Mathematical Modelling and Numerical Simulation

with Applications, 2024, 4(5), 187–206

https://dergipark.org.tr/en/pub/mmnsa
ISSN Online: 2791-8564 / Open Access
https://doi.org/10.53391/mmnsa.1523426

R E S E A R C H PA P E R

Credit default swaps prediction by using an FTS-ANN

model
Öznur Öztunç Kaymak ID 1,*,‡ , Yiğit Kaymak ID 2,‡ , Çağatay Mirgen ID 2,‡ ,

Süleyman Emir ID 3,‡ and Halis Can Koyuncuoğlu ID 4,‡

1 Department of Information Technology, Izmir Democracy University, Karabağlar, 35140 İzmir,
Türkiye, 2 Department of Business Management, National Defense University, 10145 Balıkesir, Türkiye,
3 Institute of Social Sciences, Mehmet Akif Ersoy University, 15200 Burdur, Türkiye, 4 Department of

Engineering Sciences, Izmir Katip Celebi University, Cigli, 35620 İzmir, Türkiye
* Corresponding Author
‡ [email protected] (Öznur Öztunç Kaymak); [email protected] (Yiğit Kaymak);
[email protected] (Çağatay Mirgen); emir_ suleyman_ [email protected] (Süleyman Emir);
[email protected] (Halis Can Koyuncuoğlu)

Abstract
Credit Default Swap (CDS) is a derivative instrument that serves as insurance against the credit risk
of countries or firms. Especially, since the 2008 global crisis, it has received much attention as a risk
indicator in financial markets. Given the role played by CDS prices in determining the creditworthiness
of banks, corporations or countries, even in predicting financial crisis, it is clear that there has been a
need for models that can produce results close to real values due to the nonlinear and chaotic nature of
CDS prices in fragile economies. In this study, Türkiye is analyzed as a fragile economy with a high
CDS premium. To do this, the artificial neural network (ANN) is combined with the fuzzy time series
(FTS) in order to construct a novel model called FTS-ANN. Based on this novel model, the predicted
results are evaluated using different well-known statistical techniques. It is found that the epoch and
regression R values of the proposed model are 8 and 0.99554. This shows that our model outperforms
other models. Finally, the expected contribution of our model is that this model, which gives very
good results for a fragile economy like Türkiye, can be adapted to the CDS values of other countries.
Keywords: CDS values; fuzzy time series; artificial neural network; Levenberg–Marquardt backpropa-
gation
AMS 2020 Classification: 91B24; 91B54; 91B99

1 Introduction
Credit default swap (CDS) is the most widely used credit derivative instrument that provides
protection against the underlying asset in case of its default [1]. CDS has been one of the most

➤ Received: 27.07.2024 ➤ Revised: 23.12.2024 ➤ Accepted: 27.12.2024 ➤ Published: 31.12.2024

187
 188 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

discussed derivatives since the global credit crunch 2008 [2]. CDS is known as an over-the-
counter credit protection instrument reducing the credit risk of the lender. It might also be a
measure of creditworthiness of underlying assets [3]. The CDS market has begun to grow over
the last years and has taken place in the financial literature [4]. CDS was firstly contracted by the
investment bank J.P. Morgan in 1994 in order to transfer credit exposure from its balance sheet
[5]. Nevertheless , its trading value was relatively low until the early 2000s [6]. Especially, in the
wake of 2008 financial crisis, CDSs have attracted greater attention in the credit market [7]. The
crisis, notably the bankruptcy of Lehman Brothers, made counter-party risk significant for market
participants. So, CDCs transactions increased significantly [8].
A CDS has two sides both protection buyer and protection seller. They can be a corporation or a
sovereign entering into a contract. It is a kind of insurance policy as well. Within the framework
of the contract, the protection borrower pays a premium until the borrower gets insolvent or the
contract matures. This premium acts as compensation for the protection seller bearing the risk of
a credit default [9]. Accordingly, CDS is a contract through which one party pays a premium at
a specified frequency and swaps the capital in case the other party defaults [10]. Further, CDS
spreads act as warning signals about the financial stability of credit institutions and the economy
as a whole [11]. CDS is said to be the most important liquid product in the credit derivatives
market which can be contracted on various underlying assets issued by a company or government
[12, 13]. On the other hand, banks use CDS to both examine their risk premium and hedge their
credit risk as well [14].
CDS value is accepted as one of the significant indicators that points out an approaching financial
crises. No doubt, financial and economic variables have profound effects on CDS spreads [15].
Also, global or regional risk factors highly affect emerging markets’ CDS [16]. These effects peaked
particularly in the 2008 financial crisis [15]. Given that CDS markets appear to be more liquid than
those of the referenced government bonds, analyzing CDS prices may be more useful to determine
sovereign credit risks as well as to measure systemic risk [17]. Moreover, CDS spreads perform
better than other ratings in determining firm-specific default probability [18]. In comparison with
other credit risk measures such as rating and bond yields, CDS spreads seem more instrumental
in credit risk assessments of rating agencies [19].
The structure of this paper is as follows: We give some relevant literature in Section 2. Then, we
introduce the Fuzzy Time Series (FTS) and the general structure of the Artificial Neural Networks
(ANNs) utilized to forecast the CDS values in Section 3. After that, we give Section 4 in which the
proposed models are introduced and the data used is explained. Finally, for analysis, the outputs
of the proposed and other models are discussed with different statistical parameters in Section 5.
Section 6 presents the conclusion of this study.
2 Literature review
In this section, we summarize a literature review of various models used for the estimation of
CDS values. With the help of ARIMAX (Integrated Autoregressive Moving Average) modeling,
in five European countries with high government debt, [20] finds that news from newspapers
contributes to the risk profile of the country. ANFIS model accurately predicts the daily prices
of Greece’s CDSs [21]. Similarly, using the ANFIS (Adaptive Neural Fuzzy Inference System)
model, a study, that gets Türkiye’s daily CDS data between 2015-2020, finds values close to real
values. According to estimation results of this model, exchange rate, stock price and interest rate
are observed to be the most effective inputs to estimate CDS values [22]. In a study that uses
the compound nerve networks such as ANFIS, NNARX, AdaBoost and Support Vector Machine
(SVM) regression to predict the price of CDS contracts of A-rated North American and European
companies, the prediction results show that the average predictive power of NNARX is better
 Öztunç Kaymak et al. | 189

than the other algorithms [23].

