A hybrid fuzzy–ensemble method for time series forecasting Authors: Ali Zafer Dalar Published: 2026 ABSTRACT This study proposes a rule-free hybrid forecasting framework that integrates type-1 fuzzy function modeling with ensemble regression via random forests. The proposed hybrid fuzzy–ensemble model first constructs lagged representations of a univariate time series and applies fuzzy c-means clustering to obtain soft partitions of the input space. For each cluster, a dedicated random forest regressor is trained, and final forecasts are produced through membership-weighted aggregation of cluster-level predictions. This architecture eliminates the need for expert-defined fuzzy rules while enabling the model to capture both local (regime-dependent) and global temporal dynamics. Key hyperparameters, including lag order, number of clusters, membership thresholds, and random forest settings, are jointly optimized using random search on the training data. The framework is evaluated on thirteen real-world time series from financial, industrial, and environmental domains. Experimental results demonstrate competitive and, in several cases, improved forecasting accuracy compared to classical statistical methods, fuzzy regression models, and plain random forest baselines. The findings suggest that combining soft fuzzy partitioning with cluster-specific ensemble learning provides a flexible and data-driven mechanism for modeling nonlinear structures and gradual regime-like behaviors within univariate time series.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-51585-w. Introduction Forecasting time series data continues to be a foundational task in numerous scientific and applied disciplines, ranging from finance and meteorology to energy systems and healthcare analytics. However, the inherent uncertainty, non-linearity, and potential non-stationarity of real-world time series often challenge the predictive power of classical statistical models. In recent years, the integration of machine learning and fuzzy inference systems has emerged as a promising avenue to overcome these limitations. Fuzzy systems have long been employed for time series forecasting due to their ability to handle uncertainty and linguistic ambiguity in data. However, conventional fuzzy inference systems (FISs) often rely on rule-based formulations, where defining and tuning fuzzy rules can be highly subjective, labor-intensive, and domain-dependent. Moreover, these systems may struggle when faced with complex, nonlinear, and high-dimensional datasets where interpretability and scalability must be balanced. Recent studies have attempted to alleviate these challenges by incorporating machine learning models such as neural networks or support vector machines (SVMs) into fuzzy frameworks. Nevertheless, such hybrid approaches still often suffer from issues related to overfitting, lack of interpretability, and the need for extensive hyperparameter tuning. In particular, defining the structure of fuzzy rule bases remains a persistent bottleneck in expanding fuzzy systems to real-world large-scale datasets. Fuzzy systems, particularly Type-1 Fuzzy Functions (T1FFs), offer a powerful framework for handling the imprecision and ambiguity often found in observational time series data. Unlike traditional rule-based fuzzy systems that rely on expert-defined logic rules1,2, T1FFs eliminate the need for a rule base by constructing fuzzy functions directly from data through regression-based learning3,4. This characteristic makes them especially suitable for large-scale or automated forecasting contexts. In parallel with the evolution of fuzzy systems, ensemble learning models have become increasingly prominent in handling high-dimensional and nonlinear time series. Random Forests (RFs), first introduced by Breiman5, remain one of the most influential ensemble learning methods, particularly known for their ability to model complex nonlinear relationships while reducing variance through bootstrap aggregation. Their robustness against noise, automatic feature selection, and reduced susceptibility to over-fitting make them highly desirable for forecasting applications where interpretability and performance must coexist. In its classical formulation, RFs construct an ensemble of decision trees over bootstrap samples and use feature randomness to enhance generalization. Owing to these properties, RFs have been widely adopted in fields ranging from financial forecasting to environmental modeling and bioinformatics6. Building on the complementary strengths of rule-free fuzzy functions and ensemble learning, this study introduces a hybrid model that eliminates rule dependency by integrating T1FFs with Random Forests—an ensemble method known for its robustness and minimal tuning requirements—hereafter referred to as the proposed hybrid fuzzy–ensemble model. Our method employs fuzzy C-means (FCMs) clustering to obtain soft partitions of the lagged input space that may correspond to regime-like behaviours, and then trains separate RFs models per cluster. Forecasts are aggregated through a fuzzy membership-weighted mechanism. In addition, a randomized hyper-parameter search framework is employed to tune both the fuzzy and ensemble learning components jointly. An analytical discussion of this mechanism is provided in the Analytical properties of the proposed model subsection. Real-world time series often exhibit a mixture of linear and nonlinear dynamics arising from structural breaks, regime changes, volatility clusters and interaction effects that cannot be captured by a single global linear model. Although classical T1FFs already eliminate the need for rule bases and provide a flexible soft-partitioning mechanism, each fuzzy region still relies on a linear regression surface. This assumption becomes restrictive for financial and industrial series, where local patterns are highly nonlinear and may vary substantially across regimes. Prior studies have attempted to address this limitation by replacing the linear