Modeling Markov Switching ARMA-GARCH Neural Networks Models and an Application to Forecasting Stock Returns

The study has two aims. The first aim is to propose a family of nonlinear GARCH models that incorporate fractional integration and asymmetric power properties to MS-GARCH processes. The second purpose of the study is to augment the MS-GARCH type models with artificial neural networks to benefit from the universal approximation properties to achieve improved forecasting accuracy. Therefore, the proposed Markov-switching MS-ARMA-FIGARCH, APGARCH, and FIAPGARCH processes are further augmented with MLP, Recurrent NN, and Hybrid NN type neural networks. The MS-ARMA-GARCH family and MS-ARMA-GARCH-NN family are utilized for modeling the daily stock returns in an emerging market, the Istanbul Stock Index (ISE100). Forecast accuracy is evaluated in terms of MAE, MSE, and RMSE error criteria and Diebold-Mariano equal forecast accuracy tests. The results suggest that the fractionally integrated and asymmetric power counterparts of Gray's MS-GARCH model provided promising results, while the best results are obtained for their neural network based counterparts. Further, among the models analyzed, the models based on the Hybrid-MLP and Recurrent-NN, the MS-ARMA-FIAPGARCH-HybridMLP, and MS-ARMA-FIAPGARCH-RNN provided the best forecast performances over the baseline single regime GARCH models and further, over the Gray's MS-GARCH model. Therefore, the models are promising for various economic applications.


