Pythagorean fuzzy time series approach based on hybrid high-order artificial neural network for nonlinear time series forecasting Authors: Ozge Gundogdu, Erol Egrioglu, Eren Bas Published: 2026 ABSTRACT AbstractFuzzy time series forecasting methods rely on the fuzzification of time-series observations through membership functions and the subsequent use of fuzzy-valued variables in the forecasting process. Although Pythagorean fuzzy time series models provide a richer representation of uncertainty than classical and intuitionistic fuzzy approaches, existing methods still face important challenges, including high-dimensional input structures generated by membership and non-membership values, a lack of systematic lag selection mechanisms, and limited capability to model complex nonlinear relationships. To address these limitations, this study employs an artificial neural network with a hybrid architecture combining a modified Pi-Sigma artificial neural network and simple exponential smoothing to capture both the linear and nonlinear dynamics and model the relationship in Pythagorean fuzzy time series. To mitigate the high-dimensional input structure arising from membership and non-membership values, principal component analysis was used for input dimension reduction and orthogonalization. Particle swarm optimization was used to estimate all the parameters in the Pythagorean fuzzy time series method. To evaluate its performance, the proposed forecasting algorithm was assessed against existing approaches in the literature using 10 distinct sub-time series generated via a rolling window approach from the S&P 500 and Dow-Jones 30 stock market composite indices, and the proposed method’s superior performance was demonstrated experimentally. From an artificial intelligence perspective, these contributions jointly define a novel Pythagorean fuzzy time series forecasting framework that advances hybrid neural network modelling and optimization, while from an engineering perspective, they provide a systematic and implementable methodology for constructing accurate forecasting models for complex nonlinear time series, particularly in financial and other engineering-oriented applications.Keywords: Pythagorean fuzzy time series, Modified pi-sigma artificial neural network, Simple exponential smoothing, Principal component analysis, Particle swarm optimizationSubject terms: Engineering, Mathematics and computing Introduction The fuzzy time series (FTS) modeling approach has been widely used in forecasting problems because it reveals the uncertainty inherent in time series, eliminates the limitations of traditional time series methods, and requires no assumptions. FTS models and modeling approaches employed in forecasting problems, along with the associated limitations of these methods, have been discussed concisely in the literature1–3. The concept of FTS, utilized in time series forecasting, originated from the adaptation of fuzzy set theory, originally introduced in4, to time series data, as presented in5. In this context, historical time series were transformed into linguistic values, and fuzzy relations were employed to model the resulting FTS. Furthermore, two distinct FTS models, namely, the first-order time-invariant model and the first-order time-variant model, were proposed to address forecasting problems, and the forecasting procedures are detailed in6,7. In time-variant FTS models, the relationships within the FTS framework vary over time, and capturing this variability in forecasting necessitates the use of different model structures. In this way, while such models offer increased flexibility and adaptability, they also entail higher computational complexity. Although time-invariant FTS models are frequently used in the literature, time-variant models are not commonly used. Nevertheless, studies such as8–15 exemplify the development and application of time-variant FTS models and related forecasting methods. The estimation steps in the FTS modeling approach can be expressed as determining the interval lengths in which the universal set will be divided, the definition of linguistic terms for intervals, fuzzification of time series data, determination of fuzzy relationships and relationship groups between the fuzzified time series data, defuzzification, and forecasting computation. The procedures performed in the remarked steps significantly affected the forecasting performance and accuracy. The FTS model introduced in5 was refined through various enhancements aimed at increasing forecasting accuracy and ease of computation. A new method based on the FTS introduced in5,6 was proposed in16. In this method, fuzzy relationship tables are used instead of operators for the fuzzy set in the process of determining fuzzy relationships. This approach has formed the basis for later studies by simplifying the construction of fuzzy relationships and making the modeling process in FTS forecasting more efficient and understandable. In the first step of the FTS modeling approach, the universal set is defined as a closed interval and is divided according to a fixed interval length in the5–7,16,17 studies. However18, stated that the forecasting accuracy can change depending on the interval length. In FTS modeling, the forecasting performance is aimed to be improved by determining the interval length based on the mean and distribution of the data. The determination of the interval length and number may cause differences in the forecasting results by affecting all fuzzy relationship identification processes. In this context, several studies have focused on determining the optimal interval length and number, as demonstrated in18–27. Optimization-based approaches have also been employed to improve fuzzy partitioning and forecasting performance. Beyond interval optimization, several studies have integrated metaheuristic optimization techniques into FTS frameworks to enhance forecasting accuracy28. proposed a variational weighted fuzzy time series model optimized using particle swarm optimization and reported improved forecasting accuracy across