consequents with neural networks or other nonlinear learners, but such models typically suffer from increased tuning complexity and reduced interpretability. Motivated by these gaps, the present study replaces the linear regressors of T1FFs with cluster-specific Random Forests, a nonparametric ensemble method capable of capturing complex nonlinear relations with minimal parameter tuning. Importantly, while Random Forests are not inherently interpretable at the individual tree level, the proposed architecture retains a form of structural interpretability through its fuzzy soft partitions. The fuzzy C-means memberships provide an explicit and transparent mechanism that governs how local models contribute to the final prediction, thereby preserving the rule-free organizational structure of T1FFs without relying on expert-defined if–then rules. Membership-weighted aggregation ensures smooth transitions between regimes, yielding a data-driven forecasting model that is both durable and capable of adapting to heterogeneous temporal structures. This study addresses three practical challenges in fuzzy time-series forecasting: (i) reliance on expert-defined rule bases that hampers scalability and portability across datasets, (ii) instability of single global models under regime shifts and local nonlinearities, and (iii) the difficulty of tuning interacting hyperparameters at scale. Our approach tackles (i) by adopting rule-free fuzzy functions, (ii) by learning cluster-specific regressors combined through membership-weighted aggregation, and (iii) by selecting all model parameters—including the number of clusters—via random search on the training data only. Conventional type-1 fuzzy rule–based forecasting systems rely on explicitly defined “if–then’’ rules whose antecedents describe regions of the input space and whose consequents are typically linear or affine functions. Such models require rule extraction, rule pruning, and rule-weight tuning, which introduce subjectivity and impose structural constraints on the final model. In contrast, the proposed framework is rule-free: fuzzy C-means is used solely to obtain soft partitions of the lagged input space, and each cluster is modeled by a data-driven random forest regressor rather than an explicit rule consequent. This eliminates the need for expert-defined rules, avoids manual rule design and optimization, and shifts interpretability to the memberships that govern how local models contribute to the final forecast. While the architecture is generic, the evidence presented in this paper is limited to offline univariate forecasting problems with moderate series length and real-valued responses. Therefore, claims of generalizability and scalability are restricted to time series with characteristics similar to the thirteen financial, industrial, and environmental datasets studied here. On this basis, we summarize the main methodological contributions as follows: We introduce a rule-free type-1 fuzzy–ensemble architecture in which fuzzy c-means is used solely to induce soft partitions of the lagged input space, while the local input–output relationships are learned nonparametrically via cluster-specific Random Forests. This differs from classical type-1 fuzzy functions and elastic-net FRFs, which directly parameterize the regression surfaces inside each fuzzy function.We propose a membership-weighted ensemble aggregation strategy that combines local forests according to their fuzzy memberships, yielding a smooth interpolation between regime-specific predictors without requiring explicit, expert-defined if–then rules.We design a joint randomized hyperparameter search that simultaneously tunes the number of clusters, the lag order, and key Random Forest parameters using the training data only, instead of fixing the fuzzy partition a priori as in most FRF-based approaches.We provide an empirical study on thirteen real-world univariate time series spanning financial, industrial, and environmental domains, together with ablation baselines (global RF, standard type-1 fuzzy functions approach and its variants, and RF with membership features) that help disentangle the contribution of each component. We introduce a rule-free type-1 fuzzy–ensemble architecture in which fuzzy c-means is used solely to induce soft partitions of the lagged input space, while the local input–output relationships are learned nonparametrically via cluster-specific Random Forests. This differs from classical type-1 fuzzy functions and elastic-net FRFs, which directly parameterize the regression surfaces inside each fuzzy function. We introduce a rule-free type-1 fuzzy–ensemble architecture in which fuzzy c-means is used solely to induce soft partitions of the lagged input space, while the local input–output relationships are learned nonparametrically via cluster-specific Random Forests. This differs from classical type-1 fuzzy functions and elastic-net FRFs, which directly parameterize the regression surfaces inside each fuzzy function. We propose a membership-weighted ensemble aggregation strategy that combines local forests according to their fuzzy memberships, yielding a smooth interpolation between regime-specific predictors without requiring explicit, expert-defined if–then rules. We propose a membership-weighted ensemble aggregation strategy that combines local forests according to their fuzzy memberships, yielding a smooth interpolation between regime-specific predictors without requiring explicit, expert-defined if–then rules. We design a joint randomized hyperparameter search that simultaneously tunes the number of clusters, the lag order, and key Random Forest parameters using the training data only, instead of fixing the fuzzy partition a priori as in most FRF-based approaches. We design a joint randomized hyperparameter search that simultaneously tunes the number of clusters, the lag order, and key Random Forest parameters using the training data only, instead of fixing the fuzzy partition a priori as in most FRF-based approaches. We provide