Introduction
In the light of the significant improvements in the econometric techniques and in the computer technologies, modeling the financial time series have been subject to accelerated empirical investigation in the literature. Accordingly, following the developments in the nonlinear techniques, analyses focusing on the volatility in financial returns and economic variables are observed to provide significant contributions. It could be stated that important steps have been taken in terms of nonlinear measurement techniques focusing on the instability or stability occurring vis-a-vis encountered volatility. Further, the determination of stability or instability in terms of volatility in the financial markets gains importance especially for analyzing the risk encountered. In addition to impact of the magnitude and the size of shocks on volatility, the financial returns are under the influence of sudden or abrupt changes in the economy. Hence, the volatility of economic data has been explored in econometric literature as a result of the need of modelling uncertainty and risk in the financial returns. The relationship between the financial returns and various important factors such as the trade volume, market price of financial assets, and the relationship between volatility, trade volume, and financial returns have been vigorously investigated [1][2][3][4].
The ARCH model introduced by Engle [5] and the Generalized ARCH (GARCH) model introduced by Bollerslev [6] are generally accepted for measuring volatility in financial models. GARCH models have been used intensively in academic studies. A tremendous amount of GARCH models exist and various studies provide extended evaluation of the development. 2 The Scientific World Journal Among many, Engle and Bollerslev [7] developed the Integrated-GARCH (I-GARCH) process to incorporate integration properties, AGARCH model, introduced by Engle [8], allows modeling asymmetric effects of negative and positive innovations. In terms of modeling asymmetries, GARCH models have been further developed by including asymmetric impacts of the positive and negative shocks to capture the asymmetric effects of shocks on volatility and return series which depend on the type of shocks, i.e. either negative or positive. Following the generalization of EGARCH model of Nelson [9] that allows modelling the asymmetries in the relationship between return and volatility, the Glosten et al. [10] noted the importance of asymmetry caused by good and bad news in volatile series and proposed a model that incorporates the past negative and positive innovations with an identity function that leads the conditional variance to follow different processes due to asymmetry. The finding is a result of the empirical analyses which pointed at the fact that the negative shocks had a larger impact on volatility. Consequently, the bad news have a larger impact compared to the conditional volatility dynamics followed after the good news. Due to this effect, asymmetric GARCH models have rapidly expanded. The GJR-GARCH model was developed independently by Zakoian [11,12] and Glosten et al. [10]. It should be noted that, in terms of asymmetry, the Threshold GARCH (T-GARCH) of Zakoian [12], VGARCH, and nonlinear asymmetric GARCH models (NAGARCH) of Engle and Ng [13] are closely related versions to model asymmetry in financial asset returns. The SQR-GARCH model of Heston and Nandi [14] and the Aug-GARCH model developed by Duan [15] nest several versions of the models taking asymmetry discussed above. Further, models such as the Generalized Quadratic GARCH (GQARCH) model of Sentana [16] utilize multiplicative error terms to capture volatility more effectively. The FIGARCH model of Baillie et al. [17] benefits from an ARFIMA type fractional integration representation to better capture the long-run dynamics in the conditional variance (See for detailed information, Bollerslev [18]). The APARCH/APGARCH model of Ding et al. [19] is an asymmetric model that incorporates asymmetric power terms which are allowed to be estimated directly from the data. The APGARCH model also nests several models such as the TGARCH, TSGARCH, GJR, and logGARCH. The FIAPGARCH model of Tse [20] combines the FIGARCH and the APGARCH. Hyperbolic GARCH (HYGARCH) model of Davidson [21] nests the ARCH, GARCH, IGARCH, and FIGARCH models (for an extended review GARCH models, see Bollerslev [18]).
Even though the ARCH/GARCH models can be applied quickly for many time series, the shortcomings in these models were discussed by certain studies. Perez-Quiros and Timmermann [22] focused on the conditional distributions of financial returns and showed that recessionary and expansionary periods possess different characteristics, while the parameters of a GARCH model are assumed to be stable for the whole period. Certain studies discussed the high volatility persistence inherited in the baseline GARCH and proposed early signs of regime switches. Diebold [23] and Lamoureux and Lastrapes [24] are two of the highly cited studies discussing high persistence in volatility due to structural changes. Lamoureux and Lastrapes [24] showed that the encountered high persistence in volatility processes resulted from the volume effects that had not been taken into account. Qiao and Wong [25] followed a bivariate approach and confirmed that the Lamoureux and Lastrapes [24] effect exists due to the volume and turnover effects on conditional volatility and after the introduction of volume/turnover as exogenous variables, it is possible to obtain a significant decline in the the persistence. Mikosch and Stǎricǎ [26] showed that structural changes had an important impact that leads to accepting an integrated GARCH process. Bauwens et al. [27,28] discussed that the persistence in the estimated single regime GARCH processes could be considered as resulting from the misspecification which could be controlled by introducing an MS-GARCH specification where the regime switches are governed by a hidden Markov chain.
