The Impact of Macroeconomic Sustainability on Exchange Rate: Hybrid Machine-Learning Approach

Abstract

This paper constructed a robust methodology to investigate the impact of news regarding macroeconomic policies on exchange rate fluctuations, and to examined the applicability of qualitative information alongside historical data to predict exchange rates. To do so, hybrid machine learning algorithms comprised of natural language processing, fuzzy logic, and support vector regression have been constructed. This study emphasizes the significance of qualitative information on investors’ subjective consideration, the decision-making process, and causality on exchange rate volatility. To perceive the causality of expected and unexpected macroeconomic news on exchange rate fluctuations, news regarding the inflation rate, interest rate, unemployment rate, balance of trade, and credit ratings has been extracted from the web. Learning automata has been adopted to construct a unique lexicon for textual analysis. Subjective considerations of decision makers based on news have been evaluated by processing using the prospect theory and composing fuzzy antecedents for the fuzzy logic phase. The fuzzy logic method attained the correlation value between the macroeconomic news and the exchange rate. Finally, support vector regression predicted the exchange rate on a daily basis. The statistical test results indicated a strong correlation between recently published macroeconomic news on daily exchange rate fluctuations and their usability for predicting exchange rates in the short term, while emphasizing the significance of sustainable macroeconomic policies on exchange rate stability.

1. Introduction

The sustainable economy and economic policies of a country are highly dependent on exchange rates [1,2], due to its significant role in international trade, investment determination, risk management, and the balance of payments [3,4]. For this reason, understanding the reasoning behind the fluctuations of exchange rates plays a substantial role in taking the right steps beforehand and maintaining the sustainability of economic policies and economic growth by predicting future fluctuations. However, exchange rates are highly volatile and substantially information dependent, and they can easily be affected by any event-based information that has a potential to influence the investors’ sentiments [5]. The factors that influence the expectations of investors, as well as industrial characteristics, such as government officials’ statements on macroeconomic indicators and political events [6,7,8], have a substantial effect on originating the non-stationary, nonlinear, dynamic, chaotic, and noisy nature of exchange rates [9,10,11], which are the most challenging, yet most important problems concerning the analysis of exchange rates, leading to the development of strategies to maintain sustainable economic policies. The significance of the exchange rate in macroeconomic management is crucial [12], as fluctuating exchange rates may cause a domino effect on macroeconomic variables. Unstable exchange rates have an adverse impact on macroeconomic variables in the short term [13]. Therefore, volatile exchange rate will cause unstable economy, which may influence sentiments of investors, and have adverse impact on macroeconomic indicators [14].

Today, the significant effects of publicly available political and financial information on financial markets and economics are broadly accepted. Any information regarding economic policy, whether it reflects the truth or completely relies on manipulation, plays a substantial role in the perception of investors under vague conditions that promote a lack of trust. The doubtful perception of investors on policy sustainability has a potential risk of increasing the influence of manipulations on the exchange rate, thus harming the economic sustainability. On the other hand, integrity, fairness, and transparency of policymakers may strengthen the trust of investors in the sustainability of the policy, and lessen the impact of information that can be interpreted negatively. That is why, in addition to objective and quantitative variables such as historical price data, recently obtained subjective and qualitative information also play an essential role in economic sustainability.

Recent technological advancements have unlocked an unlimited potential have and provided data scientists with an opportunity to intensify their research by deeply analyzing the complex combinations of datasets through the use of powerful computational intelligent forecasting models established under the name of machine learning (ML). More recently, the ML techniques have begun superseding traditional time series prediction techniques [15] by the nature of their design to particularly cope with non-linearity, uncertainty, and randomness [7]. The ML algorithms, such as Support Vector Machines (SVM), Fuzzy Logic (FL), and the Artificial Neural Network (ANN), have been used for financial time series prediction [16,17,18,19,20] and exchange rate prediction [2,3,4,5]. Studies [3,4,5,16,17,18,19,20] that compared the prediction performance of ML algorithms with traditional econometric models, such as Vector Auto-Regression (VAR), Auto Regressive Integrated Moving Average (ARIMA), and Autoregressive Conditional Heteroskedasticity/Generalized Autoregressive Conditional Heteroskedasticity (ARCH/GARCH) models have claimed superiority among the ML algorithms, and it is reported that ML methods outperformed traditional econometric models in prediction by a huge margin. Furthermore, the literature suggests that the hybrid machine learning algorithms provide the ability to encode the necessary knowledge and achieve superior prediction outcomes well matched to real world scenarios. The rise of the machine learning algorithms brings a fresh perspective and enhances economic prediction by providing the capability to use subjective and qualitative information as data.

This article aims to contribute to the literature by evaluating the investors’ perception of sustainability measures on exchange rate by investigating the psychological impact of official macroeconomic statements and recently published news on investors. The study brings a novel approach to exchange rate prediction. A hybrid machine learning algorithm is generated by the authors to unveil the significance of macroeconomic news on exchange rate fluctuations and to predict the price of the exchange rates on a daily basis. Human cognitive perception may be defined as a fuzzy process, stated in imprecise terms and developed out of quantitative meaning, in general. In order to comprehend humans’ imprecise, deliberate, and uncertain thoughts that can only be expressed in linguistic terms, interpreting the non-quantitative, approximate, dispositional, and linguistic method of human reasoning is necessary. Fuzzy logic has especially been preferred in order to evaluate the fuzzy antecedents obtained through sentiment analysis, to filter the antecedents through the determined rules to mimic the human decision-making route, and to defuzzify the correlation between variables for regression analysis. The predictive accuracy of SVM, ANN, NN, Decision Tree (DT), Multiple Regression (MR), Random Forest (RF), and Ensemble Averaging (EA) have been compared throughout the existing literature, suggesting the superiority of SVM’s predictive accuracy in financial time series [5,16,17,19,20] forecasting. Thus, SVM has been adopted for regression analysis. Macroeconomic news has been used to predict macroeconomic fundamentals and the exchange rate in the existing literature [21,22,23]; however, this paper is a pioneering study on the use of hybrid machine learning algorithms, examining both qualitative and quantitative data to predict exchange rates. News regarding official government statements, expectations of the attitude on economic policies, and the publication of macroeconomic data plays a crucial role in informing the community about recent economic situations that a country is currently facing. The information provided by the news plays a significant role in the behavior of investors in the short term. In order to comprehend the correlation between the news and exchange rate volatility, recently published macroeconomic news is examined alongside the historical exchange rate and macroeconomic data.

