Elsevier

Decision Support Systems

Volume 107, March 2018, Pages 78-87
Decision Support Systems

Automatic feature weighting for improving financial Decision Support Systems

https://doi.org/10.1016/j.dss.2018.01.005Get rights and content

Highlights

  • We propose a novel methodology for improving DSS through automatic feature weighting.

  • We show that automatic feature weighting leads to an improvement in decision-making.

  • Naïve Associative Classifier performance was improved by the proposed methodology.

  • The statistical analysis shows that metaheuristics are good for feature weighting.

  • Differential Evolution is adequate for decision-making having low computational cost.

Abstract

We propose a novel methodology for improving financial Decision Support Systems (DSS) through automatic feature weighting. Using this methodology, we show that automatic feature weighting leads to a significant improvement in the performance of decision-making algorithms over financial data, which are the key of financial DSS. The statistical analysis carried out shows that metaheuristic algorithms are good for automatic feature weighting, and that Differential Evolution (DE) offers a good trade-off between decision-making performance and computational cost. We believe these results contribute to the development of novel financial DSS.

Introduction

In the world of financial companies, risk means the danger of loss [1]. Financial companies assume a certain credit risk in each one of the operations they carry out (loans, lines of credit, and guarantees, among others). These financial companies cannot know everything about of their clients and, on the other hand, a client's compliance with its obligations to a financial company depends on events in which no one can know whether they will occur. That is, there is an uncertainty associated to whether a customer will (or will not) pay their debt [2,3].

According to the Global Association of Risk Professionals (GARP), which is a non-for-profit association for world-class financial risk certification, the presence of risks in financial companies is relevant because the mismanagement of financial risks can lead to global financial crises, such as those registered in the world in a cyclical manner, since the origins of financial systems. In this context, financial companies have developed numerous methods to value credits and determine credit risks, as well as their management; all of this, in order to improve financial results and profitability. Some of these methods are closely linked to models of computer sensitivity, which indicates the importance of computational methods in financial risk management [4].

With the tremendous advance of technology, and specifically with the arrival in the world of powerful computer systems, in the 1980s financial institutions turned to see the possibilities that this type of technology offers them in their field of action. Thus, some researchers decided to venture into this type of topics, so that computing in its different manifestations, became part of the financing processes, including desktop computers.

According to [5], several types of DSS are being applied in the world of financial risks, such as neural networks and pattern recognition techniques. These approaches, among others, have provided a theoretical foundation for developing DSS (including expert systems) to estimate, for instance, the probability of bankruptcy and predict fraud.

While it is true that DSS are not the only option to solve problems related to financial risks, their importance as a good option has been growing in recent years [6]. They have been applied in quality control, financial forecasting, targeted marketing, bankruptcy prediction, optical character recognition, among other problems [[2], [3], [4], [5], [6], [7], [8], [9]].

In the present work, we address the following problems: credits risk, bankruptcy prediction, banknote authentication and bank telemarketing. To do this, we focused on a key aspect of DSS: the improvement of the decision-making (or prediction) algorithm. Those algorithms are the kernel of DSS, due to they help the system to process knowledge and to “decide” or to predict the most likely situation in the future. For instance, the prediction of risks in the financial environment can be seen as a decision-making problem. That is, by deciding that a company is in bankruptcy risk, we can reduce the uncertainty for the potential investors, and the risk of capital loss due to an investment in a company that may be in bankruptcy in short term. In this work we are focused on the improvement of an algorithm recently proposed by us, which was designed for the financial area: the Naïve Associative Classifier (NAC) [3]. While it is true that there are several algorithms in the literature that can be used by DSS to improve decision-making [[10], [11], [12], [13], [14], [15]], we have chosen the NAC because that it is one of the topics on which we are currently directing our research efforts.

The NAC has obtained very promising preliminary results in financial decision-making, and we hypothesize that if we develop a methodology for the automatic obtaining of weights for the attributes or features that describe financial data, we can improve the performance of the NAC in financial decision-making. Thus, we aim at improving the decision-making process, by improving the learning algorithm that guides this process.

Clearly, attribute selection is not the most important process to solve financial risk problems. Its importance lies in that, by improving the NAC, we can apply this model to different problems, ensuring better efficacy in the results; therefore, by making a good feature selection the NAC improves, and consequently, its application in DSS to reduce financial risks will be more effective. On the other hand, it is not in our interest to analyze the problems that people find when selecting attributes: we focus on selecting attributes to improve the NAC, and consequently to solve and predict financial risks in different areas. The results of this work support this affirmation.

Our main objective is to contribute to the development of DSS in the financial field, through the design of a methodology for the automatic adjustment of the attribute weights of financial data, by means of metaheuristic algorithms [[7], [8], [9]]. This methodology should improve the decision-making process in the financial area.

To test our hypothesis, we designed an experimental framework, taking into consideration the characteristics of financial data (such as data imbalance, presence of missing values, and presence of both numeric and categorical attributes), the functioning of metaheuristic algorithms, as well as the functioning of the NAC. We intended to compare the algorithms over the same conditions, and then to test the performance of the proposed automatic feature weighting methodology.

This paper offers the following contributions. First, we demonstrate how we can automatically obtain attribute weights in financial datasets, and how to improve the decision-making process in a Decision Support System of the financial field. Second, we examine the effect of metaheuristic algorithms for feature weighting, in the context of the NAC. The study's third contribution concerns the determination of the best metaheuristic for attribute weighting in the financial decision-making scenario. Our results show that DE is the most adequate algorithm for financial feature weighting.