Krämer [29] evaluated the autocorrelation in the squared error terms and provided an important contribution. Accordingly, the observed empirical autocorrelations of the 2 are much larger than the theoretical autocorrelations implied by the estimated parameters through evaluating an MS-GARCH model where the autocorrelation problem could be shown to accelerate as the transition probabilities approached 1. (For a proof see e.g., Francq and Zakoïan [30]; Krämer [29]) (In particular, the empirical autocorrelations of the 2 often seem to indicate long memory, which is not possible in the GARCH-model; in fact, in all standard GARCH-models, theoretical autocorrelations must eventually decrease exponentially, so long memory is ruled out). Alexander and Lazaar [31] showed that leverage effects are due to asymmetry in the volatility responses to the price shocks and the leverage effect accelerates once the markets are in the more volatile regime Krämer and Tameze [32] showed that a single state GARCH model had only one mean reversion while by allowing regime switching in the GARCH processes, mean reverting effect diminishes. In a perspective of volatility, if these shifts are persistent, then there are two sources of volatility persistence, due to shocks and due to regime-switching in the parameters of the variance process. By utilizing a Markov transformation model, it could be shown that the relationships among the regimes between the periods of − 1 and could be explained and the most important advantage of the MSGARCH model exposes itself as there is no need for the researchers to observe the regime changes. The model allows different regimes to reveal by itself [33].
The regime switching in light of the Markov switching model has interesting properties to be examined such as the stationarity by allowing the switching course of volatility inherent in the asset prices. The hidden Markov model (HMM) developed by Taylor [34] is a switching model that benefits from including an unobserved variable to capture volatility to be modeled with transitions between the hidden states that possess different probability distributions attached to each state. Hidden Markov model has been applied successfully by Alexander and Dimitriu [35], Cheung and Erlandsson [36], Francis and Owyang [37], and by Clarida et al. [38] to capture the switching type of predictions in stock returns, interest rates, and exchange rates. Regime switching model has been used extensively for prediction of returns belonging to different stock market returns in different economies and by following the fact that the stock market indices are very sensitive to stock volatility, which accelerates especially during periods with market turbulences (see for detailed information, Alexander and Kaeck, [39]).
The conventional statistical techniques for forecasting reached their limit in applications with nonlinearities, furthermore, recent results suggest that nonlinear models tend to perform better in models for stock returns forecasting [40]. For this reason, many researchers have used artificial neural networks methodologies for financial analysis on the stock market. Lai and Wong [41] contributed to the nonlinear time series modeling methodology by making use of singlelayer neural network. Further, modeling of NN models for estimation and prediction for time series have important contributions. Weigend et al. [42], Weigend and Gershenfeld [43], White [44], Hutchinson et al. [45], and Refenes et al. [46] contributed to financial analyses, stock market returns estimation, pattern recognition, and optimization. NN modeling methodology is applied successfully by Wang et al. [47] and Wang [48] to forecast the value of stock indices. Similarly, Abhyankar et al. [49], Castiglione [50], Freisleben [51], Kim and Chun [52], Liu and Yao [53], Phua et al. [54], Refenes et al. [55], Resta [56], R. Sitte and J. Sitte [57], Tiňo et al. [58], Yao and Poh [59], and Yao and Tan [60] are important investigations focusing on the relationships between stock prices and market volumes and volatility. For similar applications, see [1][2][3][4]. Bildirici and Ersin [61] modeled NN-GARCH family models to forecast daily stock returns for short and long run horizons and they showed that GARCH models augmented with artificial neural networks (ANN) architectures and algorithms provided significant forecasting performances. Ou and Wang [62] extended the NN-GARCH models to Support Vector Machines. Azadeh et al. [63] evaluated NN-GARCH models and proposed the integrated ANN models. Bahrammirzaee [64] provided an analysis based on financial markets to evaluate the artificial neural networks, expert systems, and hybrid intelligence systems. Further, Kanas and Yannopoulos [65] and Kanas [40] used Markov switching and Neural Networks techniques for forecasting stock returns; however, their approaches depart from the approach followed within this study.
In this study, the neural networks and Markov switching structures are aimed to be integrated to augment the ARMA-GARCH models by incorporating regime switching and different neural networks structures. The approach aims at