Emerging economies are a perfect fit for the study of sustainability due to their exceptionally responsive nature to both domestic and foreign economic decisions, expectations, and events [24]. In recent years, the Turkish economy stands out as an obvious proof of this hypothesis. The Turkish lira depreciated by 152% against the US dollar, and hit record-breaking lows while becoming the worst-performing currency of the emerging markets between the dates 29 December 2017 and 1 November 2019. Furthermore, macroeconomic indicators conspicuously fluctuated significantly between the aforementioned date interval, which may clearly be observed through the lowest inflation rate, which was 8.55%, while the highest was 25.52%; the lowest unemployment rate was 9.60%, while the highest was 14.7%; and lowest interest rate was 8.0%, while highest was 24%. Meanwhile, the trade gap decreased down to USD −0.46 billion, and increased up to USD −9.21 billion. Moreover, the most significant rating agencies, namely Standard & Poor’s (S & P), Moody’s, and Fitch, downgrade Turkey’s rating from the highest band of speculative grade rating down to mid-range ratings. S&P downgraded Turkey’s credit rating from BB to B+, Moody’s downgraded Turkey’s credit rating from Ba1 to B1, and Fitch downgraded Turkey’s credit rating from BB+ to BB−. Hence, consecutive macroeconomic instability that occurred in a limited date interval, as mentioned above, provided a unique opportunity to study the instantaneous impact of the perception of the unsustainability of macroeconomic policies on exchange rate fluctuations.

2. Materials and Methods

2.1. Materials

As the causality between exchange rate and macroeconomic indicators has been broadly verified by the literature, the macroeconomic indicators, such as inflation rate, unemployment rate, interest rate, balance of trade, and credit rating, have been selected for examination in order to perceive the impact of unsustainable macroeconomic expectancy and existence on the exchange rate.

The instability of the aforementioned indicators during the selected date intervals has also been presented in the previous section of this paper. The literature suggests that the inflation rate is one of the fundamental macroeconomic indicators that has a vital role in monetary policy and a sustainable economy, as well as in guiding both occupants’ and enterprises’ financial planning by specifying solid signs of the purchasing power of residents [25]. The empirical evidence suggests that the exchange rate has a significant impact on macroeconomic stability and sustainability due to its direct influence on purchasing power, in addition to the supply and demand of imported and exported goods [26,27], especially on emerging economies that are highly import-dependent [28].

Additionally, the literature emphasizes the interconnection between the stability of the balance of trade and sustainable stability of the exchange rate. The impact and the causality of volatile exchange rates on the balance of trade in both the short and long run have been verified [29,30]; the substantial short-term causativeness between exchange rate instability and trade balance in Turkey [31] has also been found throughout the existing literature.

The interest rate is a critical macroeconomic variable that plays a vital role in monetary policy, aside from its effect on the unpremeditated instability of the exchange rate [12]. It may be necessary to take monetary action through central banks under a circumstance of unpremeditated volatility of the exchange rate to stabilize the fluctuating exchange rate. The existing literature indicates the causality of uncertain exchange rate fluctuations on interest rate, especially in Turkey [32], and verifies that the interest rate is a significant variable to effectively predict further exchange rates [33,34].

Moreover, the impact of uncertainty-based exchange rate instability on the unemployment rate has also been investigated for both emerging and developed economies throughout the literature. Relevant papers have emphasized an adverse correlation between an instable exchange rate and the unemployment rate in both the short and long run [35].

Lastly, a sovereign credit rating that has been graded by agencies indicates the capability, likelihood, and readiness of the governments to fulfill their financial obligations [36]. The stability and sustainability of the macroeconomic policies and the exchange rate play a crucial role in evaluating the creditworthiness of an obligor. One study in particular focused on the impact of sovereign credit ratings on the USD/TRY and EUR/TRY exchange rate instability, indicating a depreciation of the TRY in the case of downgrading the credit rating, while an appreciation of the currency has been observed in the case of upgrading the credit rating [37]. Furthermore, several studies have investigated the initial behavior of the exchange rate after rating announcements and observed an immediate, substantial, and asymmetric response of the exchange rate to sovereign credit rating announcements [38,39].

2.2. Methods

For sentiment analysis, HTTP REST API has extracted from daily news articles concerning official macroeconomic statements, announcements, and published news articles regarding the expectations and opinions of economists from leading news providers’ websites. Over eighty news articles have been investigated, and the textual content was transformed into numeric sentiment inputs.

To perceive the degree of responsiveness of the exchange rate to recently published macroeconomic news, the daily price, daily change in percentage, opening price, highest price, lowest price, and the difference between the highest and lowest prices of the USD/TRY have been investigated for each day. In particular, the opening price, close price, highest price, lowest price, and the difference between these prices play a crucial role in overcoming any possible accuracy errors in the ML algorithm caused by the coincidence of news content on opposite polarities that may take place on the same day. The accommodation of news on opposite polarities on the same day may neutralize the significant effect of vital news, causing a dramatic error in further calculations of the algorithm. To cope with this issue, the philosophy behind the candlestick chart has been adopted. In this case, even if neutralization occurred in any day during the selected date interval, the algorithm will detect it by observing the volatility of the exchange rate price throughout the day, and will weight related news accordingly to match the volatility and the price by considering previous/further news on both subjects. As financial markets are closed during weekends, any news published during the weekends was evaluated with the opening price of the next weekday in order to be able to observe the reaction of the market to that news. The same principle was applied to other scheduled national holidays as well. The abovementioned exchange rate data have been collected over the period between 29 December 2017 and 1 November 2019, which is equivalent to 4032 variables.