The remainder of this paper is structured as follows. In Section 2, we review the related scientific literature on some of the problems that involve some type of risk within the financial environment. Then, in Section 3 we briefly describe the NAC, while in Section 4 the most relevant part of this article is profusely explained: the proposed methodology for financial feature weighting. After that, we present the results obtained, and discuss them in Section 5. The article ends with the conclusions in Section 6.

Section snippets

Related scientific literature

This section includes, and briefly describes, some of the research works that have addressed some of the four problems that we address in the present work: credits risk, bankruptcy prediction, banknote authentication and bank telemarketing. In this sense, the relationship of each of these works with our proposal is that the purpose is common: solve any of the four mentioned problems, using a computational method based on DSS. It is worth mentioning that, as far as we know, our proposal is the

Decision-making methods: the NAC classifier

One of the most important subsystems of a DSS is the Knowledge based subsystem (KBS), which deals with data, and process it in order to obtain (or to extract) valuable knowledge about the phenomenon under study. This knowledge is used by KBS in the decision-making process. Usually, the KBS includes a classification (or learning) algorithm that supports decision-making. In this section, we explain the functioning of the learning algorithm used as a case study in the proposed methodology: the NAC.

Overview

The proposed methodology (Fig. 1) consists of two phases. In the first, the attribute weights of the datasets are computed by means of a metaheuristic algorithm. The second phase consists of the application of the NAC classifier using the weights obtained in the previous phase.

In the first phase of the methodology, the metaheuristic algorithm will be applied ten times, independently. Each time, the algorithm will return the best individual (feature weights) obtained, according to a certain

Datasets

In order to evaluate the proposed methodology, different datasets belonging to the financial field were used (Table 2). They were obtained from the Machine Learning repository of the University of California at Irvine [32]. The datasets correspond to credits risk, bankruptcy prediction, banknote authentication and bank telemarketing. As can be seen in Table 2, all the datasets used have only two classes. As for the imbalance ratio, it can be observed that in at least six cases, this value is

Conclusions

In this research, a methodology was proposed to automatically perform the adjustment of attribute weights in financial data sets, and to improve the decision-making process in this area. As can be seen in the results obtained, the performance of the NAC was improved with the weights generated by the three metaheuristics used (DE, Genetic Algorithm and Novel Bat Algorithm), by determining that there is a significant difference in their performances through the Friedman test. This test stipulated

Acknowledgements

The authors would like to thank the Instituto Politécnico Nacional (Secretaría Académica, COFAA, SIP, CIDETEC, and CIC), the CONACYT, and SNI for their economic support to develop this work.

Yosimar Oswaldo Serrano-Silva obtained his Bachelor's degree in Computer Engineering in 2014 from the School of Computing of the National Polytechnic Institute and his MSc degree in Computer Science in 2017 from the Center for Computing Research of the same institution. His research interests include classifiers, financial forecasting, evolutionary algorithms and associative models.

References (38)

  • V. López et al.

    On the importance of the validation technique for classification with imbalanced datasets: addressing covariate shift when data is skewed

    Information Sciences

    (2014)
  • J.F. Díez-Pastor et al.

    Diversity techniques improve the performance of the best imbalance learning ensembles

    Information Sciences

    (2015)
  • I. Brown et al.

    An experimental comparison of classification algorithms for imbalanced credit scoring data sets

    Expert Systems with Applications

    (2012)
  • D. Murphy

    Understanding Risk: The Theory and Practice of Financial Risk Management

    (2008)
  • F.J.P. García

    Financial Risk Management: Identification, Measurement and Management

    (2017)
  • R. Apostolik et al.

    Foundations of Financial Risk: An Overview of Financial Risk and Risk-based Financial Regulation

    (2015)
  • D. Shirreff

    Dealing with Financial Risk

    (2008)
  • R. Storn et al.

    Differential evolution - a simple and efficient heuristic for global optimization over continuous spaces

    Journal of Global Optimization

    (1997)
  • Cited by (0)

    Yosimar Oswaldo Serrano-Silva obtained his Bachelor's degree in Computer Engineering in 2014 from the School of Computing of the National Polytechnic Institute and his MSc degree in Computer Science in 2017 from the Center for Computing Research of the same institution. His research interests include classifiers, financial forecasting, evolutionary algorithms and associative models.

    Yenny Villuendas-Rey obtained his Bachelor and MSc degrees (2005, 2007) on Applied Informatics at Universidad de Ciego de Ávila, Cuba. Her Ph.D. was received at Universidad de Las Villas, Cuba, in 2014. Areas of interest: Metaheuristics, Intelligent Decision Systems, Data Mining, and Software Engineering. Currently, she is with CIDETEC- IPN, Mexico. Member of the National Researchers System (SNI).

    Cornelio Yáñez-Márquez obtained his Bachelor degree (1989) on Physics and Mathematics at National Poly- technics Institute (IPN) Physics and Mathematics Superior School. His MSc (1995) and Ph.D. (2002) degrees were received at IPN Center for Computing Research (CIC). Currently a Researcher Professor, Titular C, at IPN CIC. Member of the National Researchers System (SNI). Areas of interest: Associative Memories, Neural Networks, Mathematical Morphology, and Software Engineering.

    View full text