different benchmark datasets29. proposed a forecasting methodology that combines fuzzy particle swarm optimization with reprogrammed large language models to improve prediction performance. A novel formulation for high-order FTS models was introduced in30, targeting enhanced forecasting performance and improved computational efficiency. Subsequently31, introduced a novel high-order FTS model along with a forecasting algorithm aimed at enhancing the performance of first-order FTS models. Comparative analyses with widely recognized methods in the academic literature have demonstrated that higher-order models yield more accurate results than first-order models, which consider fuzzy relationships only one period ahead in forecasting time series. The high-order FTS model provides complex relationships in real-world data to be revealed more clearly by handling the relationships comprehensively. This feature of high-order FTS has been investigated by32–43 and used for time series forecasting in different fields. Although the high-order FTS approach improves the forecasting accuracy, it also increases the model complexity and computational burden. In FTS models based on fuzzy set theory introduced by4, only the membership degree of data points is taken into account during the fuzzification process. In this sense, uncertainty cannot be fully reflected in the model44,45. proposed intuitionistic fuzzy sets as an extended form of classical fuzzy set theory to overcome this limitation. According to the intuitionistic fuzzy set approach, there are membership and non-membership degrees to the set for each element in the universal set, and it provides a more enhanced representation of this uncertainty. In addition, the nondeterminacy degree of an element is defined. Intuitionistic fuzzy sets have also been adapted to time series analysis due to their capacity to provide more accurate and comprehensive solutions for decision-making under uncertainty. In46, a novel approach to perform fuzzy inference using intuitionistic fuzzy systems was proposed, and it was indicated that this approach could also be applied to time series forecasting. In47, an intuitionistic fuzzy model and forecasting method with simplified computational procedures were developed by utilizing the intuitionistic fuzzy set approach to identify fuzzy relationships in time series forecasting. Similarly48, proposed an intuitionistic FTS forecasting method that incorporates both intuitionistic fuzzification and defuzzification stages. In this way, the uncertainty inherent in time series data can be modeled more effectively. Furthermore, studies49–56 employed the intuitionistic fuzzy set approach to better capture uncertainty and increase the forecasting precision of time series models. Building upon these developments57, proposed a new algorithm based on intuitionistic FTS, focusing on forecasting nonstationary time series. Subsequently58, presented a multifactor high-order intuitionistic FTS model and employed a fuzzy clustering algorithm to generate unequal intervals in the analysis. Furthermore59, proposed a high-order intuitionistic FTS forecasting model and method in which fuzzy relationships were identified using a Pi–Sigma neural network60. developed a mixed-order intuitionistic fuzzy time series forecasting method integrating a novel artificial neural network architecture with the grey wolf optimization algorithm. In light of this development, based on the intuitionistic fuzzy sets approach, picture fuzzy sets were presented in61 as an extension of classical fuzzy and intuitionistic fuzzy sets. Thereafter62, proposed a new concept within the context of picture fuzzy sets. This approach allows for the modeling of situations that involve additional response options such as ‘abstention’ and ‘rejection’, in addition to conventional responses such as ‘yes’ and ‘no’. According to this approach, the degrees of positive, negative, and neutral membership and refusal to belong to a cluster are calculated for each element. The definition of uncertainty that is more flexible and extensive is obtained by using the degree of abstention as distinct from the intuitionistic fuzzy sets. From this perspective, picture fuzzy sets provide more information about the data and enhance the overall modeling capabilities. Building on this foundation63,64, proposed hybrid forecasting methods based on picture fuzzy set theory and evaluated their performance in the context of weather nowcasting. In line with the uncertainty definition provided by picture fuzzy sets65, introduced the concept of a picture FTS. In the present study, a high-order picture FTS model and a forecasting algorithm are proposed. In66, a high-order multivariate picture FTS forecasting model and its solution method were proposed, combining picture fuzzy set theory with a Pi–Sigma artificial neural network (ANN) for time series forecasting. Similarly67, presented a hybrid picture FTS forecasting model that integrates the picture fuzzy set approach with a new type of particle swarm optimization. Nevertheless, certain methodological limitations can be identified. Despite the innovative integration, the study does not propose a specific method for determining the hyperparameters of the model, nor does it employ any test data to assess its forecasting performance. Instead, the evaluation is limited to in-sample forecasting, which restricts the generalizability of the findings. Although recent studies have explored time series forecasting using ambiguous set theory and quantum optimization algorithms68–70, these approaches mainly focus on uncertainty representation and fuzzy partition optimization. In contrast, the present study concentrates on improving the predictive capability of Pythagorean fuzzy time series through dimensionality reduction, hybrid neural modelling, and metaheuristic parameter estimation. In parallel with these studies aimed at enhancing the modeling capabilities of fuzzy sets71–73, introduced the concept of Pythagorean fuzzy subsets and Pythagorean membership degrees, which are a more comprehensive