2.2.1. Learning Automata

The lexicon based sentiment analysis technique uses domain specific or domain independent libraries to classify the words individually as either positive or negative according to their semantic orientation value [40]. In order to define the semantic orientation of the text, the polarity value that defines the intensity measure for each word should be obtained. The text’s polarity may simply be calculated by averaging the polarity value of the constituent words [41].

To build an effective lexicon that analyzes sentiments efficiently in stochastic environments, the learning automata method is used. Learning automata is an adaptive decision-making algorithm that aims to figure out the optimal action in randomly fluctuating/unknown environments from a finite set of actions [42]. Learning automata has been widely accepted by disciplines that deal with highly uncertain human behaviors, such as psychology and biology, due to its classification abilities. In order to create a unique lexicon with a learning automata algorithm for sentiment analysis, the annotated text should be fed into system, and word a polarity function should be defined for each word. Polarity function defines the degree of positivity/negativity of each word in the formed lexicon. A higher polarity value indicates a highly positive word, or vice versa. A probability reinforcement scheme is applied for each word recorded, which can be formulated as;

$$p_i\left(n+1\right)=p_i\left(n\right)-\left(1-\beta \left(n\right)\right)g_i\left(p\left(n\right)\right)+\beta \left(n\right)h_i\left(p\left(n\right)\right) if a\left(n\right)\ne a_i$$

(1)

$$p_i\left(n+1\right)=p_i\left(n\right)+\left(1-\beta \left(n\right)\right){\displaystyle \sum }_{j\ne i}^{}{g}_i\left(p\left(n\right)\right)-\beta \left(n\right){\displaystyle \sum }_{j\ne i}^{}{h}_j\left(p\left(n\right)\right) if a\left(n\right)=a_i$$

where, $g_i$ and $h_i$ denote positive and negative annotations, respectively, for the nominated action $a_i$ with received environmental response $\beta \left(n\right)$ at cycle $n$, normalized in the interval (0,1). Favorable response can be achieved as the value of $\beta \left(n\right)$ decreases [43]. $p\left(i\right)$ denotes the score of each existing word in the training set, and the word polarity vector $p\left(x\right)$ can be identified as $p\left(i\right)=\frac{1}{m}$ for each word $i, 1\le i\le m$, where m denotes the total number of words appeared in the training set. Based on the aforementioned word polarity function, the polarity algorithm that is applied for training can be formulated as;

$$p\left(x\right)=p\left(x\right)+R/L·p\left(x\right) \mathrm{for} \mathrm{every} \mathrm{word} \mathrm{found} \mathrm{in} \mathrm{the} \mathrm{sentence}$$

$$p\left(x\right)=p\left(x\right)+S/(m-w)·p\left(x\right) \mathrm{for} \mathrm{every} \mathrm{other} \mathrm{word} \mathrm{not} \mathrm{in} \mathrm{the} \mathrm{sentence},$$

where R, L, S, and w denote the score of the examined sentence, coverage speed of the automaton, the produced overall increase of the sentence based on the polarity vector, and number of the words found in the particular sentence, respectively. As parameter S states all the value received by the words appearing in the training set, it can be calculated by $S=S_1+S_2+\cdots +S_w$. As the polarity vector is only capable of expressing in positive values, it needs to be normalized to achieve sentiment analysis in the rage of (−1, 1). The normalized polarity vector can be formulated as

$$P_N\left(x\right)=\frac{p\left(x\right)-min}{max-min}$$

(2)

where the smallest value (min) is the most negative recorded polarity value, while the largest value (max) is the most positive recorded polarity value in the lexicon [43].

2.2.2. Prospect Theory

Kahneman and Tversky proposed the prospect theory in order to qualify an individual’s actual decision behavior in risky situations [44]. The prospect theory addresses the inadequate rationalization of the expected utility theory on the certainty effect, isolation effect, and reflection effect [45], and aims to eliminate them by describing decision makers’ subjective outlook by weighting and valuing the functions and developed a novel decision model under uncertainty [46]. The prospect theory may be formulated as;

$$V={\displaystyle \sum }_{i=1}^k(w(p_i)v\left(\mathsf{\Delta }{\chi }_i\right))$$

(3)

where V and w(p) denote prospect value and probability weight function, respectively, and v(∆χ) denotes the value derived from the subjective feelings of the decision maker, where ${\chi }_0$ is a certain reference point and $\mathsf{\Delta }{\chi }_i$ is the deviation. Gains and losses may be determined by the $\mathsf{\Delta }{\chi }_i$. While positive $\mathsf{\Delta }{\chi }_i$ value indicates gains, negative value indicates losses. The power function of the value is represented as [47],

$$v\left(\chi \right)=\{\begin{array}{c}{\chi }^{\alpha }, \chi \ge 0\\ -\theta {(-\chi )}^{\beta }, \chi <0\end{array}$$

(4)

where α and β denote a concave-convex degree of gains and losses, respectively, regarding to the power function value of χ. Higher α and β values indicate that the decision maker tends to risk, which is presented by θ. θ indicates higher possibility for losses rather than gains, while θ > 1 indicates loss aversion [48].

The probability of weight function denotes the probability of the event outcome p regarding the subjective judgment of the decision maker. The weight function represents the corresponding weight on the probability, and it cannot be assessed as a linear function of the probability due to its responsive nature to the change in the probability. The probability weight function may be formulized as

$$w^{+}\left(p\right)=\frac{p^{\gamma }}{{(p^{\gamma }+{(1-p)}^{\gamma })}^{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\gamma $}\right.}} w^{-}\left(p\right)=\frac{p^{\delta }}{{(p^{\delta }+{(1-p)}^{\delta })}^{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\delta $}\right.}}$$

(5)

where $w^{+}\left(p\right)$ and $w^{-}\left(p\right)$ denote the nonlinear weight function of the gains and losses, respectively, while p, γ, and δ denote the probability of the event outcome, risk gain attitude coefficient, and risk loss attitude coefficient, respectively [48]. The degree of curvature and the elevation of the weighing function can be measured as γ > 0 and δ > 0, respectively [49]. As δ increases, the risk aversion for losses increases, while the risk aversion for gains decreases [50]. As the value of p denotes the probability of the event outcome, w (p) > p indicates that the decision maker overvalued the event, while w (p) < p indicates the undervaluation of the event [48].