Besides, [24] shows that the multiple estimation models composed of a wide range of explanatory
variables are good at predicting five-year CDS spreads. [25] compares the predictability of CDS
values between China and the US through the permutation entropy approach and G(M) test. It
deduces that CDS of the USA is more predictable than that of China. In another study, the results
reveal that machine-learning algorithms perform pretty better than financial models in terms of
their ability to predict CDS prices of companies within the same risk class [26]. A model operating
on four stochastic factors such as hazard rate, exchange rate, local interest rate and foreign interest
rate explains the observed inconsistencies in CDS spreads traded in domestic and foreign markets
[27]. Further, taking China’s 5-year daily sovereign CDS as sample data, [28] uses RVM (Relevance
Vector Machine) and ARIMA (Auto Regressive Integrated Moving Average) models for predicting
China’s sovereign CDS. The empirical results show that the proposed model outperforms other
benchmarking models for predicting China’s sovereign CDS. A different model, which takes into
account the characteristics of CDS such as historical CDS price, news and financial leverage for
improving present GAN-based model, predicts more effectively CDS’s future trends than other
machine learning algorithms and traditional regression GAN (Generative Adversarial Networks)
[29]. The study of [30] reveals that, out of SVM (Support Vector Machines), GMDH (Group
Method of Data Handling), LSTM (Long Short Term Memory) and MSA (Markov Switching
Autoregression), MSA model displays better forecasting performance for CDS of 513 leading
companies in the USA before and during the Covid-19 outbreak. Also, in [31], SVM models are
used to forecast CDS price changes for some companies. This study proves that the combined
model including new input variables outperforms than the models based on historical CDS
price changes. In [32], a nonparametric machine-learning model and two traditional parametric
models are used to determine the predictive performance for the daily CDS of various maturities.
Particularly, ANN (Artificial Neural Networks) is better at predicting than the other parametric
and non parametric models. On the other hand, the literature also contains a two-stage deep
learning-based methodology for time series data clustering in [33]. In order to predict CDS indices,
Mao et al. [34] used the Merton-LSTM model, which integrated the Merton model with the LSTM
model. Within this study, The North American High Yield Index (CDX.NA.HY) and the North
American Investment Grade Index (CDX.NA.IG) are predicted with the Merton-LSTM model and
these prediction results are compared with machine learning models LSTM, GRU, MLP, SVM and
a typical stochastic series model. The comparison results show that the root mean square error
(RMSE) values of the Merton-LSTM model are the lowest compared to other models.

In addition, [35] produces an artificial intelligence-supported simulation model to estimate the

BIST 100 stock price index and 5-year credit default swap (CDS) premiums. Here, the short-term
and long-term relationships between variables are examined using nonlinear econometric models
and exponential smooth transition autoregressive vector error correction model with a data set for
1211 observations from Türkiye in the period of 21.08.2015 - 17.08.2020. Based on this data set, a
multi-layer artificial neural network (ANN) model is created. Using the Levenberg-Marquardt
algorithm as the training algorithm in the network model, six input variables, which are thought
to be common parameters affecting the target values, are defined in the input layer. BIST 100 and
CDSs are also estimated in the output layer. The ANN model performs well in predicting BIST
100 with an average deviation of 0.04 and CDSs with an average error of -0.163. These error rates
show that the ANN is successful in predicting BIST 100 and CDS with ideal accuracy. Further, for
the telecommunication sector, the prediction model is created to make strategic CDS investment
decisions by using machine learning in [36].
 190 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

3 Fundamentals of artificial neural networks and fuzzy time series

Fuzzy time series (FTS), which are largely used in forecasting stock price, interest rate, crude
oil price etc., are among the oldest known models in the literature (see et al. [37–39]). There are
numerous approaches to predict time series data’s next lag in an efficient manner [40] and [41].
This model whose foundations were laid in 1965 by Zadeh [42] has been so far in use for different
methods and approaches in several fields. The basic goal of this theory is to create schemas for
uncertainties resulting from inherent ambiguity for the purpose of resolving various problems
[43]. In many disciplines, including health [44], economics [45], meteorology [46] etc., time series
analysis and forecasting techniques have remained popular for nearly 50 years.
As for ANNs, they are computer models that draw inspiration from a human brain. Currently,
the power of ANNs to find optimal solutions for complex and nonlinear data has been of great
attraction [47]. This model is superior to cope with limitations in linear machine learning methods.
Especially, A Multilayer Perceptron (MLP) architecture is widely used for prediction applications
for its capable of learning and representing highly complex non-linear relationships between input
and output data. Under this structure, the input layer is made up of an output layer and of a set of
hidden layers. Besides, each neuron in one layer is connected to neurons in the next layer [48]. This
architecture with two layers is shown by [49]. Overall, the combination of non-linear mapping,
adaptability, parallel processing, feature extraction, robustness, scalability, and versatility makes
artificial neural networks a powerful and widely applicable tool in machine learning and artificial
intelligence [50] and [51]. Also, various artificial intelligence systems, such as support vector
regressions (SVR), Gaussian Poisson regressions (GRP), algorithmic models and artificial neural
network topologies, are also successful in solving the problem of accurate predictability of Bitcoin
data [52].
In addition to this kind of artificial neural network based studies, two forecasting models - a basic
neural network model and a hybrid model based on time series - have been proposed to investigate
the predictive capabilities of neural networks for time series data. In this study, the Taiwan Stock
Exchange Capital Weighted Equity Index (TAIEX) from 1991 to 2003 is chosen as the target for
forecasting [53]. The hybrid model performed better than both the basic model and a traditional
fuzzy time series model according to the empirical results. In addition, a neural network based
fuzzy time series model numerically outperforms the first order model [54]. Multivariate model
[55] and a basic ANN model [53] in forecasting the Taiwan stock market index [56]. From the
literature, the proposed hybrid models will remain popular as advanced neural networks have a
better modeling capability for fuzzy time series methods [57].
In the literature, the models that can produce highly accurate results appropriate for the nonlinear
and chaotic nature of CDS prices are limited. In this context, the main objective of this paper is to
propose a different hybrid model by combining the traditional underpinnings of fuzzy time series
and the powerful learning capacity of artificial neural networks. To do this, a fragile economy
with very high CDS values,Türkiye is chosen as the dataset for testing this model. Thus, this paper
aims to overcome the inadequacy of traditional methods that lack to predict volatility of CDS val-
ues. More importantly, the novelty and contribution of this model to the literature are given below:

