Evaluating and forecasting banking crises through neural network models: An application for Turkish banking sector
Introduction
In recent years, banking crises have occurred more frequently in the countries which have developing and transition economies. And some of them had very high costs as they spread into other regions. And serious international institutions’ interventions have been required. Thus, the need for the studies that analyze banking and other financial crises, present various forecasting models and suggest precautions for them has increased. This paper aims to analyze the banking crises process and to establish a forecasting model through artificial neural networks.
Neural networks are technologies that acquire the ability to learn for the computers. They teach the relationships between the inputs and outputs of the events to the computers by using patterns. By the help of taught data various generalizations are made, similar events are interpreted, required decisions are made and related problems are solved.
Artificial neural networks (ANNs) are generally the software systems that imitate the neural networks of the human brain (Trippi & Turban (1996, pp. 4–5)). It is also possible to accept the ANNs as a parallel distributed data process system. And ANNs can be applied successfully in learning, relating, classification, generalization, characterization and optimization functions. Because ANNs have the ability to work with incomplete data, posses error tolerance, and show graceful degradation, they can easily form models for complex problems. Especially in the development of solutions for semi-structural or non-structural problems, artificial neural network (ANN) models can have very successful results. Moreover, they can be cheaper, faster and more adaptable than traditional methods. Most of the problems that the financial managers face are semi-structural. So the qualifications of the ANN models can be very helpful in the formulations and the solutions of them. It is possible to evaluate and forecast banking crises with the help of ANN models.
After introducing the applications of ANNs in finance in the first part of this paper, the formation of the two ANN models for evaluating and forecasting banking crises is explained in the second part. Also, the use of the Taguchi Approach in the optimization of the network topologies is proposed in this part. The results of the two ANN models is found in part three. And finally, the conclusions and future related studies are presented.
Section snippets
Application of neural networks in financial and commercial domains
Since the 1940s, ANNs have been used in various applications in engineering. As artificial intelligence technologies have improved, they also began to be used in the solution of medical, military and astronomical problems. Today, their potential success in formulating solutions for financial problems is area of interest.
Real-world problems have qualitative and quantitative dimensions. Real-world users of decision tools can benefit from more qualitative capabilities and increased user
Formation of the neural network models
In the formation of an ANN system, a sustainable neural network model based on the nature of the problem should be selected and a neural network should be constructed according to the characteristics of the application domain. Familiarity with existing applications of other financial studies can help determine the appropriate network topology (architecture) and selection of the best-suited computational model for learning and inference.
These application process of an ANN model design includes
Results of Model I
It has been found that the first ANN model could learn the relationships between the input and the output neurons successfully. So it would seem possible for this ANN model to estimate the output neuron values by using the input neuron values for the same date.
Model I had a learning rate of 0.1; a momentum rate of 0.4; and a threshold value of 80%. There were three hidden layers between the input and output layers. The first and the second hidden layer had 20 neurons each, the third 50 neurons.
Conclusion and future works
Models I and II could forecast the values of the output neurons consisting of Non-performing Loans/Total Loans, Capital/Assets, Profits/Assets and Equity/Assets ratios by using the 25 input neurons consisting of macroeconomic variables, the variables related to the external balance and financial system’s structure, and time with very small errors. Both ANN models appeared to have learned the relationships between their input and output data successfully. And then, the input and the output data
References (26)
- et al.
Neural network forecasting of quarterly accounting earnings
International Journal of Forecasting
(1996) - et al.
Forecasting exchange rates using general regression neural networks
Computers and Operations Research
(2000) - et al.
Forecasting stock indices: A comparison of classification and level estimation models
International Journal of Forecasting
(2000) Neural network forecasting of Canadian GDP growth
International Journal of Forecasting
(2001)- et al.
Using neural networks to forecast the systematic risk of stocks
European Journal of Operational Research
(1996) Probabilistic neural networks in bankruptcy prediction
Journal of Business Research
(1999)- et al.
Neural network forecasting of the British Pound/US Dollar exchange rate
Omega
(1998) - et al.