version of intuitionistic fuzzy set membership degrees. In the Pythagorean fuzzy set, for any element in the universal set, the membership degree (μ), non-membership degree (ν), and hesitation degree (π) must satisfy the condition . This condition ensures that the square sum of the membership and non-membership degrees does not exceed one. It provides a wider search space than intuitionistic fuzzy sets, where the sum is linear (i.e., ), thus allowing for more flexibility in representing uncertainty. Following this advancement74, introduced n-Pythagorean fuzzy sets to provide a more accurate expression of fuzziness by extending the search space of membership degrees. n-Pythagorean fuzzy sets can be described as a generalized form of intuitionistic fuzzy sets and Pythagorean fuzzy sets. In the n-Pythagorean fuzzy sets, for any element in the universal set, the membership degree (μ), non-membership degree (ν), and hesitation degree (π) must satisfy the condition , where n ≥ 1 is a real number. This generalized condition allows for a more flexible modeling of uncertainty compared to the intuitionistic fuzzy set (n = 1) and Pythagorean fuzzy set (n = 2) cases75. proposed a FTS model and forecasting method grounded in the fundamental properties of n-Pythagorean fuzzy sets. In their approach, the n-Pythagorean fuzzy c-means algorithm was employed for the fuzzification of data points and segmentation of the universal set. Additionally, an improved Markov weighted estimation method was utilized in the defuzzification phase of the FTS to enhance the forecasting performance of the model. More recently76, proposed a Pythagorean fuzzy time series forecasting method based on a wolf-raven optimization algorithm for partition. Their results demonstrated that optimization-based partitioning strategies can significantly improve forecasting accuracy by enhancing the quality of fuzzy partitions. Fundamentally, the main differences between fuzzy sets arise from the definition of membership degrees and the way uncertainty is formulated. In practice, the selection of the appropriate fuzzy set approach to a given problem is crucial for both estimation accuracy and a precise representation of uncertainty. These conceptual differences directly influence the construction and performance of fuzzy time series forecasting models. Despite significant advancements in uncertainty modeling, a comprehensive review of the current literature reveals that several critical research gaps remain unaddressed in the domain of fuzzy time series forecasting. Pythagorean fuzzy time series models improve uncertainty representation through membership and non-membership degrees; however, existing approaches still suffer from several methodological limitations. The simultaneous use of membership and non-membership values generates high-dimensional input structures that may increase the model complexity and reduce the forecasting efficiency. Second, existing Pythagorean fuzzy time series studies generally lack systematic and statistically grounded procedures for selecting lagged variables, leading to potential model misspecifications and suboptimal forecasting performance. Third, most existing approaches have limited capability to capture the complex nonlinear relationships inherent in real-world time series, particularly in financial and engineering applications. Consequently, there remains a need for a comprehensive Pythagorean fuzzy time series forecasting method that can simultaneously address dimension reduction, lag selection, and nonlinear modeling within a unified methodology. The motivation for this study is to present an effective Pythagorean FTS method based on Pythagorean fuzzy sets. This study proposes and evaluates a Pythagorean FTS approach that is based on artificial neural networks for the first time in the literature. By integrating neural network-based learning mechanisms into the Pythagorean fuzzy time series structure, the study aims to enhance the modeling capability for complex nonlinear time series and to extend the applicability of Pythagorean fuzzy set theory to advanced forecasting problems. The innovations and contributions of the proposed method are listed below. i.First ANN-Based Pythagorean FTS Method: The proposed method is the first Pythagorean FTS method based on a hybrid ANN in the literature. First ANN-Based Pythagorean FTS Method: The proposed method is the first Pythagorean FTS method based on a hybrid ANN in the literature. ii.Comprehensive Linear and Nonlinear Modeling: Thanks to the use of a hybrid ANN, a modeling approach that combines a complex nonlinear model with simple exponential smoothing has been preferred. Therefore, the method captures both the linear and nonlinear dynamics within time series data. Comprehensive Linear and Nonlinear Modeling: Thanks to the use of a hybrid ANN, a modeling approach that combines a complex nonlinear model with simple exponential smoothing has been preferred. Therefore, the method captures both the linear and nonlinear dynamics within time series data. iii.Systematic Lag Selection Mechanism: A new Pythagorean FTS forecasting model is defined. The proposed Pythagorean FTS forecasting model is autoregressive in structure, and confidence limits created for partial autocorrelation coefficients are used in determining the lagged variables. Thus, the problem of selecting lagged variables in the forecasting model is solved. Systematic Lag Selection Mechanism: A new Pythagorean FTS forecasting model is defined. The proposed Pythagorean FTS forecasting model is autoregressive in structure, and confidence limits created for partial autocorrelation coefficients are used in determining the lagged variables. Thus, the problem of selecting lagged variables in the forecasting model is solved. iv.Dimensionality Reduction via Principal Component Analysis: Principal component analysis (PCA) was used to model the lagged variables of membership and non-membership values. Thus, instead of a model with too many variables, a model with fewer variables was created using membership and non-membership values