• The proposed model is not an artificial neural network model used within the fuzzy time series.
Instead, it has a structure in which traditional fuzzy time series outputs are used as inputs for
the artificial neural network.
• Since the proposed model is based on the successive application of two models (FTS and ANN),
the Levenberg-Marquardt algorithm has been applied in Matlab to increase the speed of our
model.
 Öztunç Kaymak et al. | 191

4 The proposed FTS-ANN model

FTS-ANN model design involves several systematic steps. As shown in Figure 2, there are two
fundamental steps: (1) implementation of the classical FTS models (2) application of the ANN
with the Levenberg-Marquardt algorithm. In this study, we use a multilayer perceptron (MLP) to
determine the number of hidden layers and the neurons in each layer. We also use the Levenberg-
Marquardt algorithm in this study. This particular approach is called a Gauss-Newton technique
modification in [58]. As clearly seen from this figure, the outputs of the data trained with FTS
models constitute the inputs of the ANN model.
We get daily data on one year of the CDS price trend from January 2020 to September 2022 for
Türkiye. This data is obtained from the Thomson Reuters Eikon™ Datastream and includes price
indicators such as open, high, low and closing prices and volume. In this study, we only consider
the closing prices as seen in Figure 1.
Data preprocessing for this model is composed of several key steps to improve model accuracy
and handle uncertain data. These steps include handling missing values and cleaning the data.
The proposed dataset of 350 daily CDS values is analyzed with other datasets using Matlab
software. The next part explains the micro-ten steps of the proposed model. In this model, the

Figure 1. Historical CDS Data of Türkiye for 1.5 Years Data source: from 01 Jan 2020 to 08 Sep 2022

first nine of them are the application of the FTS. The last step includes the application of the
ANN (see Figure 2). In these steps, the stages of defining the universe of discourse and finding
the linguistic values for the training data set, creating the fluctuation-type matrix, determining
the standardized weights, finding the spreading center and defuzzification of the CDS values
are explained in detail without mentioning the well-known basic definitions in the literature. (If
desired, see [37, 38, 54] and the references therein for more details.)
Step 1: In this phase, U = [ Dmin − D1 , Dmax + D2 ] , the universe of discourse, is determined for
observations where D1 and D2 are positive integers. In view of approximately two years of data,
we determine the U value for Türkiye as seen in Table 1. Later, U is splitted into eleven linguistic
intervals for this data set. This value is shown in the same table.
Step 2: In this step, the linguistic values L1 , L2 , ..., Lk are defined for each observation for the
discourse universe U. In order to obtain the CDS risk values of Türkiye from 01 Jan 2020 to 06
192 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

Figure 2. Detailed steps of ANN-FTS model

Table 1. Detect of the universe of discourse and interval

Country The Universe of Discourse (U) The Number of Intervals
Türkiye [234, 653] 11

Table 2. Identification of linguistic values of CDS values for Türkiye

Time CDS Values ($) Linguistic Values
t = 9.03.2020 413.73 L6
t = 10.03.2020 408.94 L6
t = 11.03.2020 433 L6
t = 12.03.2020 496.99 L8
t = 13.03.2020 466.88 L7
t = 16.03.2020 518.4 L8
t = 17.03.2020 510.72 L8
t = 18.03.2020 586.73 L10
t = 19.03.2020 575.75 L10
 Öztunç Kaymak et al. | 193

May 2021, we get the eleven linguistic values as follows:

L1 = (very low value), L2 = (low value), L3 = (little low value),

L4 = (close to normal value), L5 = (normal value),
L6 = (slightly higher than normal value),
L7 = (higher than normal value), L8 = (high value),
L9 = (very high value) and L10 = ( pretty high value),
L11 = (the highest value).

Step 3: Now, Fuzzy Logic Relationship ( FLR) is created for each linguistic values. After obtaining
eleven linguistic values in the previous steps, we show linguistic values corresponding to these
CDS values in the Table 2 for the sample data set. For this set, we can easily create this FLRs table
as seen Table 3.

Table 3. Fuzzy logic relationship

FLRs
...
L6 → L6 , L6 → L6 , L6 → L8
L8 → L7 , L7 → L8 , L8 → L8
L8 → L10 , L10 → L10
...

Step 4: In this step, we create a fluctuation-type matrix to train these FLRs’ fuzzy relations. Then,
we compute FLR weight according to its order of occurrence to generate this matrix. Later, a
fluctuation-type matrix that contains the FLR values for Türkiye in Table 4 is formed.