The MS-GARCH Models
Over long periods, there are many reasons why financial series exhibit important breaks in behavior; examples include depression, recession, bankruptcies, natural disasters, and market panics, as well as changes in government policies, investor expectations, or the political instability resulting from regime change.
Diebold [23] provided a throughout analysis on volatility models. One of the important findings is the fact that volatility models that fail to adequately incorporate nonlinearity are subject to an upward bias in the parameter estimates which results in strong forms of persistence that occurs especially in high volatility periods in financial time series. As a result of the bias in the parameter estimates, one important result of this fact is on the out-of-sample forecasts of single regime type GARCH models. Accordingly, Schwert [66] proposed a model that incorporates regime switching that is governed by a two state Markov process, hence the model retains different characteristics in the regimes that are defined as high volatility and low volatility regimes.
Hamilton [67] proposed the early applications of HMC models within a Markov switching framework. Accordingly, MS models were estimated by maximum likelihood (ML) where the regime probabilities are obtained by the proposed Hamilton-filter [68][69][70][71]. ML estimation of the model is based on a version of the Expectation Maximization (EM) algorithm as discussed in Hamilton [72], Krolzig [73][74][75][76]. In the MS models, regime changes are unobserved and are a discrete state of a Markov chain which governs the endogenous switches between different AR processes throughout time. By inferring the probabilities of the unobserved regimes which are conditional on an information set, it is possible to reconstruct the regime switches [77].
Furthermore, certain studies aimed at the development of modeling techniques which incorporate both the probabilistic properties and the estimation of a Markov switching ARCH and GARCH models. A condition for the stationarity of a natural path-dependent Markov switching GARCH model as in Francq et al. [78] and a throughout analysis of the probabilistic structure of that model, with conditions for the existence of moments of any order, are developed and investigated in Francq and Zakoïan [30]. Wong and Li [79], Alexander and Lazaar [80], and Haas et al. [81][82][83] derived stationarity analysis for some mixing models of conditional heteroskedasticity [27,28]. For the Markov switching GARCH models that avoid the dependency of the conditional variance on the chain's history, the stationarity conditions are known for some special cases in the literature [84]. Klaassen [85] developed the conditions for stationarity of the model as the special cases of the two regimes. A necessary and sufficient stationarity condition has been developed by Haas et al. [81][82][83] for their Markov switching GARCH model. Furthermore, Cai [86] showed the properties 4 The Scientific World Journal of Bayesian estimation of a Markov switching ARCH model where only the constant in the ARCH equation is allowed to have regime switches. The approach has been investigated by Kaufman and Frühwirth-Schnatter [87] and Kaufmann and Scheicher [88]. Das and Yoo [89] proposed an MCMC algorithm for the same model (switches being allowed in the constant term) with a single state GARCH term to show that gains could be achieved to overcome path-dependence. MS-GARCH models are studied by Francq and Zakoïan [30] to achieve their non-Bayesian estimation properties in light of the generalized method of moments. Bauwens et al. [27,28] proposed a Bayesian Markov chain Monte Carlo (MCMC) algorithm that is differentiated by including the state variables in the parameter space to control the path-dependence by obtaining the parameter space with Gibbs sampling [90].
The high and low volatility probabilities of MS-GARCH models allow differentiating high and low volatility periods. By observing the periods in which volatility is high, it is possible to investigate the economic and political reasons that caused increased volatility. If a brief overview is to be presented, there are several models based on the idea of regime changes which should be mentioned. Schwert [66] explores a model in which switches between these states that returns can have a high or low variance are determined by a two-state Markov process. Lamoureux and Lastrapes [24] suggest the use of Markov switching models for a way of identifying the timing of the shifts in the unconditional variance. Hamilton and Susmel [91] and Cai [86] proposed Markov switching ARCH model to capture the effects of sudden shifts in the conditional variance. Further, Hamilton and Susmel [91] extended the analysis to a model that allows three regimes, which were differentiated between low, moderate and high volatility regimes, where the high-volatility regime captured the economic recessions. It is accepted that the proposals of Cai [86] and Hamilton and Susmel [91] helped the researchers to control for the problem of path dependence, which makes the computation of the likelihood function impossible (The conditional variance at time t depends on the entire sequence of regimes up to time t due to the recursive nature of the GARCH process. In Markov switching model, the regimes are unobservable, one needs to integrate over all possible regime paths. The number of possible paths grows exponentially with t, which renders ML estimation intractable.) (see for detail, Bauwens, et al. [27,28]).
Gray [92] study is one of the important studies where a Markov switching GARCH model is proposed to overcome the path dependence problem. According to Gray's model, once the conditional volatility processes are differentiated between regimes, an aggregation of the conditional variances for the regimes could be used to construct a single variance coefficient to evaluate the path dependence. A modification is also conducted by Klaassen [85]. Yang [93], Yao and Attali [94], Yao [95], and Francq and Zakoïan [96] derived conditions for the asymptotic stationarity of some AR and ARMA models with Markov switching regimes. Haas et al. [81][82][83] investigated a MS-GARCH model by which a finite state-space Markov chain is assumed to govern the ARCH parameters, whereas the autoregressive process followed by the conditional variance is subject to the assumption that past conditional variances are in the same regime (for details, the readers are referred to Bauwens et al. [27,28], Klaassen [85], Haas et al. [81][82][83], Francq and Zakoïan [30], Krämer [29], and Alexander and Kaeck [39]).
Another area of analysis pioneered by Haas [97] and Chang et al. [98] allow different distributions in order to gain forecast accuracy. An important finding of these studies showed that by allowing the regime densities to follow skewnormal distribution with Gaussian tail characteristics, several return series could be modeled more efficiently in terms of forecast accuracy. Liu [99] developed and discussed the conditions for stationarity in Markov switching GARCH structure in Haas et al. [81][82][83] and proved the existence of the moments. In addition, Abramson and Cohen [100] discussed and further evaluated the stationarity conditions in a Markov switching GARCH process and extended the analysis to a general case with m-state Markov chains and GARCH( , ) processes. An evaluation and extension of the stationarity conditions for a class of nonlinear GARCH models are investigated in Abramson and Cohen [100]. Francq and Zakoïan [30] derived the conditions for weak stationarity and existence of moments of any order MS-GARCH model. Bauwens et al. [27,28] showed that by enlarging the parameter space to include space variables, though maximum likelihood estimation is not feasible, the Bayesian estimation of the extended process is feasible for a model where the regime changes are governed with a hidden Markov chain. Further, Bauwens et al. [27,28] accepted mild regularity conditions under which the Markov chain is geometrically ergodic and has finite moments and is strictly stationary.
Yang [103], Yao and Attali [94], Yao [95], and Francq and Zakoïan [96] derived conditions for the asymptotic stationarity of models with Markov switching regimes (see for detailed information Bauwens and Rombouts [104]. The major differences between Markov switching GARCH models are the specification of the variance process; that is, the conditional variance 2 = Var( / ). To consider the conditional variance as in the Bollerslev's [105] GARCH model and to consider the regime dependent equation for the conditional variance in Frömmel [106] are accepted that The coefficients , , correspond to respective coefficients in the one-regime GARCH model, but may differ depending on the present state.
Klaassen [85] (Klassen [85] model is defined as 2 , − , − ) suggested to use the conditional expectation of the lagged conditional variance with a broader information set than the model derived in Gray [92]. Accordingly, Klaassen [85] suggested modifying Gray's [92] model by replacing Another version of MS-GARCH model is developed by Haas et al. [81][82][83]. According to this model, Markov chain controls the ARCH parameters at each regime ( , , ) and the autoregressive behavior in each regime is subject to the assumption that the past conditional variances are in the same regime as that of the current conditional variance [100].
In this study, models will be derived following the MS-ARMA-GARCH specification in the spirit of Blazsek and Downarowicz [107] where the properties of MS-ARMA-GARCH processes were derived following Gray [92] and Klaassen [85] framework. Henneke et al. [108] developed an approach to investigate the model derived in Francq et al. [78] for which the Bayesian framework was derived. The stationarity of the model was evaluated by Francq and Zakoïan [96] and an algorithm to compute the Bayesian estimator of the regimes and parameters was developed. It should be noted that the MS-ARMA-GARCH models in this paper were developed by following the models developed in the spirit of Gray [92] and Klaassen [85] similar to the framework of Blazsek and Downarowicz [107].
The MS-ARMA-GARCH model with regime switching in the conditional mean and variance are defined as a regime switching model where the regime switches are governed by an unobserved Markov chain in the conditional mean and in the conditional variance processes as where, Thus, the parameters have nonnegativity constraints , , , , , > 0 and the regimes are determined by , and the probability Pr[ = | −1 ] is calculated through iteration: Accordingly, the two models, the Henneke et al. [108] and the Francq et al. [78] approaches, could be easily differentiated through the definitions of 2 −1 and −1 . Further, asymmetric power terms and fractional integration will be introduced to the derived model in the following sections.