2.2.3. Fuzzy Logic

Fuzzy logic was introduced by Lofti Zadeh [51] as an instrument that transforms linguistically articulated information into practicable mathematical algorithms [52]. The model aims to mimic the human decision-making route by utilizing approximate reasoning logic [51]. IF-THEN rules shape the linguistic form of knowledge that can practically entirely express the structured human knowledge. The appropriate distribution of membership functions, designated fuzzy rules, and fuzzy set operation plays a crucial role in fuzzy logic’s aptitude to deal with uncertainty and vagueness [53].

The fuzziness can be characterized by the membership function. The membership function establishes the correlation between the quantitative variable values and qualitative linguistic variables that takes place using IF-THEN rules. Imprecise and subjective statements define classes of membership. The membership function associated with fuzzy sets may be described by μ(x) → (0, 1), where degree of membership is denoted as μ, and the membership function for x in fuzzy set A is denoted as ${\mu }_A\left(x\right)$, while the universal set is denoted as X. The degree of membership of element x in set A is represented by the value between 0 and 1. The transition from x to ${\mu }_A\left(x\right)$ is known as fuzzification. This study uses sigmoidal and trapezoid membership functions. The sigmoidal membership function has right and left shoulders. The right shoulder may be formulated as $\mu =\frac{1}{1+e^{-\beta \left(x-\alpha \right)}}$, and the left shoulder may be formulated as $\mu =\frac{1}{1+e^{\beta \left(x-\alpha \right)}}$, where β controls the steepness of the sigmoid, and α denotes the crossover point [54]. The trapezoid membership function may be formulated as

$$\mu \left(x\right)=\{\begin{array}{c}\begin{array}{c}0 if x<a\\ \frac{x-a}{m-a} if x\in \left[a,m\right]\end{array}\\ \begin{array}{c}1 if x\in \left[m,n\right]\\ \frac{b-x}{b-n} if x\in \left[n,b\right]\end{array}\\ 0 if x>b\end{array}$$

(6)

where the upper and lower bounds and the center of triangle are denoted as a, b, and c, respectively, while the coordinates of tolerance are denoted as m and n. Unlike in crisp sets, it is possible to be a member of a fuzzy set up to a degree. To do so, fuzzy set operation is necessary; the connections between different sets are expressed as fuzzy relations. While the presence or absence of association, interaction, or interconnectedness between the elements of two or more sets represent the crisp relation [55], various degrees or strengths of relations between elements form the fuzzy relations [52]. As relation of the element to itself is a set, all operations may be applied to it without any adjustments required. The set of all ordered pairs $X\times Y=\{\left(x, y\right) | x\in X and y\in Y\}$ has been defined as the Cartesian products, or product sets by [52]. The Cartesian product is the simplification of n-tuples (relation matrix), characterized by a function ${\mu }_R : X_1\times \cdots \times X_m\to \left[0, 1\right]$, where ${\mu }_R$, $X_i$ and $X_1\times \cdots \times X_m$ denote the membership function of a multidimensional fuzzy set, universes of discourse, and the product space, respectively [54]. The Cartesian product X × Y can be defined as R (X,Y), or simply R, which represents the fuzzy relation between two sets [52]. The composition can be obtained by combining different product spaces of fuzzy relations with each other. The MAX-MIN composition of two fuzzy relations $R_1\left(x,y\right),\left(x,y\right)\in X\times Y$ and $R_2\left(y,z\right),\left(y,z\right)\in Y\times Z$ with a membership function of ${\mu }_{R_1\circ R_2}$ may be represented as

$$R_1\circ R_2\left(x,z\right)=\left\{\left[\left(x,z\right),\begin{array}{c}max\\ y\end{array}\{min\left\{{\mu }_{R_1}\left(x,y\right),{\mu }_{R_2}\left(y,z\right)\right\}\}\right]| x\in X, y\in Y, z\in Z\right\}$$

(7)

The interpretation of the resulting relation matrix R in a linguistic way may be obtained by IF-THEN rules. IF-THEN rules are widely used to model structured human knowledge [52]. The clarification of the sets of rule-based values in the input vector to an output vector is called fuzzy inference. The fuzzy inference methods most widely used to calculate the relational matrix are the Mamdani and Larsen methods. The Mamdani method ${\mu }_{A\Rightarrow B}\left(x,y\right)=MIN\left({\mu }_A\left(x\right),{\mu }_B\left(y\right)\right)$ uses the minimum operator as a fuzzy implication and the max-min operator for the composition, while the Larsen method ${\mu }_{A\Rightarrow B}\left(x,y\right)={\mu }_A\left(x\right){\mu }_B\left(y\right)$ uses the product operation as a fuzzy implication and the max-product operator for the composition. Input variables ${\mu }_A\left(x\right) \mathrm{and} {\mu }_B\left(y\right)$ vary between 0 and 1. The main objective of this method is to transform qualitative knowledge into the form of IF-THEN rules by using the membership function for linguistic variables as low, medium, and high.

$$R_1:\mathrm{IF} x_1=low \mathrm{AND} x_2=medium \mathrm{THEN} y=high$$

(8)

$${\mu }_R\left(x_1, x_2,y\right)=MIN\left({\mu }_L\left(x_1\right),{\mu }_M\left(x_2\right){\mu }_H\left(y\right)\right)$$

MIN operator should be used while calculating rule $R_1$ as the association between the small and medium antecedents will be made by the AND operator.

$$H=MIN\left({\mu }_L\left(x_1^{\prime }\right), {\mu }_M\left(x_2^{\prime }\right)\right)$$

(9)

where H denotes fuzzy inference and $x^{\prime }$ denotes fuzzified crisp value or transformation into a membership vector ${\mu }_{L^{\prime }}$. At this point, the defuzzification method should be applied in order to transfer the fuzzy value of y into a crisp value. Defuzzification can be formulated as;

$$y^{\prime }=\frac{{\sum }_{i=1}^N{y}_i{H}_i}{{\sum }_{i=1}^N{H}_i}$$

(10)

2.2.4. Support Vector Regression

Support vector machines transform the nonlinear input into a higher dimensional feature space when the input cannot be separated linearly. The transformation constructs a linear model in the new space, which signifies the nonlinear decision boundary of the original space. Subsequently, the linear optimal separating hyperplane is featured in new space [20]. The algorithm that composes the optimum location for the decision boundaries is called the support vector machine.