Table 4. The CDS values’ fluctuation-type matrix for Türkiye, the years 2020 and 2021
P
L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11 wk
L1 27 3 30
L2 3 43 6 1 53
L3 6 21 2 29
L4 2 13 2 17
L5 3 101 8 112
L6 8 57 4 1 70
L7 1 4 32 5 1 43
L8 6 56 7 2 71
L9 7 43 4 54
L10 2 4 13 6 25
L11 6 7 13

Step 5: Here, to each FLR group, we identify the fluctuation weight assignment. By the adding up
each row in the Table 4, the fluctuation-type matrix of all FLR is found.

Step 6: The standardized weights matrix, Wn (t), for each Lk , k = 1, 2, .., 11 is obtained by using
Table 4. The second column in Table 4 shows this matrix.
 194 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

Table 5. The standardized weight matrix for all linguistic values

′ ′ ′ ′ ′
Linguistic values The standardized weight matrix: [W1 , W2 , ..., W9 , W10 , W11 ] The spread center of linguistic values (Ld
L1 [27/30,3/30,0,0,0,0,0,0,0] 250.7923333
L2 [3/53,43/53,6/53,0,0,0,1/53,0,0] 297.8629057
L3 [0,6/29,21/29,0,2/29,0,0,0,0,0,0] 326.0610345
L4 [0,0,2/17,13/17,2/17,0,0,0,0,0,0] 359.0941176
L5 [0,0,0,3/112,101/112,8/112,0,0,0,0,0] 383.8162281
L6 [0,0,0,0,8/70,57/70,4/70,1/70,0,0,0] 426.3549275
L7 [0,0,0,0,1/43,4/43,32/43,5/43,1/43,0,0] 469.5970233
L8 [0,0,0,0,0,0,6/71,56/71,7/71,2/71,0] 511.647662
L9 [0,0,0,0,0,0,0,7/54,43/54,4/54,0] 549.8172182
L10 [0,0,0,0,0,0,0,2/25,4/25,13/25,6/25] 590.703600
L11 [0,0,0,0,0,0,0,0,0,6/13,7/13] 627.6669231

Step 7: In this step, the spread center for each observation for all Türkiyes’ CDS values is found.
(please see the third column in Table 5)
Step 8: Here, in light of the values at the previous steps, we obtain the forecasting values using
the standardized weight matrix in Table 4 and fluctuation-type matrix in Table 5. We use these
values while applying the models in step 9. With the use of the fuzzified data contained in Table 6,
Forecast(t + 1) is calculated. This data will now serve as the input for the subsequent equations.

Table 6. Identification of linguistic values of CDS for Türkiye

Time CDS Values ($) Linguistic Values Defuzzify: (Ld f (t)) XWn (t))) Forecast (t + 1)
t=1 413.73 L6 425.182 386.192
t=2 408.94 L6 425.182 425.182
t=3 433 L6 425.182 425.182
t=4 496.99 L8 504.745 425.182
t=5 466.88 L7 470.334 504.745
t=6 518.4 L8 504.745 470.334
t=7 510.72 L8 504.745 504.745
t=8 586.73 L10 586.708 504.745
t=9 575.75 L10 586.708 586.708

Step 9: Now, in order to generate the forecasting results, we use the following equations given in
the relevant references:

GoldenForecast(t) = P(tlast ) + τ ( Forecast(t) − P(tlast − 1))

−τ ( Forecast(t − 1) − P(tlast − 2)), (1)
√
where ±τ = ± 1−2 5
= ±0.61803 in [59].

FibonacciForecast(t + 1) = P(t) + α( Forecast(t + 1) − P(t)) − β( Forecast(t) − P(t − 1)), (2)

where α and β are in the range -1 and 1 but not including 0 in [60].
In addition to these equations, the greatest integer function-based equation in [61] is used to
obtain very close values to realistic forecasts of crude oil prices. The outputs for the final step are
produced by substituting the estimation values in the fifth column of Table 6.
Step 10: As mentioned above, 350 daily CDS values from January 1, 2020 to May 6, 2021 are used
as the training data set. In this step, these sets are subjected to training with the using of proposed
 Öztunç Kaymak et al. | 195

models in [61], in [59], in [60] through the steps mentioned above. Then, we subject this forecast
data to the ANN model again. Briefly, this stage covers the evaluation process of retraining all
data sets trained with the fuzzy time series method with ANN. The detailed diagram covering
the whole process of the proposed model is given in Figure 3. Here, we name the ANN model
being structured by using the Fibonacci equation (2) as "Fibonacci based-ANN Model". Similarly,
we form a model using the equation in (1) called as Golden Ratio based-ANN Model. Finally, the
equation in [61] is also used in our novel model called FTS-ANN.

x1
W1
Model 1 Bias

x2
W2

Summarization F
Output
Model 2
Activation
Function
Daily CDS xn
Wn
Values
Inputs

Model 3

ARTIFICIAL NEURAL NETWORK

FUZZY TIME SERIES

Figure 3. The structure of the Golden Based-ANN, Fibonacci Based-ANN and FTS-ANN models

5 Results anddiscussions
In this section, out of 350 daily CDS values, 244 of them are allocated as training, 53 for validation
and 53 for testing. The results were fitted by using Matlab’s Levenberg-Marquardt algorithm
with hidden neurons and linear output neurons in a 10-layer feedback artificial neural network
model. In order to better examine the performance of our proposed model, only the ANN model
is applied and the results are discussed as well.
The experimental values are compared with the predictions made by the proposed FTS-ANN
model and others. It is clear from Figure 4, Figure 5, Figure 6 and Figure 7 that Golden based-ANN
Model and the proposed model have a very small number whose epoch is 8. On the other hand, it
is known that the convergence of R-value to 1 means that there is a close relationship between the
forecasted and the targeted data in [62]. So, for all these data, we compute the R values of 0.73277,
0.71636, 0.66917 and 0.99554 as shown in these figures. When the R values of these models are
evaluated, we see that FTS-ANN has a superior modeling capacity compared to the others. Thus,
the four graphs at the end of the paper for training, verifying, testing, and combining data show
that the data fit the model well through effective training.
The histogram representing the first simple ANN model shows a wide error spread (-323.6,338.6)
and higher errors in the validation and test sets, indicating reduced generalization. In the second
histogram, the Fibonacci-based ANN model has a lower error range (-318.9,305.9) and a less
balanced distribution than the first model. On the other hand, the golden-based ANN model
has the highest error range (-329,337.7) and a more balanced distribution. Finally, the model
representing the proposed FTS-ANN model is clearly superior to all models in minimizing errors
and maintaining consistency. In fact, the model has an error range of (-66.4, 48.68), with errors
concentrated close to zero and a balanced distribution between training, validation and test sets.
This asserts that the model has the best performance as well as having strong accuracy and
 196 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

generalization. (see Figure 8)

