The ANN method has been applied by means of multilayered feedforward neural networks (MLFNs) by using different macroeconomic variables such as the exchange rate of USD/TRY, gold prices, and the Borsa Istanbul (BIST) 100 index based on monthly data over the period of January 2000 and September 2014 for Turkey. Vector autoregressive (VAR) method has also been applied with the same variables for the same period of time. In this study, different from other studies conducted up to the present, ENCOG machine learning framework has been used along with JAVA programming language in order to constitute the ANN. The training of network has been done by resilient propagation method. The ex post and ex ante estimates obtained by the ANN method have been compared with the results obtained by the econometric forecasting method of VAR. Strikingly, our findings based on the ANN method reveal that there is a possibility of financial distress or a financial crisis in Turkey starting from October 2017. The results which were obtained with the method of VAR also support the results of ANN method. Additionally, our results indicate that the ANN approach has more superior prediction performance than the VAR method.
From the Great Depression in 1929 until the recent financial crisis of 2008, the world economy has experienced several financial crises that had grievous impacts on both industrialized and developing countries such as the Tequila crisis in 1994 in Mexico, the 1984 savings and loans (S&L) crisis in the US [
As Berg and Pattilo [
With the emergence of the 2008 global financial crisis, researchers accelerated their empirical works on prediction of financial crises and developing early warning systems. Due to the systemic risk and the contagious effect of the recent crisis, economists and policymakers realized how financial shocks can cause devastating consequences on the real economy. Small investors suffered a significant decline in the value of their stocks, which had serial negative impacts on such economic variables as household consumption spending, business investment spending, total demand, and employment rate [
If the crisis had been predicted, economists and policymakers would have provided an opportunity to reduce the negative impacts of the crisis or even totally avoid them [
The major purpose of this study is to predict stock market crash and financial distress before financial crisis by using different macroeconomic variables such as the exchange rate of USD/TRY (USD), gold prices (GOLDPRICE), and the Borsa Istanbul 100 index (BIST). Our specific objectives in this study can be listed as follows: To identify (if any) the causal relationship between macroeconomic variables such as gold prices, the exchange rate of USD/TRY, and the BIST 100 index. To predict a potential stock market crash or a financial crisis by using different variables in the future in order to minimize their costs to the economy.
This paper provides a highly expandable approach to neural networks and examines the connection between classical econometric techniques. An extensive review of the literature indicates that ANNs are generally more suitable than linear models for forecasting macroeconomics and financial variables. Although this method has several weak aspects, ANNs exhibit excellent results in many fields. In contrast to many researchers in the field, who tend to adopt an all-or-nothing approach to this issue, we argue that neural networks should be considered as the most powerful processor as compared to classical econometric methods. The full potential of neural networks can be exploited by using them in conjunction with linear regression models. Hence, as Gonzalez [
This paper is composed of four sections including introduction. The second section examines the main empirical literature which employs either the ANN method or the VAR method to test the relationships between different variables to predict financial distress or stock market crashes before the financial crises. The third section explains the empirical methodology used in the study and presents a detailed explanation of the empirical findings with their interpretations. The fourth section is the conclusion part.