Bankruptcy prediction of financially stressed firms: An examination of the predictive accuracy of artificial neural networks
International Journal of Intelligent Systems in Accounting, Finance and Management
(2001) Bankruptcy prediction for credit risk using neural networks: A survey and new results
IEEE Transactions on Neural Networks
(2001)- et al.
Application of neural networks to an emerging financial market: Forecasting and trading the taiwan stock index
Computers and Operations Research
(2003)
Semiparametric ARX neural-network models with an application to forecasting inflation
IEEE Transactions on Neural Networks
A comparison of linear regression and neural network methods for predicting excess returns on large stocks
Annals of Operations Research
A comparison of non-linear methods for predicting earnings surprises and returns
IEEE Transactions on Neural Networks
Cited by (58)
Does machine learning help private sectors to alarm crises? Evidence from China's currency market
2023, Physica A: Statistical Mechanics and its ApplicationsFinancial crises, bank efficiency and survival: Theory, literature and emerging market evidence
2021, International Review of Economics and FinanceA novel approach for panel data: An ensemble of weighted functional margin SVM models
2021, Information SciencesAn integrated early warning system for stock market turbulence
2020, Expert Systems with ApplicationsCitation Excerpt :Due to the non-linear nature of financial data, machine-learning algorithms are also recognized tools in the general field of stock market prediction (Krollner, Vanstone, & Finnie, 2010), e.g. stock price prediction (Dash & Dash, 2016a; Patel, Shah, Thakkar, & Kotecha, 2015b; Schöneburg, 1990), stock trend prediction (Dash & Dash, 2016c; Nelson, Pereira, & de Oliveira, 2017; Patel, Shah, Thakkar, & Kotecha, 2015a) and volatility prediction (Dash & Dash, 2016b; Oliveira, Cortez, & Areal, 2017). In the literature of EWS, artificial neural networks (Celik & Karatepe, 2007; Kim et al., 2004a; Kim, Oh, Sohn, & Hwang, 2004b; Oh et al., 2006; Sevim et al., 2014; Yu et al., 2010), fuzzy inference (Fang, 2012; Giovanis, 2012; Lin et al., 2008; Nan et al., 2012) and support vector machines (SVM) (Ahn, Oh, Kim, & Kim, 2011; Hu & Pang, 2008; Hui & Wang, 2006) have been proven accurate models for financial crisis prediction. Despite the promising results demonstrated by those studies in terms of accuracy, few investigates the test-set early warning power of the model, that is the duration of forewarned period before the crisis onset.
Measuring systematic risk with neural network factor model
2020, Physica A: Statistical Mechanics and its ApplicationsCitation Excerpt :One is to study the exposure of individual asset prices to systematic risk, for example, successful estimation or prediction of the beta for the CAPM (Wittkemper and Steiner [17] and Yuan and Lee [18]). The other is to reflect on the risk itself such as accurate predictions of financial crises (Celik and Karatepe [19] and Dabrowski et al. [20]). However, it seems that these methods can be improved by removing ad hoc properties or feature engineering from them.
Systemic banking crisis early warning systems using dynamic Bayesian networks
2016, Expert Systems with ApplicationsCitation Excerpt :Various expert systems have been proposed as EWSs. These include expert systems based on self-organising maps (Jagric, Bojnec, & Jagric, 2015), neural network models (Celik & Karatepe, 2007; Iturriaga & Sanz, 2015; Nag & Mitra, 1999; Sevim, Oztekin, Bali, Gumus, & Guresen, 2014), regression trees (Manasse & Roubini, 2009), binary classification trees (Duttagupta & Cashin, 2008), hybrid models (Lin, Khan, Chang, & Wang, 2008; Lin, 2009), grey rational analysis (Lin & Wu, 2011), support vector machines (Ahn, Oh, Kim, & Kim, 2011; Feki, Ishak, & Feki, 2012), random forests (Alessi & Detken, 2014), and the multiple-indicator-multiple-cause method (Rose & Spiegel, 2012). Comparisons between such methods have been conducted by Boyacioglu, Kara, and Baykan (2009).