Output ~= 0.51*Target + 2.6e+02 Output ~= 0.58*Target + 2.2e+02

Output ~= 0.5*Target + 2.5e+02 Output ~= 0.5*Target + 2.7e+02
Training: R=0.71079 Validation: R=0.79887

Data Data
800 800
Fit Fit
700 Y=T Y=T

600 600
500

400 400
400 600 800 400 600 800
Target Target
Test: R=0.78207 All: R=0.73277

Data Data
800 800
Fit Fit
700 Y=T Y=T

600 600

500

400 400

400 600 800 400 600 800

Target Target

Best Validation Performance is 14690.2114 at epoch 9

106
Train
Validation
Test
Best
105
Mean Squared Error (mse)

104

103

102
0 5 10 15
15 Epochs

Figure 4. For only ANN Model, regression of training, testing, validation, and combination with all sets, the
R-value
 Öztunç Kaymak et al. | 197

Output ~= 0.38*Target + 3.5e+02 Output ~= 0.51*Target + 2.7e+02

Output ~= 0.49*Target + 2.8e+02 Output ~= 0.55*Target + 2.5e+02

Training: R=0.71315 Validation: R=0.82365

Data 800 Data

800
Fit Fit
700 Y=T 700 Y=T

600 600

500 500

400 400

400 600 800 400 600 800

Target Target
Test: R=0.62945 All: R=0.71636

Data Data
800 Fit 800 Fit
Y=T Y=T

600 600

400 400

400 600 800 400 600 800

Target Target

Best Validation Performance is 10071.8065 at epoch 22

106
Train
Validation
Test
Best
105
Mean Squared Error (mse)

104

103

102
0 5 10 15 20 25
28 Epochs

Figure 5. For Fibonacci based-ANN Model, regression of training, testing, validation, and combination with all
sets, the R-value
 198 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

Output ~= 0.42*Target + 3.4e+02 Output ~= 0.43*Target + 3.1e+02

Output ~= 0.43*Target + 3.2e+02 Output ~= 0.38*Target + 3.3e+02

Training: R=0.65732 Validation: R=0.65118

Data 800 Data

800
Fit Fit
700
Y=T 700 Y=T

600 600

500 500

400 400
400 600 800 400 600 800
Target Target
Test: R=0.72743 All: R=0.66917

Data Data
800 Fit 800 Fit
Y=T Y=T

600 600

400 400
400 600 800 400 600 800
Target Target

Best Validation Performance is 16453.2331 at epoch 8

106
Train
Validation
Test
Best
Mean Squared Error (mse)

105

104

103
0 2 4 6 8 10 12 14
14 Epochs

Figure 6. For Golden based-ANN Model, regression of training, testing, validation, and combination with all
sets, the R-value
 Öztunç Kaymak et al. | 199

Training: R=0.99556 Validation: R=0.99594

Output ~= 0.99*Target + 1.1

Output ~= 1*Target + 3.6
Data Data
800 800
Fit Fit
Y=T Y=T
700

600 600

500

400 400
400 600 800 400 600 800
Target Target
Test: R=0.99612 All: R=0.99554

Output ~= 1*Target + -0.58

Output ~= 1*Target + -18

Data Data
800 800
Fit Fit
Y=T Y=T
700

600 600

500

400 400
400 600 800 400 600 800
Target Target

Best Validation Performance is 301.5655 at epoch 8

106
Train
Validation
Test
105 Best
Mean Squared Error (mse)

104

103

102

101
0 2 4 6 8 10 12 14
14 Epochs

Figure 7. For the proposed ANN-FTS Model, regression of training, testing, validation, and combination with all
sets, the R-value
200 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