an empirical study on thirteen real-world univariate time series spanning financial, industrial, and environmental domains, together with ablation baselines (global RF, standard type-1 fuzzy functions approach and its variants, and RF with membership features) that help disentangle the contribution of each component. We provide an empirical study on thirteen real-world univariate time series spanning financial, industrial, and environmental domains, together with ablation baselines (global RF, standard type-1 fuzzy functions approach and its variants, and RF with membership features) that help disentangle the contribution of each component. Taken together, these elements yield a practically effective hybrid fuzzy–ensemble forecaster that retains the interpretive structure of fuzzy partitions while leveraging the robustness of tree-based ensembles. The remainder of this paper is organized as follows: the next section reviews related work in fuzzy time series forecasting, followed by the methodology section introducing the proposed approach. The experimental setup, results, and discussion are then presented, and the paper concludes with final remarks and future research directions. Literature review Initial studies on fuzzy time series (FTS) models, such as those pioneered by7–9, primarily utilized type-1 fuzzy logic with predefined rules. Over the years, several extensions have been proposed, including high-order FTS, weighted FTS, and hybrid models that combine fuzzy logic with neural networks, ARIMA, or support vector machines. More recent advances include intuitionistic fuzzy time series models (e.g.10,11, type-2 fuzzy systems for uncertainty modeling12, and deep learning-fuzzy hybrids such as those incorporating LSTM13 architectures. While these approaches enhance forecasting flexibility, they often suffer from interpretability issues or require intensive parameter tuning. The fusion of fuzzy inference systems and machine learning models has catalyzed significant advancements in time series forecasting over the past two decades. This integration is particularly visible in the evolution of fuzzy regression functions (FRFs), which bypass traditional fuzzy rule bases by using data-driven regressors built over fuzzy clusters. Classical fuzzy rule-based systems, while intuitive, often rely on subjective rule extraction and lack scalability for large datasets. This led researchers to explore rule-free paradigms such as fuzzy functions, which bypass rule elicitation by directly modeling the input-output relationships through parametric or non-parametric functions. This historical evolution from rule-based systems to fuzzy functions is thoroughly reviewed by Turksen14, outlining the paradigm shift in fuzzy inference architectures. The foundational concept of fuzzy functions was pioneered by Turksen4 and further developed in subsequent works15. These methods construct fuzzy functions for each cluster, using membership values as input features, and have shown effectiveness across various forecasting tasks. Additional contributions expanding the application of fuzzy functions for time series forecasting include the works by Aladag et al.16 and Dalar et al.17, who demonstrated the effectiveness of T1FFs using covariate-based structures and benchmark datasets such as Australian beer production and electricity consumption. Extensions such as the Meta Fuzzy Functions (MFFs) and Meta Possibilistic Fuzzy Functions (MPFFs) proposed by Tak18,19 improved upon classical T1FFs by incorporating ensemble and possibilistic frameworks. Recent advancements in fuzzy clustering have extended the classical FCM algorithm used in T1FFs. For instance, Bas and Egrioglu20 proposed the use of the Gustafson–Kessel (GK) clustering algorithm to enhance the adaptability of fuzzy regression functions, especially in detecting ellipsoidal clusters. Similarly, Tak21 developed an autoregressive moving average type possibilistic fuzzy forecasting model (ARMA-PFFs), combining autoregressive structure with possibilistic clustering and metaheuristic optimization via the Grey Wolf Optimizer, offering robustness against noisy data and improved forecasting flexibility. Recent extensions have also incorporated metaheuristic optimization strategies. For example, Dalar22 employed particle swarm optimization (PSO) to fine-tune T1FF parameters for improved performance across multiple benchmark time series, including stock and index data. The robustness of FRFs in the presence of outliers has been a recurring concern. Several researchers have proposed robust versions of fuzzy regression functions by introducing noise clusters or applying robust loss functions. For instance, Chakravarty et al.23 introduced Modified Fuzzy Regression Functions with Noise Clusters (MFRFN), demonstrating enhanced accuracy in wind speed estimation tasks. Earlier work by the same authors24 benchmarked this approach against classical machine learning models and highlighted the performance gains under adversarial data conditions. Similarly, Bas25 and Egrioglu and Bas26 explored robust FRFs using robust estimation techniques such as Huber and Cauchy loss functions to mitigate the influence of outliers in both dependent and independent variables. From a modelling architecture standpoint, hybrid frameworks that integrate fuzzy systems with neural networks or support vector machines have also garnered attention. Aktoprak and Cagcag Yolcu27 proposed Fuzzy Regression Network Functions (FRNFs), where non-linear dependencies are modelled via multiple feedforward neural networks, each associated with a fuzzy cluster. Their architecture provides a flexible and interpretable alternative to black-box deep learning models. The FRF-SVM model proposed by Baser and Demirhan28 provided improvements in solar radiation estimation. Beyond neural and SVM-based hybrids, recent studies have explored recurrent and bootstrapped versions of fuzzy functions. Egrioglu et al.29 introduced recurrent fuzzy time series functions and demonstrated their success across