with a reduced dimension. In addition, the correlations between membership and non-membership values were eliminated, and orthogonal principal components were used. Dimensionality Reduction via Principal Component Analysis: Principal component analysis (PCA) was used to model the lagged variables of membership and non-membership values. Thus, instead of a model with too many variables, a model with fewer variables was created using membership and non-membership values with a reduced dimension. In addition, the correlations between membership and non-membership values were eliminated, and orthogonal principal components were used. The main advantage of the proposed method is that it improves uncertainty representation, reduces model complexity, provides statistically grounded lag selection, and enhances forecasting performance through nonlinear learning capabilities. PCA reduces the complexity of membership and non-membership information, PACF-based lag selection provides an objective autoregressive structure, and the hybrid ANN captures both linear and nonlinear patterns of the data. Consequently, the proposed method produces a statistically grounded flexible forecasting framework. The remaining sections of the article are organized as follows. The second section provides brief information on the autocorrelation and partial autocorrelation coefficients, principal component analysis, particle swarm optimization, simple exponential smoothing, and Pythagorean fuzzy c-means methods used as tools in the proposed method. The third section introduces the proposed hybrid artificial neural network. In the fourth chapter, the Pythagorean FTS definitions and the proposed FTS forecasting model are introduced, and the algorithm of the proposed Pythagorean FTS method is presented. The findings of performance comparisons and applications of the proposed method are presented in the fifth chapter. The discussion in the final chapter focuses on the results that were obtained, and an attempt is made to determine a route for future studies. Tools used in developing the proposed method Many tools previously proposed in the literature were used in developing the proposed method. These tools are mathematical and statistical tools, which are briefly introduced in this section. Autocorrelation and partial autocorrelations The most effective method for investigating the internal relationship of a time series is to examine the autocorrelation and partial autocorrelation coefficients. Autocorrelation coefficients are important tools that reveal the internal relationships within a time series. Partial autocorrelations, on the other hand, are a type of partial correlation coefficient calculated by fixing the effect of the lagged variables in between when the step length of the calculated autocorrelation coefficient is 2 or more. Autocorrelation coefficients represent the relationship between the error term and lagged variables, while partial autocorrelation coefficients can be used to determine the number of autoregressive lags. The number of lags in a linear autoregressive model can be determined by partial autocorrelation coefficients that are not equal to zero. This approach, which is used for linear models, can also be used as a preliminary tool in nonlinear modeling. The autocorrelation coefficients are calculated using Eq. (1), while the partial autocorrelation coefficients are calculated using Eq. (2). Hypothesis tests can be used to determine whether the autocorrelation and partial autocorrelation coefficients are different from zero for the population. Although many test statistics with a chi-square distribution are defined in the literature, an easier approach is to use the standard errors of the autocorrelation and partial autocorrelation. For a white noise process with all autocorrelations equal to zero, the variance of the partial autocorrelations is 1/n. Therefore, partial autocorrelations outside the limits can be considered to be different from zero. Principal component analysis In multivariate statistical analyses, high correlations between independent variables and a large number of independent variables cause the models created to be ineffective and to work with too many parameters, resulting in low statistics in the degrees of freedom used in model testing. In principal component analysis, new variables consisting of linear combinations of independent variables that can represent a large portion of the total variation are obtained by utilizing the eigenvalues of the covariance or correlation matrix. These linear components, which are fewer than the total number of variables, are also uncorrelated, allowing models with fewer parameters to be used as the inputs. Σ, where Σ is the variance–covariance matrix calculated for the independent variables, and the eigenvalues of this matrix are obtained from Eq. (3). The smallest number that satisfies the following inequality in dimension reduction is determined as k. This number is the smallest number of principal components that explains at least 95% of the variance. Thus, using principal component analysis, it is possible to work with variables, which are unrelated and fewer in number, instead of variables (p≥ k). Particle swarm optimization Particle swarm optimization is an artificial intelligence optimization algorithm that mimics the foraging methods of bird flocks. In this algorithm, a set of candidate solutions for the problem is used, and this set employs the intelligent strategies of a bird flock. Just like in a flock of birds, the candidate solutions use each other’s information to determine their new positions in the search space. Each candidate solution has a position and the current speed at which it uses to change this position. When a candidate solution changes its position, it uses the memory created by the steps it takes in the search space and the shared memory of all candidate solutions. While both memories play a critical role in determining the speed of each candidate solution, the weighting of each memory is determined randomly in order to explore different points