Support vector machines have also been improved to overcome prediction problems [56]. This method is called support vector regression. The main objective of support vector regression is to minimize the prediction error and the risk of over-fitting by using the ε-intensive zone [57]. Support vector regression can be formulated as $f\left(x\right)=w^T\varphi \left(x\right)+b$, where $f \left(x\right)$, w, ${\varphi }_i\left(x\right)$, and b denotes the function of support vectors, weight parameters that determine the hyperplane, mapping function, and bias term, respectively. As the main objective of support vector regression is to minimize the risk R and the width and flatness of ε-tube ${\Vert w\Vert }^2$ simultaneously, the risk function R should be minimized.

$$R_{w,\xi ,{\xi }^{\ast }}=\left[\frac{1}{2}{\Vert w\Vert }^2+C\left({\displaystyle \sum }_{i=1}^l{\xi }_i+{\displaystyle \sum }_{i=1}^l{\xi }_i^{\ast }\right)\right]$$

(11)

$$Subject to\{\begin{array}{c}{y}_i-w^T{x}_i-b\le \epsilon +{\xi }_i\\ w^T{x}_i+b-y_i\le \epsilon +{\xi }_i^{\ast }\\ \begin{array}{c}{\xi }_i\ge 0\\ {\xi }_i^{\ast }\ge 0\end{array}\end{array}$$

where $\xi$ and ${\xi }^{\ast }$ denote slack variables that measure above and below the ε-tube, l denotes the number of training data, and C is the unit of regularization that controls the trade-off between the empirical risk and misclassification. The user will select parameter C with a trade-off between an approximation error and the weights vector norm ‖w‖ [52]. Constructing a Lagrangian function can solve the quadratic optimization problem. Lagrange multipliers ${\alpha }_i \mathrm{and} {\alpha }_i^{\ast }$ are the unique minimum and maximum values of the function equivalent to $\xi \mathrm{and} {\xi }^{\ast }$ that are above and below the ε-tube. The Lagrange function can be formulated as

$$L_p\left(w,\xi ,{\xi }^{\ast },{\alpha }_i,{\alpha }_i^{\ast },{\beta }_i,{\beta }_i^{\ast }\right):$$

(12)

$$L=\frac{1}{2}{w}^{T}w+C\left({\displaystyle \sum }_{i=1}^l{\xi }_i+{\displaystyle \sum }_{i=1}^l{\xi }_i^{\ast }\right)-{\displaystyle \sum }_{i=1}^l{\alpha }_i^{\ast }\left[y_i-w^T{x}_i-b+\epsilon +{\xi }_i^{\ast }\right]-{\displaystyle \sum }_{i=1}^l{\alpha }_i\left[w^T{x}_i+b-y_i+\epsilon +{\xi }_i^{\ast }\right]-{\displaystyle \sum }_{i=1}^l\left({\beta }_i^{\ast }{\xi }_i^{\ast }+{\beta }_i{\xi }_i\right)$$

Lagrangian variables $L_p\left(w,\xi ,{\xi }^{\ast },{\alpha }_i,{\alpha }_i^{\ast },{\beta }_i,{\beta }_i^{\ast }\right)$ have to be minimized in regards to the primal variables $w,b,\xi , \mathrm{and} {\xi }^{\ast }$ and maximized regarding the non-negative Lagrange multipliers ${\alpha }_i,{\alpha }_i^{\ast },{\beta }_i,\mathrm{and} {\beta }_i^{\ast }.$ Lagrange multipliers $L_d\left(\alpha , {\alpha }^{\ast }\right)$ can be maximized by applying the Karush–Kuhn–Tucker (KKT) conditions for regression, which may be formulized as;

$$L_d\left(\alpha ,{\alpha }^{\ast }\right)=-\epsilon {\displaystyle \sum }_{i=1}^l\left({\alpha }_i^{\ast }+{\alpha }_i\right)+{\displaystyle \sum }_{i=1}^l\left({\alpha }_i^{\ast }-{\alpha }_i\right)y_i-\frac{1}{2}{\displaystyle \sum }_{i,j=1}^l\left({\alpha }_i^{\ast }-{\alpha }_i\right)\left({\alpha }_j^{\ast }+{\alpha }_j\right)x_i^T{x}_j$$

(13)

$$Subject to\{\begin{array}{c}{\displaystyle {\displaystyle \sum }_{i=1}^l}{\alpha }_i^{\ast }={\displaystyle {\displaystyle \sum }_{i=1}^l}{\alpha }_i\\ 0\le {\alpha }_i^{\ast }\le C\\ 0\le {\alpha }_i\le C\end{array}$$

As Lagrange multipliers $L_d\left(\alpha , {\alpha }^{\ast }\right)$ are obtained, the optimal desired weights vector w of the regression hyperplane can be solved, where w may be obtained by $w=\sum _{i=1}^l({\alpha }_i^{\ast }-{\alpha }_i)x_i$, while optimal bias b of the regression hyperplane may be found as $b=\frac{1}{l}(\sum _{i=1}^l{y}_i-x_i^{T}w)$. Therefore, after applying the Lagrange theory and the Karush–Kuhn–Tucker condition, the support vector regression for non-linear data may be expressed as;

$$f\left(x\right)={\displaystyle \sum }_{i=1}^l\left({\alpha }_i^{\ast }-{\alpha }_i\right)K\left(x_i,x\right)+b$$