energy and economic datasets. Tak30 incorporated recurrent intuitionistic fuzzy structures and used grey wolf optimization to improve performance under uncertainty. Recent advances have emphasized recurrent and elastic-net-based fuzzy systems (E-FRFs, Tak and Inan31, ridge-regularized variants to mitigate multicollinearity issues32, and hybrid architectures combining elastic-net fuzzy regression with deep models like long short-term memory (LSTM)33. Although the Tak and Inan31‘s E-FRFs model addresses overfitting via regularization and provides interpretability, it still operates on a global modeling structure, which may overlook localized data dynamics. In contrast, the proposed method leverages fuzzy clustering to decompose the series and learns localized models through random forests, thus capturing heterogeneous temporal behaviors more effectively. Gourdazi et al.34 introduced recurrent fuzzy functions structures within nonlinear autoregressive models with exogenous inputs (NARX) for modeling chaotic time series, demonstrating improved performance on systems with hidden attractors. In particular, Dalar and Egrioglu35,36 proposed hybrid frameworks incorporating subsampling bootstrap mechanisms for generating confidence intervals within fuzzy regression functions. This enhancement not only improves forecast reliability but also enables probabilistic scenario modeling—especially valuable in financial and high-volatility domains. Several researchers have focused on rule-based fuzzy systems, such as Takagi and Sugeno2 and Mamdani and Assilian37 models, for time series forecasting. While these systems provide interpretability, their reliance on expert-defined rules often limits scalability and generalizability across heterogeneous datasets. To mitigate this, fuzzy functions approaches—such as those proposed by Aladag et al.16 and Dalar et al.17—have emerged, offering rule-free architectures. However, these models may struggle to adapt to local fluctuations in high-variance series due to their single-model nature. In parallel, ensemble learning methods, particularly RFs, have gained popularity for capturing nonlinear dynamics and handling noise, but their integration with fuzzy reasoning remains underexplored. Previous hybrid methods (e.g., fuzzy clustering with neural networks or SVMs) have shown improvements but often require complex tuning or lack interpretability. These gaps underscore the need for a flexible yet interpretable hybrid approach that leverages the advantages of both fuzzy reasoning and ensemble learning. In the context of time series forecasting, RFs have gained significant traction due to their non-parametric, data-driven nature. Despite not being explicitly designed for fuzzy systems, their capability to handle non-linearities and interactions complements fuzzy methods well. Recent studies (e.g., Wei et al.38 have shown that combining RFs with fuzzy inference structures can yield highly competitive forecasting results, especially under uncertainty and heterogeneity. Beyond hybrid fuzzy-neural architectures, recent studies have increasingly adopted RFs as core regressors for time series prediction tasks. Owing to their inherent ensemble nature, RFs are particularly well-suited for handling nonlinearity, noise, and variable interactions in high-dimensional temporal data. For instance, Tihi and Popov39 conducted a comparative study of auto regressive integrated moving average (ARIMA) and RF models for urban drought prediction, concluding that although classical models performed well under specific assumptions, RFs offered competitive accuracy and greater adaptability to changing climate conditions. Similarly, Altincop and Oktay40 applied RF models for air quality forecasting in Turkish metropolitan regions, demonstrating that RFs outperformed artificial neural networks in predicting PM10 and CO concentrations across fluctuating meteorological conditions. These findings underscore the versatility of RFs in real-world forecasting tasks, especially when data exhibit structural heterogeneity and require robust nonparametric learning mechanisms. As such, the integration of RFs into fuzzy inference frameworks—as pursued in our study—aligns with this broader trend in the literature, combining the explainability of fuzzy partitioning with the predictive strength of tree-based ensembles. In addition to their general strengths in forecasting, recent studies have increasingly recognized RFs as effective components in fuzzy system architectures. In line with this trend, ensemble models—particularly RFs—have attracted attention as robust, nonparametric alternatives to conventional regressors in fuzzy systems. While most fuzzy forecasting systems rely on traditional statistical or neural components, integrating RFs introduces additional flexibility and interpretability. Additionally, algorithmic variants such as Extremely Randomized Trees41 and efficient implementations like ranger42 point to further opportunities for scaling fuzzy-ensemble hybrids in real-world settings. Our work situates itself at the intersection of these research streams. While previous studies have explored fuzzy–neural hybrids or fuzzy–SVM integrations, our approach distinguishes itself by embedding RF regression into the structure of Type-1 Fuzzy Functions, enabling a transparent, interpretable, and high-performing model architecture. Moreover, the use of a randomized hyper-parameter optimization strategy allows the model to dynamically adapt to the characteristics of each dataset, enhancing generalizability across domains. To summarize, while several fuzzy computing models have been proposed for time series forecasting, many suffer from either rule dependency, lack of flexibility, or limited model interpretability. This study aims to bridge these gaps by integrating fuzzy partitioning with ensemble regression to deliver a more adaptive and rule-free framework for time series forecasting. The proposed methodology This section introduces the mathematical and algorithmic formulation