in the search space and reduce the risk of getting stuck in a local optimum. Let the speed value V(i,j) of the i-th candidate solution be indicated for the j-th position. Let PB be the matrix containing the memories of each candidate solution, and GB be the vector containing the common memory of the candidate solution set. The velocity of a candidate solution is determined using Eq. (5). In this equation, represents the j-th position of i-th candidate solution. and show distinct random numbers from the standard uniform distribution. is a cognitive coefficient and this coefficient is used to control the effect of the candidate solution’s own memory in finding the new velocity value. is a social coefficient, and this coefficient is used to control the effect of shared memory in finding the new velocity value. w is the inertia weight, and this coefficient is very similar to the momentum term in the back propagation algorithm in neural network training algorithms. This coefficient is used to increase the convergence speed of the algorithm. The new position values of the particles are computed using Eq. (6). It is expected that all candidate solutions will converge as the iterations progress. Although candidate solutions generally converge into a single solution, in the case of local optima, candidate solutions may converge into two or more solution points as the iterations progress. One way to stop the iterations of the particle swarm optimization is to determine whether the solutions are sufficiently close to each other. However, it is also possible to specify different stopping conditions that are specific to the problem. Simple exponential smoothing One of the most basic forecasting methods is simple exponential smoothing, where forecasts are a weighted average of the previous true value and the previous forecast. The simple exponential smoothing method works by assigning more weight to the observations closest to the period for which a forecast is to be obtained. The weights of the observations decrease exponentially. The simple exponential smoothing method is known to be equivalent to a first-order integrated moving average model. In the simple moving average model, forecasts are calculated by Eq. (7). The average of the series can be used to obtain an initial forecast. α is a smoothing parameter and its value lies on the unit interval. If α equals 1, the model transforms to a random walk. Since the method has the potential to adapt to recent observations, it can produce successful forecasting results. Pythagorean fuzzy c-means clustering method Fuzzy clustering methods are employed to derive membership values by clustering the observations of variables according to their similarity. A clustering method based on n-Pythagorean fuzzy sets is presented in72. The method works with logic similar to that of the intuitionistic fuzzy clustering method. The objective of the method is to minimize the objective function in Eq. (8). In (8) equation, is n-Pythagorean membership, d(.) is the Euclidean distance and is the cluster center. Moreover, N is the number of learning samples and c is the number of clusters. n-Pythagorean membership values are calculated by using (9). In Eq. (9), m is the fuzziness index value, and its value can be taken as 2. Another membership value used in the n-Pythagorean fuzzy clustering method is the membership value combined with the degree of hesitation given in Eq. (10). denotes the degree of n-Pythagorean hesitation and is calculated by Eq. (11). In Eq. (11), is a non-membership value. The degrees of membership, non-membership, and hesitation collectively sum to one. In Eq. (9), cluster centers are calculated by using Eq. (13). In the clustering method, membership and non-membership values are initially generated randomly. Iterations are advanced by successively calculating formulas (9) and (13). When the iterations are stopped, the change in membership values is observed. When the change is small enough, the algorithm is stopped. The proposed hybrid artificial neural network In this study, a novel artificial neural network is proposed for determining fuzzy relationships in Pythagorean FTS. The architectural structure of the proposed ANN is shown in Fig. 1. The proposed ANN is a combination of a modified pi sigma ANN and simple exponential smoothing. Previously, an architectural structure combining a pi-sigma ANN with simple exponential smoothing has been presented in the literature77. The proposed architectural structure in this study contains a significant difference from the structure in77. In77, the output is obtained from the feedback prediction pi-sigma ANN in the exponential smoothing section, whereas in this study, it is obtained from the output of the exponential smoothing section. Thus, the proposed new architectural structure offers a complete combination of predictions from the pi-sigma ANN modified with exponential smoothing. Another difference from77 is that instead of pi-sigma architecture, its generalized form, modified pi-sigma architecture, is used. The architecture of the proposed hybrid artificial neural network. The equations to be employed for calculating the outputs of the proposed ANN are given below in order. First, the signals arriving and output at the hidden layer units in the Modified Pi-Sigma section are obtained using Eq. (14). The output of the modified pi-sigma ANN section is calculated using Eq. (15). The output of the hybrid ANN is obtained using Eq. (17). Here, ξ is the parameter that provides the combination of each part. The proposed hybrid ANN has number of trainable weights and biases. The weights and biases of the proposed hybrid ANN must be estimated during training. This study used particle swarm optimization for training. The optimization problem can generally be expressed as follows. The variables to be optimized in the optimization problem are given in Eq. (19) in the θ parameter. The proposed pythagorean fuzzy time series forecasting method study In this study, a new Pythagorean FTS method is proposed. The proposed method can be considered a