In the last few decades, ANNs have been frequently used to solve complicated problems and to test the relationship between different variables. Empirical researches clearly indicate that they can be used as efficient tools to analyze critical issues in business decisions. In this context, ANNs have many advantages which may help to solve problems in finance as well as in other fields with poorly defined models. As Tu and Oztemel [
Many researchers prefer to utilize ANN to conduct their studies in different fields. However, only a limited number of researchers prefer to use ANN in their empirical researches to identify financial distress, to develop early warning systems, or to predict financial crisis. In one of these studies, Nag and Mitra [
Franck and Schmied [
Ozkan-Gunay and Ozkan [
Fioramanti [
Ban and Mazibas [
Several other researchers choose to employ VAR model to determine the relationships between different variables which might be used to predict stock market crashes or financial crises. As Shinkai and Kohsaka [
Dhakal et al. [
Sengonul et al. [
Bayraci et al. [
Koc and Akgul [
Zhou and Zhang [
VAR model is an easy to use model for the analysis of multivariate time series. It has been proven to be especially useful for describing the dynamic behavior of economic and financial time series and for forecasting. With VAR models, it is possible to approximate the actual process by arbitrarily choosing lagged variables. Thereby, one can form economic variables into a time series model without an explicit theoretical idea of dynamic relations. The easiest multivariate time series model is the bivariate VAR model with two dependent variables,
VAR analysis assumes that all the stochastic processes in the autoregressive system are variance-covariance stationary. If this condition is not met, little confidence can be placed in the VAR regression results. Since most if not all of the data here are highly autocorrelated and tend to drift over time, some form of detrending is obviously required before the models can be estimated. Our choice was influenced in part by the recent work of Dickey and Fuller [
Results for stable and unstable situations based on simulations for the chosen VAR(
Lag length | ||||||
---|---|---|---|---|---|---|
Information criterion | 0 | 1 | 2 | 3 | 4 | 5 |
Frequency distribution of estimated VAR orders, |
||||||
HJC, stable VAR | 3.25 | 4.0 | 97.5 | 4.5 | 2.0 | 0.3 |
HJC, unstable VAR | 0.1 | 3.3 | 95.3 | 4.3 | 3.1 | 0.2 |
HJC signifies the Hatemi-J information criterion presented by (
However, some situations exist when SBC has the highest rate of picking the right lag order compared to HCQ and there are also situations when HCQ outperforms SBC. Nevertheless, compared to Monte Carlo simulation studies, the true model is not known in empirical research. Thus, when these two information criteria are selected according to the two different lag orders, it is quite difficult to determine which criterion one should rely on. Some researchers suggest linking these two criteria to obtain the performance of the above information criterion based on Monte Carlo simulation experiments for choosing the optimal lag order in stable and unstable VAR models [
Table
Augmented dickey fuller (ADF) stationary test results.
Variables | Level value | First difference | ||||
---|---|---|---|---|---|---|
None | Intercept | Trend and intercept | None | Intercept | Trend and intercept | |
USD | 1.0078( |
−2.0151( |
−2.7693( |
−9.22638( |
−9.3790( |
−9.357046( |
0.9172 ( |
0.2802 ( |
0.2107 ( |
0.000 ( |
0.000 ( |
0.000 ( |
|
|
||||||
BIST | 1.5502( |
0.0584( |
−3.0669( |
−11.1903( |
|
−11.3634( |
0.9702 ( |
0.9616 ( |
0.1176 ( |
0.000 ( |
0.000 ( |
0.000 ( |
|
|
||||||
GP | 1.3376( |
0.3920( |
−2.0456( |
−9.9681( |
−10.2438( |
−10.2221( |
0.9543 | 0.9066 ( |
0.5719 | 0.000 | 0.000 | 0.000 |
“
The impulse response analysis quantifies the reaction of every single variable on an exogenous shock to our model. The impulse response function of VAR is to analyze dynamic effects of the system when the model receives the impulse. Impulse response analysis based on vector autoregressions (VARs) has an important place in modern empirical macroeconomics (for reviews of this literature, see Christiano et al. [
The results of impulse response analysis with VAR model.
The right column shows the impulse responses of other variables to the gold prices. When the impulse is GOLDPRICE, the response of USD is firstly positive. There is the highest positive effect in the second month, and the lowest negative effect is in the sixth month. The response of BIST is a smooth fluctuation, and more varieties in terms of fluctuations exist starting from the second month. In other words, a standard deviation shock to the GOLDPRICE causes gold prices suddenly to create a negative effect and this effect remains until the third month. After the turning point, the value fluctuates around the line zero. One standard deviation shock to USD causes USD to peak about 1-2 months. Then, it begins to decrease eventually leading to a substantial drop in USD after about 5-6 months. A standard deviation shock to GOLDPRICE increases the USD, and the effect becomes statistically significant 2 months after the shock. USD decreases to its previous value after about 6 months.
Variance decomposition analysis tries to show how a change of a variable is affected from other variables. In this study, we will explain the variance of the variable until the 10th period in variance decomposition here. According to the results of the variance decomposition of GOLDPRICE in Table
The results of variance decomposition.