(14)

where $K\left(x_i,x\right)$ defines the kernel function. Nonlinear input data can be separated linearly by mapping $\varphi \left(x_i\right)$ the inputs into the high-dimensional feature space by using the kernel function [20,25]. Function $K\left(x_i, x_j\right)$ defines the kernel function, where the value of the kernel function is equal to the inner products $x_i$ and $x_j$ in the feature space $\varphi \left(x_i\right)$ and $(x_j)$ [58]. Kernel function may be expressed as;

$$K\left(x,x_j\right)={\displaystyle \sum }_{i=1}^{\infty }{\lambda }_i{\varphi }_i\left(x_i\right){\varphi }_i\left(x_j\right)$$

(15)

Diverse kernels may be adopted to generate inner products to construct support vectors machines with a nonlinear decision surface in the input space [57]. Most commonly used kernel functions are Gaussian and polynomial kernel functions.

$$K_{gaussian}=e^{-\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{${\sigma }^2$}\right.\left[{\left(x-x_i\right)}^T\sum _{}^{-1}\left(x-x_i\right)\right]}$$

(16)

$$K_{polynomial}={\left[\left(x^T{x}_i\right)+1\right]}^d$$

where d and ${\sigma }^2$ denote the degree and the bandwidth of the kernel, respectively.

3. Results

3.1. Empirical Workflow

Primarily, every published news article related to macroeconomic indicators and decisions on economic policies are collected from the leading news publishers, in addition to authorities’ publishing services. Collected articles form the foundation of sentiment analysis. The annotated text of collected news is fed into learning automata to construct a lexicon dedicated predominantly to macroeconomic indicators, which will be authorized by the word polarity functions of existing lexicons dedicated to economics. Generated lexicons extract sentiments that annotated text comprises by examining the text word by word, producing an overall polarity vector value for every text feed into the algorithm. The obtained polarity vector will be evaluated by the historical macroeconomic data alongside examples of the USD/TRY prices that have been mentioned in Section 2.2 to obtain a prospect value that expresses the subjective feelings and valuations of the decision maker regarding news considering each macroeconomic indicator individually, which are designated as fuzzy antecedents, namely, interest rate, inflation rate, balance of trade, unemployment rate, and credit ratings. There are three universal outcomes for macroeconomic news and policy announcements directly perceived by the decision maker (which also have a direct psychological impact on the subjective sentiments and the process of decision making), namely: increase, steady, and decrease. These three universal consequences are appointed as membership functions for fuzzy antecedents. As interpretations of news are completely subjective, decision makers may also interpret consequences of macroeconomic announcements subjectively. For that reason, 245 fuzzy rules have been designated by the authors and implemented into the fuzzy engine to generate the fuzzy consequent. The fuzzy consequent is principally a dependency rate that signifies the correlation between The USD/TRY and news. The attained correlation value is used as input, along with historical the USD/TRY and macroeconomic data, to run support vector regression. The Gaussian radial basis function operated as the kernel function to project nonlinear historical data into high-dimensional space. The Monte Carlo cross validation method was adopted to split presented data into training and test groups to run a prediction. The accuracy of the prediction is investigated by the statistical tests. The prediction consistency of the model is evaluated against five widely used statistical tests. The empirical workflow described above is presented as a flowchart in Figure 1.

3.2. Empirical Results

The main objective of the sentiment analysis is to derive a subjective perspective of decision makers on particular events that took place in the news. To do so, a unique lexicon for macroeconomics is built by learning automata, which is supplemented by the polarity functions of existing lexicons on economics. The probability of sentiment outcome is extracted for each published news article. The prospect theory enhanced the outcome by observing the historical exchange rate prices. The subjective feelings of the decision makers on a variety of scenarios of macroeconomic indicators have been used to understand the valuation behavior of decision makers in particular scenarios. This procedure provided an enhanced overview of the content of the news, the perspective of decision makers on each scenario, and fuzzy antecedents for the next step.

The sentiment analysis has produced sentiment scores from 1 to 5 for every analyzed news item, where the value 5 defines substantial appreciation of the USD/TRY, values between 5 and 3 represent diminishing appreciation of the USD/TRY, a value of 3 defines steady state, values between 3 and 1 represent gradual depreciation of the USD/TRY, and a value of 1 defines very strong depreciation of the USD/TRY. The obtained dataset as the output of sentiment analysis is not fed into the fuzzy logic algorithm directly, but separated and allocated under the title of the related macroeconomic variable to form fuzzy antecedents.

Due to the acute effects of change on macroeconomic variables, the trapezoidal membership function has been selected for use. For each fuzzy antecedent, membership functions, namely ‘increase,’ ‘steady,’ and ‘decrease,’ representing appreciation, steady state, and depreciation of the USD/TRY, respectively, have been assigned to determine the degree of belonging to a fuzzy set based on the outcome of the sentiment analysis.

The fuzzy engine determined that the inflation rate, interest rate, and unemployment rate are positively correlated with the USD/TRY, while the balance of trade and credit ratings are inversely correlated with the USD/TRY. For that reason, values that denote the degree of belonging alter. For inflation rate, interest rate, and unemployment rate, values from 1 to 3 denote the degree of belonging to the fuzzy set ‘increase,’ a value 3 defines the degree of belonging to fuzzy set ‘steady,’ and values from 3 to 5 denote the degree of belonging to fuzzy set ‘decrease.’ Meanwhile, for the balance of trade and credit ratings, values from 1 to 3 denote the degree of belonging to fuzzy set ‘decrease,’ a value 3 defines the degree of belonging to fuzzy set ‘steady,’ and values from 3 to 5 denote the degree of belonging to fuzzy set ‘increase.’ The degree of membership to a fuzzy set denoted between 0 and 1, (while 0 and 1 represent 0% and 100% membership degree, respectively) is considered as a partial membership for belonging to a set. As each macroeconomic indicator has unique dynamics, significance, and influence on exchange rate, the membership function of the fuzzy logic algorithm is diversified and constructed specifically for each indicator, while fuzzy rules comprise macroeconomic indicators collectively, and they function regarding their weights. The dependency of the USD/TRY on the inflation rate, interest rate, balance of trade, unemployment rate, and credit ratings have been presented in Figure 2, Figure 3, Figure 4, Figure 5 and Figure 6, respectively.