new generation forecasting method. The process of obtaining forecasts is based on revealing the nonlinear mathematical function between the forecasts and the inputs determined in relation to the forecasts. Inputs with linear or nonlinear relationships with the forecast may be deterministic functions of time, lags in errors produced by the system, lagged variables in the time series, or nonlinear transformations of all these variables. Determining what inputs will be used in the forecasting method is one of the key factors influencing the forecasting method’s performance. Therefore, one of the issues studied and resolved in this study is the selection of method inputs. In the proposed method, membership and non-membership values, which are nonlinear combinations of lagged variables, have been added to the system using Pythagorean fuzzy clustering. In determining lagged variables, the method of identifying significant partial autocorrelation coefficients was preferred. Thanks to this method, lagged variables with the least linear relationship were included in the model, and these variables were used in nonlinear modeling. Although this approach is not optimal, it can be used as an effective method for reducing the number of inputs. Another important issue is the high correlation between the inputs of the method. The high correlation of commonly used inputs shows that it is possible to work with fewer inputs. In this study, a solution was developed by decoupling the inputs associated with principal component analysis and reducing the dimension. Estimating the nonlinear function between inputs and predictions is the most challenging part of the modeling process. A new hybrid ANN is proposed for this task. Since the proposed ANN is combined with a simple exponential smoothing method, it is a mixture of both complex and simple models in this study. Unlike other FTS in the literature, this approach combines the strengths of classical and contemporary modeling methods in modeling. The optimization problem that arose in the proposed method was solved using particle swarm optimization. In the literature, FTS is defined in different ways. This study is based on the definition of Pythagorean FTS given in75. The Pythagorean FTS forecasting method is implemented using a newly defined forecasting model, which is described below. Definition 1 Modeling the Pythagorean FTS is the process of determining the f function given below. This model is referred to as a high-order complex Pythagorean FTS forecasting model. Here, is the zero mean and constant variance error term, and f is a nonlinear function. A hybrid ANN is used to determine f. shows the number of lags q determined according to the partial autocorrelation functions. membership and non-membership values for lags are the reduced forms obtained as a result of the k basic components of the variables created. The algorithm of the proposed method is presented in detail, step by step below. In addition, the flowchart of the proposed method is given in Fig. 2. The flowchart of the proposed forecasting method. Algorithm 1 The proposed forecasting algorithm Step 1. The initial parameters of the recommended method are determined. These parameters are determined by the user. : In the proposed method, time series observations are divided into three separate clusters. The division is performed in a block structure in a way that does not disrupt the autocorrelation structure of the time series observations. The training time series is used to predict the forecast model f, while the validation set is used to determine the hyperparameters. Finally, the test set is used to demonstrate the performance of the proposed method and compare it with other forecasting methods. The parameters show the minimum and maximum values of the fuzzy set, respectively. The parameters are the minimum and maximum values for the number of hidden layer units in the hybrid ANN to be used in modeling. The parameters are, in order, the number of particles used in particle swarm optimization, the speed limit value, and the maximum number of iterations. Step 2. The time series is split according to the specified values to create training, validation, and test sets. Step 3. Partial autocorrelation coefficients for the training time series ( are calculated using formula (2), and correlations outside the limits of non-zero autocorrelations are determined. The lag numbers of non-zero partial autocorrelations are determined as . Step 4. Once the hyperparameter set has been selected, steps 4.1–4.6 are applied to this set. Step 4.1. By iteratively applying the Pythagorean fuzzy clustering method (9–13) formulas to the training time series, membership values and non-membership values are obtained. The membership values matrix is denoted as M, and the non-membership values matrix is denoted as NM, as shown below. Step 4.2. The M and NM matrices are created for the t time index. Each column of the matrices is like a time series consisting of membership and non-membership values. The lagged variables of each column are created according to and brought together in the L-N&NM matrix. Step 4.3. Dimension reduction and independence are achieved by applying principal component analysis (3–4) formulas to the L-N&NM matrix. The columns of the L-N&NM matrix are replaced by the reduced matrix (IM) principal component scores. The IM matrix is an n× k dimensional matrix. Step 4.4. XLM matrix consisting of lagged variables from the time series is obtained. Lagged variables are created according to set . The system inputs matrix (SI) is created from the combination of the IM and XLM matrices. Step 4.5. Using the system inputs, the ANN shown in Fig. 1 is applied by taking the observations for time t as target values. Particle swarm optimization, whose formulas are given in (5–6), is used in the training of the time series. Formula (17) is used to calculate the forecast values. The fitness function formula used in training with particle swarm optimization is the mean square error given in (18). During training, particle swarm optimization controls