(a) The Simple ANN Model

(b) The Fibonacci based-ANN Model

(c) The Golden based-ANN Model

(d) The Proposed ANN-FTS Model

Figure 8. The histogram diagrams for all the models

 Öztunç Kaymak et al. | 201

6 Conclusion
CDS is a crucial indicator reflecting economic stability and risk perception and has gained im-
portance since the 2008 global crisis. In this context, this paper aims to suggest a novel model
for predicting CDS values. To this end, a FTS-ANN model is structured using Türkiye’s CDS
data. In this model, the outputs obtained from the proposed models’ results are compared with
other models. These results display good forecasting accuracy for CDS values. Additionally, what
matters is that this model has a couple of important features that differentiate from it others. First
of all, although the previous studies focus on the combination of many models, no modeling
has been so far constructed with the using of fuzzy time series that is old but still very popular.
Secondly, Türkiye’s daily CDS values from January 2020 to September 2022, when CDS volatility
was quite high, are included in this model. On top of that, given their capability of learning
complex patterns, neural networks are incredibly practical to use. Further, time series forecasting
is another useful tool in the model to make decisions as well. At this point, our model, which is a
combination of these two models, is more instrumental in cope withing non-linear and chaotic
patterns of CDS values. So, this model turns out to be much more predictive even in highly volatile
periods, which enables market actors such as traders, lenders and investors etc. to make more
efficient decisions.
In conclusion, this work will have some implications for various economic actors and policy-
makers in practice. The model could serve as a means of early warning in reducing counterpart risk,
notably sovereign risk through its power of providing information and of predicting CDS premium.
Now that CDS premiums interact with other economic and financial variables such as exchange
rate, interest rate, stock prices, debt costs, expectations or other risk premium,s etc., the prediction
accuracy of the model can also undoubtedly reflect some “preliminary information” about these
variables. Put it differently, given that these variables react to sovereign CDS market, the model
helps economic actors indirectly forecast the direction of them in line with the fluctuations in CDS
values.
Most notably, since high uncertainty in the market is a serious issue negatively affecting almost
all macroeconomic variables, the indicators of economic uncertainty are of utmost importance
for policy makers and economic agents in decision-making process. In this context, CDS is one of
the basic indicators that reflects the financial conditions of the whole economy and the level of
volatility. Consequently, the future direction of CDS values and their volatility are worth being
monitored by many economic agents for several reasons. Though the proposed model is not a
complete financial crisis forecasting model, it may indirectly signal an approaching financial crisis
in the market accompanied by other macroeconomic and financial indicators. In this regard, a
continued increased trend in CDS values points that the systemic risk in the economy is on the
verge of increasing. For example, the expectations of extreme increases in CDS price can inform
lenders and borrowers of the future course of credit costs. For this reason, this model can be a
dynamic tool to guide potential debt relationships in the future.
Accurate forecasts are also a crucial tool for banks in liquidity and credit risk management. The
model remarkably contributes to planning of the trading strategies by its high degree of prediction
accuracy as well. Further, the more accurate predictions shape the risk and return expectations of
bond or stock investors including real investment decisions. No doubt, it adds remarkably value
to risk prediction studies in the financial sector.
This paper leaves a fertile area for future studies, which could integrate our model with other
machine learning models. In doing so, the robustness and predictive capabilities of the model will
be more enhanced in the days to come.
 202 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

Declarations
Use of AI tools
The authors declare that they have not used Artificial Intelligence (AI) tools in the creation of this
article.

Data availability statement

The data associated with the manuscript is taken from "Thomson Reuters Eikon™ Datastream".

Ethical approval
The authors declare that this research complies with ethical standards. This research does not
involve human participants or animals.

Consent for publication

Not applicable

Conflicts of interest
The authors declare that they have no conflict of interest.

Funding
No funding was received for this research.

Author’s contributions
Ö.Ö.K.: Conceptualization, Methodology, Software. Y.K.: Writing-Original draft preparation,
Investigation. Ç.M.: Data curation, Methodology. S.E.: Investigation, Validation. H.C.K.: Software,
Validation, Writing-Reviewing and Editing. All authors discussed the results and contributed to
the final manuscript.

Acknowledgements
Not applicable
References
[1] Garcia, J. and Goossens, S. The Art of Credit Derivatives: Demystifying the Black Swan. John
Wiley & Sons: New York, (2010).
[2] Tabassum and Yameen, M. Why do banks use credit default swaps (CDS)? A systematic
review. Journal of Economic Surveys, 38(1), 201-231, (2024). [CrossRef]
[3] Hui, C.H. and Fong, T.P.W. Price cointegration between sovereign CDS and currency option
markets in the financial crises of 2007–2013. International Review of Economics & Finance, 40,
174-190, (2015). [CrossRef]
[4] Galil, K., Shapir, O.M., Amiram, D. and Ben-Zion, U. The determinants of CDS spreads.
Journal of Banking & Finance, 41, 271-282, (2014). [CrossRef]
[5] Augustin, P., Subrahmanyam, M.G., Tang, D.Y. and Wang, S.Q. Credit default swaps: Past,
present, and future. Annual Review of Financial Economics, 8, 175-196, (2016). [CrossRef]
[6] Abrahams, C.R. and Zhang, M. Credit Risk Assessment : The New Lending System for Borrowers,
Lenders, and Investors. John Wiley Sons: New Jersey, (2009).
 Öztunç Kaymak et al. | 203

[7] Son, Y., Byun, H. and Lee, J. Nonparametric machine learning models for predicting the
credit default swaps: An empirical study. Expert Systems with Applications, 58, 210-220, (2016).
[CrossRef]
[8] Choudhry, M. An Introduction to Credit Derivatives. Butterworth-Heinemann: USA, (2013).
[9] Weistroffer, C., Speyer, B., Kaiser, S. and Mayer, T. Credit default swaps. Deutsche Bank
Research, 1-26, (2009).
[10] Mazzi, B. Treasury Finance and Development Banking,+ Website: A Guide to Credit, Debt, and Risk.
John Wiley & Sons: New York, (2013).
[11] Hu, N., Li, J. and Meyer-Cirkel, A. Completing the Market: Generating Shadow CDS Spreads by
Machine Learning. International Monetary Fund: USA, (2019).
[12] Chan-Lau, J.A. and Kim, Y.S. Equity prices, credit default swaps, and bond spreads in
emerging markets. International Monetary Fund, 1-30, (2004).
[13] Schofield, N.C. and Bowler, T. Trading the Fixed Income, Inflation and Credit Markets: A Relative
Value Guide. John Wiley & Sons: United Kingdom, (2011).
[14] Amato, J.D. Risk aversion and risk premia in the CDS market. BIS Quarterly Review, 55-68,
(2005).
[15] Doshi, H., Jacobs, K. and Zurita, V. Economic and financial determinants of credit risk
premiums in the sovereign CDS market. The Review of Asset Pricing Studies, 7(1), 43-80, (2017).
[CrossRef]
[16] Daehler, T.B., Aizenman, J. and Jinjarak, Y. Emerging markets sovereign CDS spreads during
COVID-19: Economics versus epidemiology news. Economic Modelling, 100, 105504, (2021).
[CrossRef]
[17] Stolbov, M. The causal linkages between sovereign CDS prices for the BRICS and major
European economies. Economics, 8(1), 20140026, (2014). [CrossRef]
[18] Hilscher, J. and Wilson, M. Credit ratings and credit risk: Is one measure enough?. Management
Science, 63(10), 3147-3529, (2017). [CrossRef]
[19] Das, S.R., Hanouna, P., and Sarin, A. Accounting-based versus market-based cross-sectional
models of CDS spreads. Journal of Banking & Finance, 33(4), 719-730, (2009). [CrossRef]
[20] Apergis, N. Forecasting Credit Default Swaps (CDSs) spreads with newswire messages:
Evidence from European countries under financial distress. Economics Letters, 136, 92-94,
(2015). [CrossRef]
[21] Atsalakis, G.S., Tsakalaki, K.I. and Zopounidis, C. Forecasting the Prices of Credit Default
Swaps of Greece by a Neuro-fuzzy Technique. Financial Engineering Laboratory, 1-14, (2012).
[22] Karabıyık, B.K. ANFIS metodu ile CDS primi tahminlemesi: Türkiye örneği. International
Journal of Management Economics and Business, 17(4), 1175-1197, (2021). [CrossRef]
[23] Beytollahi, A. and Zeinali, H. Comparing prediction power of artificial neural networks
compound models in predicting credit default swap prices through Black-Scholes-Merton
model. Interdisciplinary Journal of Management Studies, 13(1), 69-93, (2020). [CrossRef]
[24] Driesum, S. Forecasting Credit Default Swap Spreads. Ph.D. Thesis, Department of Econometric
Institute, The Erasmus University, (2016). [https://thesis.eur.nl/pub/33966/]
[25] Feng, Q., Hao, J., Sun, X. and Li, J. Predictability of sovereign CDS: permutation entropy
method. Procedia Computer Science, 199, 866-870, (2022). [CrossRef]
 204 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