The fuzzy consequents for each macroeconomic indicator, along with the consequent combination of macroeconomic fundamentals, have been obtained. The fuzzy consequents provided the correlation value, which is necessary to run the prediction using support vector regression. The obtained correlation values for each macroeconomic indicator specify that published news regarding the increase in inflation rate of Turkey, and the downgrading credit rating of Turkey, have significant impacts on the appreciation of the USD/TRY. However, published news on the increase in the unemployment rate of Turkey has an insignificant impact on the appreciation of the USD/TRY. On the other hand, the impact of published news regarding the decreased interest rate in Turkey on the depreciation of the USD/TRY is inevitably significant. Meanwhile, published news on the increase in balance of trade for Turkey has an insignificant impact on the depreciation of the USD/TRY. In the light of the correlation values of macroeconomic news for each indicator, the impulsive dependency of the exchange rate on macroeconomic news has been simulated. The simulation results present the significant impulsive dependency of the USD/TRY on recently published macroeconomic news. The simulation results verify and emphasize the importance and significance of the consideration of published macroeconomic news on understanding the fluctuations of the exchange rate.

The correlation values that have been obtained as fuzzy consequents have been used as input, in addition to historical the USD/TRY data and historical data of each macroeconomic indicator, to run support vector regression. As a final step, the USD/TRY prices are predicted in the light of the consequents achieved during sentiment analysis and fuzzy logic processes regarding the correlation between macroeconomic news and exchange rate prices. Nonlinear historical data have been projected into high-dimensional space by using the kernel function. The constructed hyperplane provides adjustability to converge optimal set of weights [17]. The Gaussian radial basis function was utilized as a kernel function due to its good fit on non-linear inputs [58]; the selected kernel function was tuned by C, ε, and γ hyperparameters. In order to set the hyperparameters, a grid search with cross validation was adopted. Grid search ranges of C, ε, and γ hyperparameters were assigned as: C ranges from $2^{-5} \mathrm{to} 2^{10}$, ε from $2^{-5} \mathrm{to} 2^3$, and γ from $2^{-10} \mathrm{to} 2^3$. The detection of potential misleading data, such as information leakage, is crucial to eliminate possible prediction error. To overcome possible prediction error along with the over-fitting issue, the ε-intensive zone was assigned by considering the absurd fluctuations in data [57]. The heat map that illustrates the accuracy of selected hyperparameters C and γ is presented as Figure 7.

The accuracy of the parameter selection using a grid search with cross validation is evaluated by the consideration of three metrics, namely precision, recall, and F1-score. Precision indicates the ratio of correctly predicted positive observations compared to the sum of positive observations. The ratio of correctly predicted positive observations compared to all classified samples is named recall. The weighted average of precision and recall can be expressed as the F1 score. The F1 score considers both false positives and false negatives. For that reason, in situations such as uneven distribution into classes, the F1 score is the preferred metric instead of accuracy. However, if false positives and false negatives are evenly distributed, the accuracy metric should be taken into consideration. The attained evaluation metrics have been presented in

Table 1. Evaluation of Hyperparameter Selection.

	Precision	Recall	F1-Score
0	1.00	1.00	1.00
1	0.93	0.97	0.94
2	0.97	0.96	0.96
Accuracy			0.96
Macro Avg	0.95	0.98	0.95
Weighted Avg	0.95	0.97	0.95

The critical values for accuracy, precision, recall, and F1-score are 0.20, 0.40, 0.60, and 0.80, denoting slight, fair, moderate, substantial, and almost perfect agreement, respectively.

below. The results indicated the accuracy of 0.96 for parameter selection, which indicates almost perfect agreement.

Macroeconomic indicators provided restricted historical data due to their nature. Therefore, splitting presented data into training and test groups is essential to appraise the performance of the regression model. To do so, the Monte Carlo cross validation (MCCV) method was adopted. A total of 85% of the provided data was used as trained data, while 15% was reserved as test data. MCCV randomly splits reserved portions of data into sub-samples and assigns them as test sets. The process of selecting random independent partitions was repeated multiple times. The prediction consistency of the model is evaluated against five widely used statistical metrics, namely V measure score (VMS), explained variance score (EVS), root mean squared error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination ($R^2$). VMS has been developed to evaluate the performance of clustering tasks. To do so, the technique requires two measures, homogeneity and completeness. Homogeneity is a measure that identifies the number of data points clustered from the same class. Completeness is a measure that features clustering the data points from the same class into same cluster. The main goal of V measure is to obtain a harmonic mean between homogeneity and completeness. EVS aims to measure inconsistency between the true values and predicted values. Additionally, it clarifies the total variance of the model, which can be explained by the factors presented by the actual data and not a part of variance error. EVS shares the same principle with $R^2$. $R^2$ is a probabilistic investigation metric that indicates the percentage of the variance for a dependent variable from the regression model. Conversely, EVS uses the biased variance, while $R^2$ uses the raw sum of squares. The difference between the mentioned metrics identifies the level of prediction bias. RMSE measures the dissimilarity between the predicted value and actual value. RMSE is very likely to contribute relatively high weight to greater error values due to squaring the errors prior to averaging, which makes RMSE advantageous in cases where high errors are exceptionally unfavorable. As RMSE output shares the identical unit with the predicted values, it is difficult to interpret if the value is significant, acceptable, or unacceptable. To do so, a scatter index (SI) is used to convert the RMSE value into an interpretable value. MAPE computes the accuracy of the prediction made through calculating the average of absolute error in percentage terms for each time period. MAPE calculates each variable individually, achieving normalized absolute error before averaging. For the purpose of validating the results, along with measuring the sustainability of the results, the Monte Carlo cross validation has repeated 40 times; the mean of the achieved statistical test results has been taken and presented below as

Table 2. Averaged Monte Carlo Cross Validation Statistical Test Results.