of the proposed hybrid fuzzy–ensemble model, designed for univariate time series forecasting. The methodology integrates soft partitioning through FCM clustering and regression modeling through RFs, yielding a hybrid architecture capable of capturing both local and global nonlinearities without requiring expert-defined fuzzy rules. The forecasting process proceeds through three primary stages: (i) fuzzy partitioning of the lagged time series using FCM, (ii) feature transformation and cluster-specific RF model training, and (iii) aggregation of predictions based on fuzzy memberships weights. Fuzzy partitioning via fuzzy c-means Let the input time series be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x=\left[{x}_{1},\:{x}_{2},\dots\:,{x}_{n}\right]\in\:{\mathbb{R}}^{n}$$\end{document}. To incorporate temporal dependencies, a lag matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{D}\in\:{\mathbb{R}}^{\left(n-p\right)\times\:(p+1)}$$\end{document} is constructed, where each row corresponds to a vector of lagged observations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{d}}_{t}=\left[{x}_{t},\:{x}_{t-1},\:\dots\:,\:{x}_{t-p}\right],\:\:\mathrm{f}\mathrm{o}\mathrm{r}\:t=p+1,\:\dots\:,\:n.$$\end{document}. Prior to model estimation, all input variables are standardized to zero mean and unit variance using statistics computed from the training set only, in order to prevent information leakage into the test data. The same transformation is subsequently applied to the test set. No missing observations are present in the datasets considered, and no imputation is required. The time series are modeled in their original scale (levels) as in the referenced studies from which the data were obtained. Given a predefined number of clusters, the FCM algorithm aims to minimize the following objective function43,44:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J\left( {U,V} \right) = \sum\limits_{{i = 1}}^{{n - p}} {\sum\limits_{{j = 1}}^{c} {u_{{ij}}^{m} } } \left\| {{\mathbf{d}}_{i} - {\mathbf{v}}_{j} } \right\|^{2} ,$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{ij}\in\:\left[0,\:1\right]$$\end{document} is the membership degree of observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{d}}_{i}$$\end{document} to cluster j, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{v}}_{j}$$\end{document} is the centroid of cluster j, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m>1$$\end{document} is the fuzziness parameter (commonly set to 2). Membership degrees are updated iteratively using:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_{{ij}} = \left( {\sum\limits_{{k = 1}}^{c} {\left( {\frac{{\left\| {{\mathbf{d}}_{i} - {\mathbf{v}}_{j} } \right\|}}{{\left\| {{\mathbf{d}}_{i} - {\mathbf{v}}_{k} } \right\|}}} \right)} ^{{2/\left( {m - 1} \right)}} } \right)^{{ - 1}} .$$\end{document} After convergence, a threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}\in\:\left[0,\:1\right]$$\end{document} is applied to the matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:U=\left[{u}_{ij}\right]$$\end{document} to prune low-membership associations and derive cluster-specific training subsets. From an analytical perspective, minimizing the objective \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:J(U,V)$$\end{document} in (1) yields a soft partition of the lagged input space in which observations with similar lag vectors are encouraged to share high memberships in the same clusters43,44. Because each row of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{D}$$\end{document} encodes a short history segment of the series, this partitioning can be interpreted as grouping time points that exhibit similar recent dynamics, without assuming explicit regime boundaries. The fuzziness parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m$$\end{document} controls the overlap between clusters: for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\to\:1$$\end{document} the partition approaches a hard assignment, whereas larger \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m$$\end{document} values distribute memberships more evenly across clusters. Applying the threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{cut}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:U$$\end{document} removes very small memberships while still allowing an observation to contribute to multiple clusters, so that each cluster-specific regressor is estimated from a subset of lag vectors that is relatively coherent in the Euclidean sense. These properties will be exploited in the subsequent stages to construct local models associated with different regions of the lag space and to combine them smoothly through membership-weighted aggregation. Feature construction and model training For each cluster \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j\in\:\left\{1,\dots\:,c\right\}$$\end{document}, a transformed feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{X}}^{\left(j\right)}\in\:{\mathbb{R}}^{{n}_{j}\times\:d}$$\end{document} is constructed using the subset of rows from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{D}$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{ij}>{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document}. The row vectors in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{X}}^{\left(j\right)}$$\end{document} are augmented feature vectors that combine fuzzy membership \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{j}$$\end{document} with lagged observations:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{X}}^{\left(j\right)}=\left[1,{\mu\:}_{j},\sqrt{{\mu\:}_{j}}\:,\:{{\mu\:}_{j}}^{2},\mathrm{e}\mathrm{x}\mathrm{p}\left({\mu\:}_{j}\right),\:{x}_{t-1},\dots\:,{x}_{t-p}\right],$$\end{document} where 1 is the intercept vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{j}={u}_{ij}$$\end{document} for selected samples, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d=p+5$$\end{document} is the total number of features (including one intercept