[26] Gündüz, Y. and Uhrig-Homburg, M. Predicting credit default swap prices with financial and
pure data-driven approaches. Quantitative Finance, 11(12), 1709-1727, (2011). [CrossRef]
[27] Itkin, A., Shcherbakov, V. and Veygman, A. New model for pricing quanto credit default
swaps. International Journal of Theoretical and Applied Finance, 22(03), 1950003, (2019). [CrossRef]
[28] Li, J., Hao, J., Sun, X. and Feng, Q. Forecasting China’s sovereign CDS with a decomposition
reconstruction strategy. Applied Soft Computing, 105, 107291, (2021).[CrossRef]
[29] Lin, S.Y., Liu, D.R. and Huang, H.P. Credit default swap prediction based on generative
adversarial networks. Data Technologies and Applications, 56(5), 720-740, (2022). [CrossRef]
[30] Vukovic, D.B., Romanyuk, K., Ivashchenko, S. and Grigorieva, E.M. Are CDS spreads pre-
dictable during the Covid-19 pandemic? Forecasting based on SVM, GMDH, LSTM and
Markov switching autoregression. Expert Systems With Applications, 194, 116553, (2022). [Cross-
Ref]
[31] Zhang, T.Y. Predicting credit default swap (cds) returns with machine learning. University of
Florida Journal of Undergraduate Research, 20(1), 1-11, (2018). [CrossRef]
[32] Son, Y., Byun, H. and Lee, J. Nonparametric machine learning models for predicting the
credit default swaps: An empirical study. Expert Systems with Applications, 58, 210-220, (2016).
[CrossRef]
[33] Tavakoli, N., Siami-Namini, S., Adl Khanghah, M., Mirza Soltani, F. and Siami Namin, A.
An autoencoder-based deep learning approach for clustering time series data. SN Applied
Sciences, 2, 937, (2020). [CrossRef]
[34] Mao, W., Zhu, H., Wu, H., Lu, Y. and Wang, H. Forecasting and trading credit default swap
indices using a deep learning model integrating Merton and LSTMs. Expert Systems with
Applications, 213, 119012, (2023). [CrossRef]
[35] Koy, A. and Çolak, A.B. Predicting stock market index and credit default swap spreads using
artificial intelligence and determining non-linear relations. Archives of Advanced Engineering
Science, 00(00), 1-12, (2023). [CrossRef]
[36] Wu, C., Li, J., Xu, J. and Bouvry, P. Strategic predictions and explanations by machine learning:
The prediction model of credit default swaps for the telecommunication service sector. In
Proceedings, 2024 International Conference on Information Networking (ICOIN), pp. 268-273, Ho
Chi Minh City, Vietnam, (2024, January). [CrossRef]
[37] Song, Q. and Chissom, B.S. Forecasting enrollments with fuzzy time series—Part I. Fuzzy Sets
and Systems, 54(1), 1-9, (1993). [CrossRef]
[38] Song, Q. and Chissom, B.S. Forecasting enrollments with fuzzy time series—Part II. Fuzzy
Sets and Systems, 62(1), 1-8, (1994). [CrossRef]
[39] Karasu, S., Altan, A., Bekiros, S. and Ahmad, W. A new forecasting model with wrapper-
based feature selection approach using multi-objective optimization technique for chaotic
crude oil time series. Energy, 212, 118750, (2020). [CrossRef]
[40] Siami-Namini, S., Tavakoli, N. and Namin, A.S. A comparison of ARIMA and LSTM in
forecasting time series. In Proceedings, 17th IEEE International Conference on Machine Learning
and Applications (ICMLA), pp. 1394-1401, Orlando, USA, (2018, December).[CrossRef]
[41] Baz, J., Ferrero-Jaurrieta, M., Díaz, I., Montes, S., Beliakov, G. and Bustince, H. Probabilistic
study of Induced Ordered Linear Fusion Operators for time series forecasting. Information
Fusion, 103, 102093, (2024). [CrossRef]
 Öztunç Kaymak et al. | 205