MS-ARMA-APGARCH Model.
Liu [99] provided a generalization of the Markov switching GARCH model of Haas et al. [81][82][83] and derived the conditions for stationarity and for the existence of moments. Liu [99] proposes a model which allowed for a nonlinear relation between past shocks and future volatility as well as for the leverage effects. The leverage effect is an outcome of the observation that the reaction of stock market volatility differed significantly to the positive and the negative innovations. Haas et al. [109,110] complements Liu's [99] work in two ways. Firstly, the representation of the model developed by Haas [109] allows computational ease for obtaining the unconditional moments. Secondly, the dynamic autocorrelation structure of the power-transformed absolute returns (residuals) was taken as a measure of volatility.
Haas [109] model assumes that time series { , ∈ Z} follows a k regime MS-APGARCH process, The stationary distribution of Markov-chain is shown as According to the Liu [99] notation of MS-APGARCH model, the conditional variance 2 of jth regime follows a univariate APGARCH process as follows: The Scientific World Journal where, > 0, 1 , 2 , ≥ 0, = 1, . . . , . For the power term = 2 and for 1 = 2 , the model in (8) reduces to MS-GARCH model. Similar to the Ding et al. [19], the asymmetry, which is called "leverage effect, " is captured by 1 ̸ = 2 [109]. If the past negative shocks have deeper impact, parameters are expected to be 1 < 2 so that the leverage effect becomes stronger.
The MS-ARMA-GARCH type model specification in this study assumes that the conditional mean follows MS-ARMA process, whereas the conditional variance follows regime switching in the GARCH architecture. Accordingly, MS-ARMA-APGARCH architecture nests several models by applying certain restrictions. The MS-ARMA-APGARCH model is derived by moving from MS-ARMA process in the conditional mean and MS-APGARCH( , ) conditional variance process as follows: where the regime switches are governed by ( ) and the parameters are restricted as ( ) > 0, ,( ) , ,( ) ≥ 0 with ,( ) ∈ (−1, 1), = 1, . . . , . One important difference is that MS-ARMA-APGARCH model in (10) allows the power parameters to vary across regimes. Further, if the following restrictions are applied, = 1, = 1, ( ) = , the model reduces to the model of Haas [109] given in (9).
In applied economics literature, it is shown that many financial time series possess long memory, which can be fractionally integrated. Fractional integration will be introduced to the MS-ARMA-APGARCH model given above.