the percentage change in MSE as an early stopping condition. If there is no improvement in the percentage change for 20 consecutive iterations, training is completed. In addition, every 50 iterations, position values are randomly regenerated to jump to various regions of the search space in an attempt to avoid local optima traps. Step 4.6. The forecasts of the method for the validation set are computed using the trained ANN, and the root mean square error value for the validation set is calculated based on these forecasts. Step 5. The hyperparameter set with the lowest value is determined. For this parameter set, the training and validation sets are combined and retrained, and the RMSE value is calculated for the test set. Step 6. Step 5 is repeated 30 times with 5 different random starting values, minimum and maximum, mean and standard deviation statistics of the RMSE test values are calculated. The proposed forecasting method consists of the following stages. First, Pythagorean fuzzy c-means clustering is applied to obtain membership and non-membership values from the original time series. Next, significant lagged variables are identified using partial autocorrelation analysis. Principal component analysis is employed to reduce the dimension and eliminate correlations among the generated fuzzy inputs. The resulting inputs are processed by the proposed hybrid artificial neural network to model the underlying nonlinear relationships. Particle swarm optimization is used to estimate the network parameters and optimize forecasting performance. The overall workflow of the proposed forecasting method is summarized in Fig. 2 and detailed in Algorithm 1. Applications In the application, two separate stock market composite indices were studied. These composite indices are the S&P 500 and Dow-Jones 30 stock market composite indices. The time series used were obtained by selecting blocks with a rolling window approach from these two separate composite indices. The graphs of the composite index time series are shown in Figs. 3 and 4, respectively. Both time series consist of opening prices between January 3, 2023, and December 1, 2023. The time series graph of “S&P 500 opening prices”. The time series graph of “Dow-Jones 30 opening prices”. The start and end dates of the selected sub-time series with the rolling window approach used in the application are given in Table 1. Table 1Start and end dates of randomly selected time series.Stock price nameSeries NoStart dateEnd date#ObservationsS&P 500 Opening Price1Jan 03Aug 091512Sep 071713Oct 051914Nov 022115Dec 01231Dow-Jones 30 Opening Price1Jan 03Aug 091512Sep 071713Oct 051914Nov 022115Dec 01231 Start and end dates of randomly selected time series. A total of 10 time series were used to investigate the proposed method’s performance. The time series were obtained using a rolling window approach based on data from two globally recognized stock exchanges. This allowed the training and test sets of the time series to be shifted forward in time. Additionally, the lengths of the training data sets were varied to examine how the method’s performance changed when trained with fewer or more data points. First, five time series belonging to the S&P 500 stock index were analyzed. A similar hyperparameter selection strategy was used for all alternative methods in the analysis. The time series was divided into blocks, with the first part serving as training, the second part as validation, and the last part as a test set. Validity and test set lengths are equal and set to 20. In comparing the performance of the proposed method, the FTS network approach (FTS-N) proposed in78, the FTS approach (FTS-SMN) based on the single multiplicative neuron model proposed in80, and the fuzzy picture fuzzy function method (PFF) and classical FTS approach proposed by79 the picture fuzzy function method (PFF), and31 the high-order FTS approach (HO-FTS) as a classical FTS approach. The results obtained from applying the methods to the S&P 500 series are summarized in Table 2. The and rows in the table show the best hyperparameter values. In the HO-FTS method, only a single result is shown in the table because random initial values have no effect, and the results of the method do not vary. In the application of all methods, the fuzzy set number and the number of lagged variables in the model input were varied between 1 and 12, and one of a total of 144 hyperparameter sets was selected. Table 2 shows that the recommended method has an optimal fuzzy set number of 2 and an optimal input number of 1 for all series. In addition, the proposed method has the lowest RMSE average value in all five S&P 500 series. While hyperparameter values vary from series to series in other methods, the proposed method achieves the best performance with the same set of hyperparameters. Table 2Results of the analysis performed for the S&P 500 time series.SeriesStatisticsFTS-N78FTS-SMN80PFF79HO-FTS31The proposed methodSeries 1Mean ± SD31.06± 0.2934.40± 1.2433.44± 2.0137.1029.35± 0.22Min30.5430.6130.0728.73Max31.5836.1938.9129.787637237621Series 2Mean ± SD45.85± 0.6449.32± 3.9043.49± 0.5554.9042.41± 0.02Min43.7942.7242.3442.30Max47.3554.5644.3242.443642265381Series 3Mean ± SD36.70± 0.6839.50± 2.5737.34± 0.6441.2433.06± 1.00Min35.4037.9135.8030.30Max38.0845.3538.3033.8056522662101Series 4Mean ± SD43.39± 0.3143.39± 3.3349.70± 18.0646.9041.84± 0.02Min42.7741.4242.8441.82Max43.9351.92143.9341.9075106257131Series 5Mean ± SD15.18± 0.3320.89± 3.0217.69 ± 1.2317.6213.82± 0.16Min14.5615.5215.2913.66Max15.8723.3120.2614.4646772742121 Results of the analysis performed for the S&P 500 time series. According to the minimum statistics in four of the five time series analyzed, the recommended method is seen to be the most successful method. According to the maximum statistic, the recommended method appears to give the best results among all methods. When examining the standard deviation statistics in Table 2, the proposed method is the method with the lowest variability. The Friedman test was used for statistical comparison of the results obtained, and