[42] Zadeh, L.A. Fuzzy sets. Information and Control, 8(3), 338-353, (1965).
[43] Montes, I., Miranda, E. and Montes, S. Decision making with imprecise probabilities and
utilities by means of statistical preference and stochastic dominance. European Journal of
Operational Research, 234(1), 209-220, (2014). [CrossRef]
[44] Tricahya, S. and Rustam, Z. Forecasting the amount of pneumonia patients in Jakarta with
weighted high order fuzzy time series. In Proceedings, IOP Conference Series: Materials Science
and Engineering, pp. 052080, Malang, Indonesia, (2019, March). [CrossRef]
[45] Yu, H.K. Weighted fuzzy time series models for TAIEX forecasting. Physica A: Statistical
Mechanics and its Applications, 349(3-4), 609-624, (2005). [CrossRef]
[46] Chen, S.M. and Hwang, J.R. Temperature prediction using fuzzy time series. IEEE Transactions
on Systems, Man, and Cybernetics, Part B (Cybernetics), 30(2), 263-275, (2000). [CrossRef]
[47] Singh, S., Bansal, P., Hosen, M. and Bansal, S.K. Forecasting annual natural gas consumption
in USA: Application of machine learning techniques-ANN and SVM. Resources Policy, 80,
103159, (2023). [CrossRef]
[48] Attux, R.R.D.F., Duarte, L.T., Ferrari, R., Panazio, C.M., de Castro, L.N., Von Zuben, F.J.
and Romano, J.M.T. MLP-based equalization and pre-distortion using an artificial immune
network. In Proceedings, 2005 IEEE Workshop on Machine Learning for Signal Processing, pp.
177-182, Mystic, USA, (2005, September). [CrossRef]
[49] Mustaffa, Z. and Sulaiman, M.H. Stock price predictive analysis: An application of hybrid
barnacles mating optimizer with artificial neural network. International Journal of Cognitive
Computing in Engineering, 4, 109-117, (2023). [CrossRef]
[50] Pelegrina, G.D., Duarte, L.T. and Grabisch, M. A k-additive Choquet integral-based approach
to approximate the SHAP values for local interpretability in machine learning. Artificial
Intelligence, 325, 104014, (2023). [CrossRef]
[51] Yavuz, M. and Özdemir, N. A feed-forward neural network approach to Istanbul stock
exchange. Journal of Applied Computer Science & Mathematics, 12(26), 31-36, (2018). [CrossRef]
[52] Lahmiri, S. and Bekiros, S. Intelligent forecasting with machine learning trading systems in
chaotic intraday Bitcoin market. Chaos, Solitons & Fractals, 133, 109641, (2020). [CrossRef]
[53] Huarng, K. and Yu, T.H.K. The application of neural networks to forecast fuzzy time series.
Physica A: Statistical Mechanics and its Applications, 363(2), 481–491, (2006). [CrossRef]
[54] Chen, S.M. Forecasting enrollments based on fuzzy time series. Fuzzy Sets and Systems, 81(3),
311-319, (1996). [CrossRef]
[55] Huarng, K.H., Yu, T.H.K. and Hsu, Y.W. A multivariate heuristic model for fuzzy time series
forecasting. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 37(4),
836–846, (2007). [CrossRef]
[56] Yu, T.H.K. and Huarng, K.H. A neural network-based fuzzy time series model to improve
forecasting. Expert Systems with Applications, 37(4), 3366–3372, (2010). [CrossRef]
[57] Bas, E., Egrioglu, E. and Tunc, T. Multivariate picture fuzzy time series: New definitions
and a new forecasting method based on Pi-sigma artificial neural network. Computational
Economics, 61(1), 139–164, (2023). [CrossRef]
[58] Haykin, S. Neural Networks: A Comprehensive Foundation. Prentice Hall PTR: New Jersey,
(1998).
[59] Kaymak, Ö.Ö., Kaymak, Y. and Özdemir, N. Forecasting of the Istanbul stock exchange (ISE)
 206 | Mathematical Modelling and Numerical Simulation with Applications, 2024, Vol. 4, No. 5, 187–206

return with a golden ratio model in the epidemic of COVID-19. Applied and Computational
Mathematics, 20(1), 95-107, (2021).
[60] Chen, T.L., Cheng, C.H. and Teoh, H.J. Fuzzy time-series based on Fibonacci sequence for
stock price forecasting. Physica A: Statistical Mechanics and its Applications, 380, 377-390, (2007).
[CrossRef]
[61] Kaymak, Ö.Ö. and Kaymak, Y. Prediction of crude oil prices in COVID-19 outbreak using
real data. Chaos, Solitons & Fractals, 158, 111990, (2022). [CrossRef]
[62] Kannaiyan, M., Karthikeyan, G. and Thankachi Raghuvaran, J.G. Prediction of specific wear
rate for LM25/ZrO2 composites using Levenberg-Marquardt backpropagation algorithm.
Journal of Materials Research and Technology, 9(1), 530-538, (2020). [CrossRef]

Mathematical Modelling and Numerical Simulation with Applications (MMNSA)

(https://dergipark.org.tr/en/pub/mmnsa)

Copyright: © 2024 by the authors. This work is licensed under a Creative Commons Attribution
4.0 (CC BY) International License. The authors retain ownership of the copyright for their article,
but they allow anyone to download, reuse, reprint, modify, distribute, and/or copy articles in
MMNSA, so long as the original authors and source are credited. To see the complete license
contents, please visit (http://creativecommons.org/licenses/by/4.0/).

How to cite this article: Öztunç Kaymak, Ö., Kaymak, Y., Mirgen, Ç., Emir, S. and
Koyuncuoğlu, H.C. (2024). Credit default swaps prediction by using an FTS-ANN
model. Mathematical Modelling and Numerical Simulation with Applications, 4(5), 187-206.
https://doi.org/10.53391/mmnsa.1523426