MS-ARMA-FIAPGARCH Model .
Andersen and Bollerslev [111], Baillie et al. [17], Tse [112], and Ding et al. [19] provided interesting applications in which the attention had been directed on long memory. Long memory could be incorporated to the model above by introducing fractional integration in the conditional mean and the conditional variance processes.
MS-ARMA-FIAPGARCH derived is a fractional integration augmented model as follows: where the lag operator is denoted by , autoregressive parameters are ( ) , and ( ) shows the moving average parameters, (1) > 0 denotes the optimal power transformation, the fractional differentiation parameter varies between 0 ≤  [17]. To differentiate between the GARCH specifications, forecast performance criteria comparisons are assumed.

Neural Network and MS-ARMA-GARCH Models
In this section of the study, the MultiLayer Perceptron, Radical Basis Function, and Recurrent Neural Network models that belong to the ANN family will be combined with Markov switching and GARCH models. In this respect, Spezia and Paroli [113] is another study that merged the Neural Network and MS-ARCH models.

MS-ARMA-GARCH-MLP Model. Artificial Neural
Network models have many applications in modeling of functional forms in various fields. In economics literature, the early studies such as Dutta and Shektar [114], Tom and Kiang [115], Do and Grudinsky [116], Freisleben [51], and Refenes et al. [55] utilize ANN models to option pricing, real estates, bond ratings, and prediction of banking failures among many, whereas studies such as Kanas [40], Kanas and The Scientific World Journal 7 Yannopoulos [65], and Shively [117] applied ANN models to stock return forecasting, and Donaldson and Kamstra [118] proposed hybrid modeling to combine GARCH, GJR, and EGARCH models with ANN architecture. The MLP, an important class of neural networks, consists of a set of sensory units that constitute the input layer, one or more hidden layers, and an output layer. The additional linear input which is connected to the MLP network is called the Hybrid MLP. Hamilton model can also be considered as a nonlinear mixture of autoregressive functions, such as the multilayer perceptron and thus, the Hamilton model is called Hybrid MLP-HMC models [119]. Accordingly, in the HMC model, the regime changes are dominated by a Markov chain without making a priori assumptions in light of the number of regimes [119]. In fact, Hybrid MLP accepts the network inputs to be connected to the output nodes with weighted connections to form a linear model that is parallel with nonlinear Multilayer Perceptron.
In the study, the MS-ARMA-GARCH-MLP model to be proposed allows Markov switching type regime changes both in the conditional mean and conditional variance processes augmented with MLP type neural networks to achieve improvement in terms of in-sample and out-ofsample forecast accuracy.
The MS-ARMA-GARCH-MLP model is defined of the form: where, the regimes are governed by unobservable Markov process: In the MLP type neural network, the logistic type sigmoid function is defined as and ( = | −1 ), the filtered probability with the following representation, if , transition probability ( = | −1 = ) is accepted; → max{ , } recursive procedure is started by constructing ( = | −1 ), where ( ℎ ) is of the form 1/(1 + exp(− )), a twice-differentiable, continuous function bounded between [0, 1]. The weight vector = ; = logistic activation function and input variables are defined as ℎ = , where ℎ is defined as in (16).

MS-ARMA-APGARCH-MLP Model. Asymmetric power GARCH (APGARCH) model has interesting features.
In the construction of the model, the APGARCH structure of Ding et al. [19] is followed. The model given in (13) where, regimes are governed by unobservable Markov process. The model is closed as defining the conditional mean as in (12) and conditional variance of the form equation's (14)- (19) and (20)

(22)
Following the division of regression space into two subspaces with Markov switching, the model allows two different asymmetric power terms, (1) and (2) , and two different fractional differentiation parameters, (1) and (2) ; as a result, different long memory and asymmetric power structures are allowed in two distinguished regimes.
It is possible to show the model as a single regime NN-FIAPGARCH model if = 1: Further, the model reduces to NN-FIGARCH if = 1 and ( 1 ) = = 2 in the fashion of Bildirici and Ersin [61]: models are trained to generate both time series forecasts and certainty factors. Accordingly, RBF neural network is represented as a composition of three layers of nodes; first, the input layer that feeds the input data to each of the nodes in the second or hidden layer; the second layer that differs from other neural networks in that each node represents a data cluster which is centered at a particular point and has a given radius and in the third layer, consisting of one node.

MS-ARMA-GARCH-RBF
where = 1, . . . , regimes are governed by unobservable Markov process: A Gaussian basis function for the hidden units given as ( ) for = 1, 2, . . . , where the activation function is defined as, With defining the width of each function. is a vector of lagged explanatory variables, + < 1 is essential to ensure stationarity. Networks of this type can generate any real-valued output, but in their applications where they have a priori knowledge of the range of the desired outputs, it is computationally more efficient to apply some nonlinear transfer function to the outputs to reflect that knowledge.
is the filtered probability with the following representation: → max{ , } recursive procedure is started by constructing ( = | −1 ).

MS-ARMA-APGARCH-RBF Model.
Radial basis functions are three-layer neural network models with linear output functions and nonlinear activation functions defined as Gaussian functions in hidden layer utilized to the inputs in light of modeling a radial function of the distance between the inputs and calculated value in the hidden unit. The output unit produces a linear combination of the basis functions to provide a mapping between the input and output vectors:

Recurrent Neural Network MS-GARCH Models.
The RNN model includes the feed-forward system; however, it distinguishes itself from standard feed-forward network models in the activation characteristics within the layers. The activations are allowed to provide a feedback to units within the same or preceding layer(s). This forms an internal memory system that enables a RNN to construct sensitive internal representations in response to temporal features found within a data set. The Jordan [122] and Elman's [123] networks are simple recurrent networks to obtain forecasts: Jordan and Elman networks extend the multilayer perceptron with context units, which are processing elements (PEs) that remember past activity. Context units provide the network with the ability to extract temporal information from the data. The RNN model employs back propagation-through-time, an efficient gradient-descent learning algorithm for recurrent networks. It was used as a standard variant of cross-validation referred to as the leave-one-out method and as a stopping criterion suitable for estimation problems with sparse data and so it is identified the onset of overfitting during training. The RNN was functionally equivalent to a nonlinear regression model used for time-series forecasting (Zhang et al. [124]; Binner et al. [125]). Tiňo et al. [126] merged the RNN and GARCH models.