Period | S.E. | DGOLDPRICE | DBIST | DUSD |
---|---|---|---|---|
Variance decomposition of DGOLDPRICE | ||||
1 | 2.326149 | 100.0000 | 0.000000 | 0.000000 |
2 | 2.418729 | 99.52846 | 0.337312 | 0.134227 |
3 | 2.426568 | 99.31653 | 0.455137 | 0.228337 |
4 | 2.431537 | 99.20369 | 0.481523 | 0.314782 |
5 | 2.431664 | 99.20029 | 0.481729 | 0.317978 |
6 | 2.431714 | 99.19671 | 0.483779 | 0.319516 |
7 | 2.431728 | 99.19564 | 0.484172 | 0.320188 |
8 | 2.431729 | 99.19562 | 0.484173 | 0.320207 |
9 | 2.431729 | 99.19561 | 0.484181 | 0.320211 |
10 | 2.431729 | 99.19560 | 0.484183 | 0.320213 |
|
||||
Variance decomposition of DBIST | ||||
1 | 3907.536 | 14.89173 | 85.10827 | 0.000000 |
2 | 3958.518 | 14.65356 | 84.73305 | 0.613395 |
3 | 4007.277 | 16.04923 | 83.27992 | 0.670842 |
4 | 4015.353 | 16.37279 | 82.94525 | 0.681959 |
5 | 4015.574 | 16.37697 | 82.93850 | 0.684534 |
6 | 4015.730 | 16.37801 | 82.93413 | 0.687867 |
7 | 4015.753 | 16.37849 | 82.93319 | 0.688321 |
8 | 4015.755 | 16.37852 | 82.93316 | 0.688321 |
9 | 4015.756 | 16.37852 | 82.93315 | 0.688337 |
10 | 4015.756 | 16.37852 | 82.93314 | 0.688340 |
|
||||
Variance decomposition of DUSD | ||||
1 | 0.055513 | 15.78019 | 14.56730 | 69.65252 |
2 | 0.059682 | 17.94952 | 15.26003 | 66.79045 |
3 | 0.059814 | 18.17990 | 15.27787 | 66.54223 |
4 | 0.060009 | 18.59504 | 15.27692 | 66.12804 |
5 | 0.060021 | 18.59939 | 15.29406 | 66.10655 |
6 | 0.060026 | 18.60796 | 15.29353 | 66.09851 |
7 | 0.060027 | 18.61097 | 15.29291 | 66.09612 |
8 | 0.060027 | 18.61101 | 15.29296 | 66.09603 |
9 | 0.060027 | 18.61101 | 15.29298 | 66.09601 |
10 | 0.060027 | 18.61102 | 15.29298 | 66.09600 |
|
||||
Cholesky ordering: DGOLDPRICE DBIST DUSD |
Over the last decade, a number of techniques have been developed, which allow estimation of general nonlinear models without specifying an exact functional form. Neural networks are one of these most popular techniques. White [
In this paper, the data of 177 months between January 2000 and September 2014, obtained from the Electronic Data Distribution System of the Central Bank of the Republic of Turkey (CBRT) and the website of the World Bank, were used. In the empirical analysis part, three different variables, which are USD/TRY exchange rates (USD), gold prices (GP), Borsa Istanbul (BIST) 100 index, were used. In the prediction of the next 36 months of crisis indicator data in Turkey, we preferred to use feed-forward, backpropagation, multilayered ANN. It consists of an input layer, two hidden layers, and an output layer as shown in Figure
A brief multilayered feedforward neural network (MLFN) architecture of the proposed methodology for ANN.
In this study, in terms of applied method, a feed-forward and multilayered ANN has been used which is composed of
As seen in Figure
Because of easiness in its utilization, time series, which take part in ENCOG framework, have been given preference in the Temporal Neural Dataset class of software that are constituted for the data sets. Additionally, we benefited from several strategies in order to assist coding in the framework such as “if certain number of iterations have been passed in the training of the network, restart to learning” or “if the error value does not progress a significant proportion, restart to learning.”
In this study, the normalization and the denormalization of data have been done with the help of built-in functions found in framework. While the network is being trained, the twelve-month lagged values of data were given as prologue and we demanded the forecasting of the 37th month. In the same manner, we tried to estimate results starting from the 37th month of 176 months that are in the data set. For each month to be predicted, ANN has been trained iteratively with its previous three-year data. For the 178th and the following months, again its previous three-year data has been given to the ANN as training set, and the missing months have been filled by the prediction values obtained in previous iterations. In other words, different networks have been used in our analysis in the calculation of each prediction value as compared to other studies which use a single network.