Statistical Test	Results
VMS	0.919392 ***
EVS	0.928443 ***
RMSE	0.225206 **
MAPE	0.039060 ***
$R^2$	0.924597 ***

The critical values of VMS are 0.70, 0.80, and 0.90. Therefore, **, and *** denote significant, and highly significant at 0.80, and 0.90, respectively [59]. The critical values of EVS are 0.60, 0.80, and 0.90. Therefore, **, and *** denote significant, and highly significant at 0.80, and 0.90, respectively [60]. The critical values of RMSE are set by SI at ≤10%, and ≤5%, which indicate an RMSE value of 0.520908 and 0.260454. Therefore, ** indicate the significance at respective critical points [61]. The critical values of MAPE are 0.25, 0.10, and 0.05. Therefore, **, and *** denote low but acceptable, and highly acceptable accuracy at 0.25, 0.10, and 0.05, respectively [62]. The critical values of $R^2$ are 0.50, 0.75, and 0.90. Therefore, **, and *** denote moderate, and substantial prediction at 0.50, 0.75, and 0.90, respectively [63,64]. The critical value of the difference between EVS and $R^2$ is ≤2%.

.

The MCCV statistical test results indicate that:

The VMS result is 0.919392, which is higher than 0.90 criteria. Therefore, it is possible to denote VMS value as highly significant, thus indicating a harmonic mean between homogeneity and completeness, indicating an accomplished clustering task for the future weight support vector regression.

The EVS result is 0.928443, which is higher than the 0.90 criteria. Therefore, it is possible to signify the EVS value as highly significant. The results indicate that the variance can be explained by the factors presented by the actual data. The highly significant EVS result, 0.928443, denotes that the variance can be explained by the factors presented by the actual data.

The RMSE value is in the same unit with the original data, indicating the standard deviation of errors emerged during prediction, thus signifying the accuracy of the model. The RMSE result is 0.225206, which denotes that the standard deviation of error is 0.225206 when compared to the actual the USD/TRY value. In order to evaluate the RMSE result, the value was transformed into an interpretable value, SI. The transformed RSME SI value is 0.043233, which is lower than 0.05 criteria. Thus, it is possible to denote the RMSE value as significant.

The accuracy of the prediction has also been measured by the MAPE results. The MAPE result, 0.039060, is equivalent to 3.90%, which is lower than the 0.05 criteria, equivalent to 5%. Thus, it is possible to say that the model has highly acceptable accuracy.

The $R^2$ result is 0.924597, which is higher than the 0.90 criteria. These highly significant results have verified that the prediction is substantial with a superior accuracy, as the highly significant majority of the predicted data points are on the regression line.

Moreover, the deviation between the EVS and $R^2$ results, 0.928443 and 0.924597, respectively is only 0.42%, which is lower than the 2% criteria. Thus, the deviation between EVS and $R^2$ results signifies that the prediction is unbiased.

When the obtained statistical test results are compared with the existing literature, it is possible to observe the enhanced performance of the composed hybrid machine learning algorithm. Bollen et al. [18] achieved a predictive accuracy of 86.7% by using sentiment analysis, Weng et al. [16] achieved a predictive accuracy of 85.8% by using hybrid DT, NN, and SVM, Schumaker and Chen [19] achieved 0.04261 directional accuracy by using SVM, Sharma et al. [3] achieved MAPE 8.68% by using hybrid ANN and FL, while the top 10 performing algorithms of the M4 Time Series Forecasting Competition achieved a MAPE between 3% and 10% [65].

The prediction efficiency is graphically illustrated in Figure 8 below. The actual exchange prices have been illustrated in blue, while the predicted prices are illustrated in orange.

When evaluated in daily matters, macroeconomic news may react deficiently in comprehending the relationship between actual and predicted data, due to a lack of continuity of macroeconomic news taking place in the media. In order to understand the reasoning behind the detected deficiency in comprehending the relationship between actual and predicted data, and figuring out a solution for the attained reason, a detailed observation has been made. After detailed observations of the historical data and events, reputable VMS, EVS, RMSE, MAPE, and $R^2$ results revealed the importance and usability of recently published macroeconomic news on understanding and predicting the fluctuations of exchange rate prices.

4. Conclusions

This article has developed a novel approach to evaluate the importance of the sustainable macroeconomic policies, causality of exchange rate stability on a daily basis, and the significance of predicting future rates by implementing a robust methodology. Daily news regarding the inflation rate, unemployment rate, interest rate, balance of trade, and credit rating have been examined alongside the historical data in order to observe the causality between news concerning macroeconomic policies and exchange rate volatility. To achieve this, related news and historical data have been collected to form the necessary database for their examination using the robust machine learning methodology. The machine learning methodology encapsulates learning automata for textual analysis, prospect theory for associating the textual analysis scores with change in exchange rate prices, fuzzy logic for comprehending the degrees of dependency of exchange rate volatility on macroeconomic news concerning each macroeconomic indicator, and support vector regression for predicting the exchange rate by using variables obtained by previous phases of the methodology, besides historical data. Evaluations on news articles to derive subjective sentiments and perceptions of the decision makers, along with historical prices, indicated a significant correlation between exchange rate volatility and the news regarding the inflation rate, interest rate, and credit ratings, while insignificant correlation was detected for the balance of trade and the unemployment rate, providing an enhanced understanding of the reasons behind the economic fluctuations that are causing unsustainable and unstable economic policies. By considering the correlation values obtained in the fuzzy logic phase, support vector regression analysis was conducted as a final step. A successful and strong prediction for the USD/TRY exchange rate was obtained in the final phase.

The result of this study verifies the significant impact of the perceived macroeconomic condition, expectation, and events by investors on the sustainability of macroeconomic policies, the stability of exchange rate, and the prediction future rates. The stable attitude and implementation of required policies of the institutions and organizations subject to the government play a significant role in economic sustainability, as they play a vital role in eliminating uncertainty and lack of trust, thus boosting the confidence of investors in their investment decisions. If the trust of investors in the sustainability of policies and the economy cannot be obtained, the responsiveness of the market, and thus, the influence of manipulations, become unnecessarily huge, even if the allegations are baseless. For that reason, a commitment to coping with policy instability and uncertainty plays a vital role in sustainable exchange rate stability, thus causing the economic sustainability of emerging economies. Furthermore, this study reveals that not only the objective quantitative variables, but also subjective qualitative information can be used to obtain enhanced prediction performance by examining decision makers’ behaviors in particular scenarios.