MS-ARMA-GARCH-RNN
Similar to the models above, (32) is shown for = 1, . . . , regimes which are governed by unobservable Markov process. Activation function is taken as the logistic function.

The Data.
In order to test forecasting performance of the abovementioned models, stock return in Turkey is calculated by using the daily closing prices of Istanbul Stock Index ISE 100 covering the 07.12.1986-13.12.2010 period corresponding to 5852 observations. To obtain return series, the data is calculated as follows: = ln( / −1 ), where ln(⋅) is the natural logarithms, is the ISE 100 index, and is taken as a measure of stock returns. In the process of training the models, the sample is divided between training, test, and outof-sample subsamples with the percentages of 80%, 10%, and 10%. Further, we took the sample size for the training sample as the first 4680 observations, whereas the sample sizes for the test and out-of-sample samples are 585 and 587. The statistics of daily returns calculated from ISE 100 Index are given in Table 1.
In order to provide out-of-sample forecasts of the ISE100 daily returns, two competing nonlinear model structures are used, the univariate Markov switching model and Neural Network Models, MLP, RBF, and RNN. In order to assess the predictability of models, models are compared for their outof-sample forecasting performance. Firstly, by calculating RMSE and MSE error criteria, the forecast comparisons are obtained.

Econometric
Results. At the first stage, selected GARCH family models taken as baseline models are estimated for evaluation purposes. Results are given in Table 2. Included models have different characteristics to be evaluated, namely, fractional integration, asymmetric power, and fractionally integrated asymmetric power models, namely, GARCH, APGARCH, FIGARCH, and FIAPGARCH models. Random walk (RW) model is estimated for comparison purpose. Furthermore, the models given in Table 2 will provide basis for nonlinear models to be estimated.
It is observed that, though all volatility models perform better than the RW model in light of Log Likelihood criteria, as we move from the GARCH model to asymmetric power GARCH (APGARCH) model, the fit of the models improve accordingly. The sum of ARCH and GARCH parameters is calculated as 0.987 and less than 1. The results for the APGARCH model show that the calculated power term is 1.35 and the asymmetry is present. Further, similar to the findings of McKenzie and Mitchell [127], it is observed that the addition of the leverage and power terms improves generalization power and thus show that squared power term may not necessarily be optimal as Ding et al. [19] study suggested.
In Table 3, transition matrix and the MS model were estimated. The standard deviation takes the values of 0.05287 and 0.014572 for regime 1 and regime 2. It lasts approximately 75.87 months in regime 1 and 107.61 months in regime 2. By using maximum likelihood approach, MS-GARCH models are tested by assuming that the error terms follow studentdistribution with the help of BFGS algorithm. Number of regimes are taken as 2 and 3. GARCH effect in the residuals is tested and at 1% significance level, the hypothesis that there are no GARCH effects is rejected. Additionally, the normality in the residuals is tested with Jacque-Berra test, at 1% significance level, it is detected that the residuals are not normally distributed. As a result, MS-GARCH model is estimated under the distribution assumption. In the MS-GARCH model, the transition probability results are calculated as Prob( = 1 | −1 = 1) = 0.50 and Prob( = 2 | −1 = 2) = 0.51 and show that the persistence is low in the MS-GARCH model.
Statistical inference regarding the empirical validity of two-regime switching process was carried out by using nonstandard LR tests [128]. The nonstandard LR test is statistically significant and this suggests that linearity is strongly rejected.
The volatility values tend to be calculated at higher values than they actually are in GARCH models. As the persistence coefficients obtained for the MS-GARCH, MS-PGARCH, and MS-APGARCH models are compared to those obtained for the GARCH models,; the persistence in the GARCH models are comparatively higher and for the GARCH models; a shock in volatility is persistent and shows continuing effect. This situation occurs as a result of omitting the importance of structural change in the ARCH process. In the MS-GARCH, MS-PGARCH, and MS-APGARCH models which take this situation into consideration, the value of persistence parameter decreases.
Power terms are reported comparatively lower for developed countries than the less developed countries. Haas et al. [109,110]    On the other hand, though the improvement by shifting to modeling the conditional volatility with regime switching is noteworthy, the desired results are still not obtained, therefore, MS-GARCH models are extended with MLP, RBF, and RNN models and their modeling performances are tested.

