective function of the neural network, we focus the learning process on more interesting areas of the implied volatility surface. The results from this transformation are encouraging.

Wavelet Neural Network for ground resistance estimation

This paper presents the results of a computational approach for the ground resistance of grounding systems, used for the safe operation of electrical installations, substations and power transmission lines and aspires to build a forecasting model for the ground resistance values. The proposed model consists of a Wavelet Neural Network, which has been trained and validated by field measurements, performed for the last three years. Several grounding rods, encased in ground enhancing compounds and natural soil, have been tested, so that a wide data set for the training of the network can be obtained, covering various soil conditions. The input variables of the network are the soil resistivity within various depths of the tested field, varying with respect to time and the rainfall height during the year. This work introduces the wavelet analysis in the field of ground resistance estimation and attempts to take advantage of the benefits of artificial intelligence.

On Detecting the Optimal Structure of a Neural Network Under Strong Statistical Features in Errors

The purpose of this article is to investigate the empirical performance of various statistical techniques for detecting the optimal structure of a neural network (NN) regression model. We are particularly concerned with the specification of the NN architecture when the error component is characterized by special statistical properties, such as heteroskedasticity and non-normality. We consider the sequential testing procedure based on standard Lagrange multiplier (LM) tests for neglected nonlinearity and also examine three modifications of this test that are robust to heteroskedasticity. By means of Monte Carlo simulations, we investigate the ability of these procedures to detect the right structure of the NN under different types of heteroskedasticity and noise distributions. Simulation results show that robustified LM tests allow the researcher to control the complexity of the NN without having to explicitly model all statistical aspects of the data-generating process, something which is not generally feasible with the standard LM test. The combination of robust regression-based testing with bootstrapping and generalized autoregressive conditional heteroskedasticity modelling techniques increases the efficiency of the statistical sequential procedure in eliciting the optimal NN architecture.

Optimisation of Complex Financial Models Using Nature-Inspired Techniques

This paper discusses applications of nature-inspired computational techniques in optimisation problems encountered in portfolio selection and applied econometrics. By means of an empirical study, we show how particle swarm intelligence can be effectively used in the estimation of a GARCH and an EGARCH model, two popular econometric parametrisations for the volatility of financial prices. We discuss several issues emerging from the application of nature-inspired techniques in financial optimisation

On the Use of a Combination Approach to Automated Knowledge Acquisition Based on Neural Networks and Fuzzy Logic with Regard to Credit Scoring Problems

This paper presents an application of automated knowledge acquisition in classification problems. The problem under examination is the classification of a sample of 130 enterprises into 5 different categories of credit risk. The data are both numerical and linguistic in nature, adding up to a total of 70 attributes. All data are obtained from an application form, which the company completes at a bank when seeking a loan and the only field which is provided by the bank is the final class. This classification task deals with statistical models, which try to give a clear view of the importance of every attribute by assigning a weight to it. Then, by multiplying each weight with the value of each corresponding attribute, the method produces a score. According to the range of the score the bank makes its decision. The problem, however, is that this method is very time consuming as it requires a long adaptation period in order to achieve the right results. The alternative solution presented in this paper is a combination of the fuzzy c-means clustering algorithm in conjunction with the Kohonen artificial neural network, which has a self-organizing Kohonen map. The basic idea behind this combination is to take advantage of the benefits offered by the two individual methods mentioned above and to compensate for each method’s short-comings. The c-means process can be integrated into the Kohonen algorithm by replacing the learning rate with membership values thus combining the fuzzy c-means algorithm with the structure and adaptive rules of the Kohonen network.