the Nemenyi test was used for pairwise comparison of the methods. In Friedman’s two-way analysis of variance, the p-value was 0.005, and it was concluded that there was a significant difference between the methods. Figure 5 shows the rank statistics. As can be seen, the method that ranks first is the recommended method. Bar charts of statistics obtained from the analysis of the S&P 500 time series. The results of the Nemenyi multiple comparison test, which compares the proposed method with other methods, are given in Table 3. Table 3Results of the analysis performed for the S&P 500 Time series.Sample 1–Sample 2Test statisticSigThe Proposed Method -FTS-N781.300.194The Proposed Method -PFF792.400.016The Proposed Method -FTS-SMN802.900.004The Proposed Method -HO-FTS313.400.001 Results of the analysis performed for the S&P 500 Time series. As shown in Table 3, the proposed method produced statistically better results than the PFF, FTS-SMN, and HOFTS methods. The difference between the results obtained with the FTS-N method in terms of prediction performance is insignificant. Subsequently, the proposed method, along with several alternative methods, was applied to the Dow-Jones 30 time series, with the outcomes summarized in Table 4. The content of Table 4 is prepared in a similar manner to Table 2. Table 4Results of the analysis performed for Dow-Jones 30 time series.SeriesStatisticsFTS-N78FTS-SMN80PFF79HO-FTS31The proposed methodSeries 1Mean ± SD170.62± 0.93186.24± 0.00235.46± 122.42174.25165.53± 7.27Min168.48186.24170.17162.27Max172.36186.24663.62195.392487212371Series 2Mean ± SD193.05± 2.72204.63± 0.00182.55± 13.15262.81187.71± 9.01Min186.02204.63168.41185.10Max199.09204.63237.44235.436383272171Series 3Mean ± SD257.33 ± 2.18307.36± 2.31302.74± 216.18233.76243.97 ± 35.37Min253.67300.31246.95209.53Max261.13317.631446.06427.9547942666121Series 4Mean ± SD276.14± 1.87285.00± 0.00284.32± 8.84358.86280.93± 1.05Min272.71285.00269.49277.94Max281.28285.00310.13285.202636272551Series 5Mean ± SD146.97± 2.9298.84± 0.00154.32± 52.60125.82151.94± 82.90Min140.7498.8492.88105.84Max154.3398.84415.82580.0764832621111 Results of the analysis performed for Dow-Jones 30 time series. The analysis results presented in Table 4 reveal that the proposed method ranks first once, second three times, and fourth once according to the average statistic. However, based on the minimum and maximum statistics, the proposed method is seen to be more successful. When the Friedman test is applied to the results obtained based on the average statistics, p>0.05, and no significant difference is found between the methods. According to Table 4, when examining the standard deviation results, it can be said that the variability of the proposed method may be higher than that of other methods, but it does not have a low change rate. In general, when the results in Table 4 are examined based on the average RMSE values, the fact that the proposed method ranks among the top two in four out of five series can be interpreted as evidence of the method’s success. When the rank statistics of the methods are examined in Fig. 6, it is clear that the proposed method performs best, as demonstrated by the results. Bar charts of statistics obtained from the analysis of Dow-Jones 30 time series. Conclusions and discussion Alternative forecasting methods to statistical forecasting methods are intensively discussed in the academic literature. In fuzzy set-based forecasting methods, a fully systematic approach is required for the selection of model inputs and hyperparameters. A novel prediction algorithm based on Pythagorean fuzzy sets is presented in this study. In the proposed Pythagorean FTS forecasting method, a new hybrid ANN used in modeling fuzzy relationships is presented. This neural network is a feedback and combined ANN that incorporates a simple exponential smoothing method, which is a classic forecasting method. One of the most important problems in fuzzy set-based prediction approaches is the use of large input data provided by membership values in modeling. In the proposed method, instead of using large data such as membership and non-membership values, principal component analysis is used to include smaller and unrelated data in the ANN. In the proposed method, a validity data set-based approach is used for hyperparameter selection, thereby eliminating the influence of hyperparameters in evaluating the method’s performance and ensuring that the method’s performance is presented in a fairer manner. In evaluating the performance of the proposed method, two distinct stock market datasets were employed. The rolling window approach was preferred to obtain different time series from stock market data. When the obtained results are examined, it is clear that the predicted performance of the proposed method is the best in each series compared to the compared methods for the S&P 500 series and has a very good performance for the Dow-Jones 30 series. Although statistically different prediction forms are obtained between the compared methods in the S&P 500 time series applications, it is evident that the performance differences in the Dow-Jones 30 series are statistically insignificant. It is observed that the proposed method produces promising results in both data set applications. It is clear that the input selection based on partial autocorrelations saves the researchers from a significant parameter selection problem and creates an effective input selection method. In future studies, it is planned to transform the proposed method into a robust method that can also work in the presence of outliers. In addition, it seems possible to develop the proposed method for high-order n-Pythagorean FTS (n > 2), which is another type of Pythagorean FTS and is another future research topic. Author contributions O.G. Methodology, Software, Conceptualization, Writing, Editing. E.E. Methodology, Software, Conceptualization, Writing, Data Analysis. E.B. Methodology, Software, Conceptualization, Data Analysis. Funding No Funding. Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.