MS-ARMA-GARCH-Neural Networks Results.
In the study, model estimation is gathered through utilizing backpropagation algorithm and the parameters are updated with respect to a quadratic loss function, whereas the weights are iteratively calculated with weight decay method to achieve the lowest error. Alternative methods include Genetic Algorithms [132][133][134] and 2nd order derivative based optimization algorithms such as Conjugate Gradient Descent, Quasi-Newton, Quick Propagation, Delta-Bar-Delta, and Levenberg-Marquardt, which are fast and effective algorithms but may be subject to overfitting (see [135][136][137]). In the study, we followed a two-step methodology. Firstly, all models were trained over a given training sample vis-àvis checking for generalization accuracy in light of RMSE criteria in test sample. The approach is repeated for estimating each model for 100 times with different number of sigmoid activation functions in the hidden layer. Hence, to obtain parsimony in models, best model is further selected with respect to the AIC information criterion. For estimating NN-GARCH models with early stopping combined with algorithm corporation, readers are referred to Bildirici and Ersin [61].
The estimated models are reported in Table 4. For comparative purposes, MSE and RMSE values for training sample are given. Among the MS-ARMA-GARCH-NN models and for the training sample, the lowest RMSE value is achieved as 0.114 by the MS-ARMA-GARCH-RBF model, followed by MS-ARMA-GARCH-MLP model with a RMSE value of 0.186 as the second best model among the models with GARCH specifications noted at the first part of Table 4. Furthermore, the above mentioned models are followed by MS-ARMA-GARCH-HYBRID MLP model that deserves the 3rd place with RMSE = 0.187, though the value is almost equal to the value obtained for the MS-ARMA-GARCH-MLP model having the 2nd place. Further, MS-ARMA-GARCH-RNN, and MS-ARMA-GARCH-ELMAN RNN models take the 4th and 5th place among the GARCH type competing models in terms of in-sample accuracy only.
The Markov-switching models with asymmetric power terms in the conditional volatility, namely, the, MS-ARMA-APGARCH-NN models are reported in the second part of Table 4

14
The Scientific World Journal  The results obtained for the test sample and the holdout sample are given in Tables 5 and 6. If an overlook is to be provided, it is observed that holding the training sample results on one side, the real improvement occurs for forecasting. When we move from MS-ARMA-GARCH-NN models (denoted as Model Group 1) towards MS-ARMA-APGARCH (denoted as Model Group 2) and fractionally integrated models of MS-ARMA-FIAPGARCH-NN (Model Group 3) models, though certain amount of in-sample forecast gain is achieved, the MSE and RMSE values are the lowest for test sample and, most importantly, for the out-ofsample forecasts.
To evaluate the test sample performance of the estimated models, MSE and RMSE error criteria are reported in Table 5. Accordingly, among the MS-ARMA-GARCH-NN models, the lowest RMSE value for the test sample is 0.1238 and is calculated for the MS-ARMA-GARCH-MLP model that is followed by the MS-ARMA-GARCH-RNN model with a RMSE of 0.1242. The MS-ARMA-GARCH-HYBRID MLP model takes the 3rd place in the test sample, similar to the results obtained for the training sample with RMSE = 0.1243, though the value is almost equal to the models taking the first and the second places. Compared with the first three models we mentioned with RMSE values that are achieved as low as around 0.124, MS-ARMA-GARCH-ELMAN RNN, and MS-ARMA-GARCH-RBF models take the 4th and 5th places with RMSE values equal to 0.1538 and 0.1546. The results show that, for the test sample, MS-ARMA-GARCH-MLP, MS-ARMA-GARCH-RNN, and MS-ARMA-GARCH-HYBRID MLP models have very similar results and well generalization capabilities, though the last 2 competing models also have promising results. Furthermore, as we evaluate the APGARCH type models, the generalization capabilities increase significantly. Among the MS-ARMA-APGARCH-NN models given in the table, MS-ARMA-APGARCH-HYBRID MLP model has the lowest RMSE (=0.00011539) followed by the MS-ARMA-APGARCH-MLP model (RMSE = 0.00011548). The third best model is the MS-ARMA-APGARCH-ELMAN RNN model with RMSE = 0.00011789 followed by the 4th and 5th models; MS-ARMA-APGARCH-RNN (RMSE = 0.0001276) and MS-ARMA-APGARCH-RBF (RMSE = 0.0001631), respectively. As discussed, models in the third group are MS-ARMA-FIAPGARCH-NN models. According to the results obtained in Table 5 Table 6, the models are compared for their respective out-of-sample forecast performances. At the first part of Table 6