term and three nonlinear transformations—square root, quadratic and exponential—of the membership degree), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{j}$$\end{document} is the number of samples in cluster j after thresholding. These nonlinear transformations of the membership degree follow the classical type-1 fuzzy functions literature and are retained here to ensure comparability with existing FRF-based models, while shifting the expressive burden from the fuzzy layer to the nonparametric RF regressors. Let the corresponding target vector be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{Y}}^{\left(j\right)}=\left[{x}_{t}\right]\in\:{\mathbb{R}}^{{n}_{j}}$$\end{document}. Each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{X}}^{\left(j\right)}$$\end{document} is used to train a cluster-specific RF regression model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{F}}^{\left(j\right)}$$\end{document}, which is an ensemble of N regression trees \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{h}_{1}^{\left(j\right)},\:{h}_{2}^{\left(j\right)},\:\dots\:,\:{h}_{N}^{\left(j\right)}\right\}$$\end{document}. For each tree, splits are chosen to minimize the mean squared error (MSE) across nodes:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{MSE}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}=\sum\limits_{i\in\:L}{\left({y}_{i}-{\stackrel{-}{y}}_{L}\right)}^{2}+\sum\limits_{i\in\:R}{\left({y}_{i}-{\stackrel{-}{y}}_{R}\right)}^{2},$$\end{document} where L and R denote the left and right child nodes of potential split, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{-}{y}}_{L}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{-}{y}}_{R}$$\end{document} are the mean target values in these nodes. The prediction for an input vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{d}}_{t}$$\end{document} by ensemble \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{F}}^{\left(j\right)}$$\end{document} is computed by averaging tree predictions:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{t}^{\left(j\right)}=\frac{1}{N}\sum\:_{n=1}^{N}{h}_{n}^{\left(j\right)}\left({\mathbf{d}}_{t}\right),$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{n}^{\left(j\right)}$$\end{document} is the prediction of the n-th tree in cluster j. The number of trees N, maximum tree depth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}, minimum samples to split a node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document}, minimum samples per leaf \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document}, and bootstrap sampling flag (True/False) are selected via a randomized hyperparameter search detailed in the hyperparameter optimization subsection. The parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{j}$$\end{document} denotes the number of training samples associated with cluster j after applying the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} threshold to the fuzzy memberships. While the RF training itself follows the standard regression formulation, its role in the proposed framework is not to introduce a new tree-based learner, but to act as a variance-reducing nonlinear approximator within each fuzzy region. This architectural placement is critical, as it decouples regime identification (handled by FCM) from local nonlinear approximation (handled by RF), which is not present in classical type-1 fuzzy functions. A schematic illustration of the RF regression process is presented in Fig. 1, where the aggregation of predictions from multiple decision trees is demonstrated. Fig. 1Illustration of the prediction mechanism in RFs, where individual tree outputs are aggregated to form a final ensemble prediction. Illustration of the prediction mechanism in RFs, where individual tree outputs are aggregated to form a final ensemble prediction. Figure 1 illustrates the internal aggregation mechanism of a RF within a single fuzzy region, whereas Fig. 2 summarizes how these cluster-specific ensembles are embedded into the overall fuzzy–ensemble forecasting pipeline. Random forests were chosen for the cluster-level regressors because they require substantially less hyperparameter tuning than boosting algorithms and are known to be more durable to noisy or short time series due to their parallel, non-iterative structure5,45. In contrast, gradient boosting methods iteratively fit residual errors, which often increases the risk of overfitting within small cluster-specific subsets. RFs also provide smoother response surfaces and more stable behaviour in repeated training across different random seeds, making them suitable for repeated training inside each fuzzy region. Although the experiments in this study focus on univariate time series, the structure of the proposed framework naturally accommodates higher-dimensional inputs. Random forests mitigate the effects of high dimensionality through the random subspace mechanism, in which each tree is trained on a randomly selected subset of features46. This procedure reduces the effective dimensionality seen by individual trees and is known to improve learning stability in high-dimensional settings5. Unlike boosting methods, which may propagate noise across iterative residual fits in large feature spaces, RFs maintain decorrelation between trees and avoid aggressive feature–residual interactions45. Consequently, if multivariate extensions are explored in future work, the cluster-specific RFs would remain computationally feasible and would not require explicit feature-selection heuristics at each fuzzy region. Fig. 2Workflow of the proposed hybrid fuzzy–ensemble model for univariate time series forecasting. The pipeline consists of: (1) lagged feature construction, (2) fuzzy C-means soft partitioning of the lag space, (3) cluster-specific random forest training within each fuzzy region, and (4) membership-weighted aggregation to produce the final one-step-ahead forecast. Workflow of the proposed hybrid fuzzy–ensemble model for univariate time series