Figure
Comparison of real and predicted values with ANN.
Turkish economy also experienced the major crises of 1994 and 2001 until 2008. During both crises, large amounts of capital entered and left Turkey, leading to instability in the exchange rate regimes. In the aftermath of the 1994 crisis, the Turkish economy contracted by 6% and the Turkish Lira was devalued by more than 50% against the US Dollar. The second crisis erupted in the midst of an exchange rate based stabilization program. In February of 2001, the fixed exchange rate system was collapsed and Turkey began to implement the floating exchange rate system. Instability in the financial markets and lack of confidence together with optimism in expectations all created speculative attacks. As a result of both of these crises, the trust to the Turkish Lira was shaken, the reserves of the Central Bank of Turkey decreased significantly, and inflation rate skyrocketed. The period of 2002 and 2007 was a stability period thanks to structural reforms applied in Turkish economy and the financial system and the positive course of events in the global economy. However, this stability period has come to an end with the 2008 financial crisis.
The 2008 financial crisis, the first global financial crisis of the world, has had a deep negative effect initially on developed countries and later on emerging countries. Turkey is one of these developing countries which have been affected negatively from this crisis. As it can be seen from Figure
Figure
The future values predicted with ANN (2014–2017).
Figure
The future values predicted with VAR model.
The current financial crisis, the dynamism of financial markets, and the globalization process have accelerated the obsolescence of financial crisis prediction models and emphasized the need to reformulate these models. Neural networks arise as a powerful tool to enhance modeling flexibility and dynamism and to identify the most outstanding properties to predict financial crisis originated in some financial variables such as gold prices, stock exchanges, and exchange rates.
In this study, the causal relationship between macroeconomic variables such as gold prices, the exchange rate of USD/TRY, and the BIST 100 index has been investigated. Additionally by using these variables, we aimed to predict potential stock market crashes or financial crises in order to minimize their costs to the economy. We used two different methods of Artificial Neural Networks (ANNs) and the Vector Autoregressive (VAR) method to conduct this study. According to the findings of the ANN method, there is a possibility of significant fluctuations for the USD/TRY exchange rate and the BIST 100 index between 2015 and 2017 except for gold prices. Strikingly, our findings based on the ANN method reveal that there is a possibility of a financial crisis starting from October 2017.
Our results that we obtained with the VAR method is different from our results that we obtained with the ANN method. According to the results of VAR method, all of these three variables will fluctuate significantly between 2015 and 2017. However, very sharp fluctuations can be observed in the fourth quarter of 2015 and the third quarters of 2016 and 2017 for all of the variables. Based on our findings with the VAR method, we interpret that there is a possibility of a potential crisis in one of these dates.
Our study indicates that predictions of VAR method do not correspond to the ANN results. In this regard, it can be said that ANN method has a better estimation capability and we obtain more accurate results as compared to the VAR method. We come to this conclusion by analyzing the real values obtained in the past for both the methods of VAR and the ANN. Some of the major literatures, which support our perspective, include Nag and Mitra [
The literature on building economic models and early warning systems for the prediction of financial crisis is promising but still new. The main difficulty in studying macroeconomic problems is how to model people’s expectations when the economic environment is volatile mainly in developing countries. Small changes in expectations can often lead to large changes in people’s behavior and, thus, in the behavior of macroeconomic variables such as GOLDPRICE, foreign exchange rates, and stock market exchanges.
As a suggestion for further research, researchers can obtain more accurate results and make more precise predictions by using more macroeconomic variables. This study can also be extended to a larger time period covering the major crises experienced in 1990s to see the prediction capabilities of different methods such as VAR, ANN, or other alternative methods. By extending this study with more variables and more countries covering a larger time interval, a potential global crisis can even be predicted prior to its occurence. Consequently, long-term forecasting will contribute positively to the global economy as a first step of providing stability which is needed in the international business and economic environment.
The authors declare no conflict of interests.
All authors have contributed in this paper. All authors have read and approved the final paper.
The authors gratefully acknowledge the constructive comments offered by the anonymous referee(s), which help to improve the quality of the paper significantly.