forecasting. The pipeline consists of: (1) lagged feature construction, (2) fuzzy C-means soft partitioning of the lag space, (3) cluster-specific random forest training within each fuzzy region, and (4) membership-weighted aggregation to produce the final one-step-ahead forecast. As shown in Fig. 2, the process integrates both data-driven partitioning and ensemble learning. Combined with fuzzy soft partitions, this design preserves a level of structural interpretability inherited from T1FFs at the partition level—where fuzzy memberships indicate regime affiliation—while allowing each cluster to learn its own nonlinear local dynamics through data-driven ensemble models. The subsequent section outlines how cluster-level predictions are aggregated into a final forecast. In the proposed model, each fuzzy cluster is equipped with a RF regressor. RFs construct multiple decision trees trained on randomly selected subsets of features, following the random subspace principle introduced by Ho46. This mechanism produces structural diversity among trees even when bootstrap sampling is disabled, as is the case in all optimal configurations identified by the random search procedure (see Table 6). The final prediction is obtained by averaging the outputs of the trees5, which is known to reduce the instability associated with individual decision trees and yields smoother aggregated responses. In this study, RF therefore contributes a flexible nonlinear approximation component at the cluster level without relying on sample-level resampling. Aggregated prediction via membership weighting The final forecast for time point t, denoted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{t}$$\end{document}, is obtained by aggregating the outputs of all cluster models using fuzzy memberships as weights:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{t}=\frac{{\sum\:}_{j=1}^{c}{\mu\:}_{j}\left({\mathbf{d}}_{t}\right)\times\:{\widehat{y}}_{t}^{\left(j\right)}}{{\sum\:}_{j=1}^{c}{\mu\:}_{j}\left({\mathbf{d}}_{t}\right)}.$$\end{document} This aggregation enables the model to blend the predictions from multiple local models according to the degree of membership, allowing for smooth transitions across structural regimes in the time series. Analytically, (6) defines a convex combination of the cluster-level predictions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{t}^{\left(j\right)}$$\end{document} whenever the memberships satisfy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{j}\left({\mathbf{d}}_{t}\right)\ge\:0$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{j=1}^{c}{\mu\:}_{j}\left({\mathbf{d}}_{t}\right)>0$$\end{document}. Under these conditions, the aggregated forecast \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{t}$$\end{document} lies in the range spanned by the local forecasts, preventing extreme values that are not supported by any cluster model. Moreover, because the memberships vary smoothly as a function of the lag vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{d}}_{t}$$\end{document} the mapping \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{d}}_{t}\mapsto\:{\widehat{y}}_{t}$$\end{document} is piecewise smooth, with gradual transitions between regimes instead of abrupt switches. This fuzzy interpolation mechanism is a key difference from hard-assignment schemes, where the prediction would jump from one local model to another when the winning cluster changes. Hyperparameter optimization The proposed model involves several hyperparameters that significantly affect its forecasting performance: p: number of lagged inputs,c: number of fuzzy clusters,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document}: membership threshold for partial cluster participation,N: number of trees in each RF,RF-specific parameters: N: number of trees (10–200), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}: maximum tree depth (2–20), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document}: minimum samples to split (2–10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document}: minimum samples per leaf (1–10), bootstrap: sampling method (True/False). p: number of lagged inputs, p: number of lagged inputs, c: number of fuzzy clusters, c: number of fuzzy clusters, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document}: membership threshold for partial cluster participation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document}: membership threshold for partial cluster participation, N: number of trees in each RF, N: number of trees in each RF, RF-specific parameters: N: number of trees (10–200), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}: maximum tree depth (2–20), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document}: minimum samples to split (2–10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document}: minimum samples per leaf (1–10), bootstrap: sampling method (True/False). RF-specific parameters: N: number of trees (10–200), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}: maximum tree depth (2–20), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document}: minimum samples to split (2–10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document}: minimum samples per leaf (1–10), bootstrap: sampling method (True/False). N: number of trees (10–200), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}: maximum tree depth (2–20), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document}: minimum samples to split (2–10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document}: minimum samples per leaf (1–10), bootstrap: sampling method (True/False). N: number of trees (10–200), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}: maximum tree depth (2–20), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document}: minimum samples to split (2–10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document}: minimum samples per leaf (1–10), bootstrap: sampling method (True/False). To identify optimal settings, a random search strategy47 is employed over the parameter space using a specified number of iterations. This allows adaptive tuning without exhaustive grid search, balancing efficiency and performance. The number of fuzzy clusters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} is not predetermined; instead, it is treated as a hyperparameter and optimized jointly with the remaining parameters within the random search procedure. The search range \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c\in\:\{\mathrm{2,3},\dots\:,7\}$$\end{document} was chosen to avoid both under-partitioning—where very small \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} values merge distinct temporal patterns into a single region—and excessive fragmentation, where large \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} values can lead to unstable memberships and insufficient data within clusters. As illustrated in the empirical results presented in the Results and discussion section, extreme values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} (i.e., close to the boundaries of the search range) tend to yield inferior performance, whereas moderate values produce stable and consistent forecasting accuracy across datasets. This indicates that the random-search mechanism effectively mitigates sensitivity to the choice of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} by adaptively selecting values that balance cluster coherence and sample availability. In contrast to grid search—which evaluates all parameter combinations and scales poorly with dimensionality—random search explores the hyperparameter space more efficiently by sampling from predefined ranges. As shown by Bergstra and Bengio47, random search allocates trials to the most influential parameters and often reaches near-optimal configurations with typically fewer evaluations. Within the proposed framework, this leads to improved scalability because both fuzzy and RF components contain multiple interacting hyperparameters. Moreover, since each sampled configuration induces a different fuzzy partition and a different set of local forests, variability across trials implicitly assesses the stability of the model structure. This mechanism reduces the risk of over-fitting to a particular cluster configuration and enables the procedure to adapt to heterogeneous datasets without exhaustive exploration. The empirical results in the Results and discussion section illustrate that random search consistently identified moderate, stable cluster numbers and tree settings, supporting its suitability for the proposed hybrid forecaster. Table 1 presents the hyperparameter search spaces used for the RF-based fuzzy forecasting model across all investigated datasets. Each time series was subjected to a random search process, where key parameters—including the number of lagged features , number of clusters , membership threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document}, and RF settings—were independently optimized. The ranges reflect the variability considered during the optimization stage, ensuring that the proposed model adapts effectively to the unique characteristics of each dataset. Table 1Parameter ranges used in the randomized hyperparameter search for each dataset.Series n \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}}$$\end{document} p c \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} N \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document} BootstrapBIST100/2009103151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2010104151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2011106151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2012106151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2013106151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/1999266451–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2000271471–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2001244431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2002248431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2003249431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2004250431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseABP154161–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTemp. of Giresun120141–152–70, 0.1, 0.210–200None + 2–202–101–10True, False Table 1Parameter ranges used in the randomized hyperparameter search for each dataset.Series n \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}}$$\end{document} p c \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} N \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document} BootstrapBIST100/2009103151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2010104151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2011106151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2012106151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseBIST100/2013106151–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/1999266451–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2000271471–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2001244431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2002248431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2003249431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTAIEX/2004250431–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseABP154161–152–70, 0.1, 0.210–200None + 2–202–101–10True, FalseTemp. of Giresun120141–152–70, 0.1, 0.210–200None + 2–202–101–10True, False Parameter ranges used in the randomized hyperparameter search for each dataset. As detailed in Table 1, the parameter ranges were selected to balance computational tractability and model flexibility. Subsequent analyses presented in the Results and discussion section evaluate the performance of these configurations in diverse forecasting scenarios. Analytical properties of the proposed model The proposed hybrid forecaster can be viewed as a three–stage mapping in which lagged vectors are first constructed, then softly partitioned by fuzzy c-means, and finally passed through a set of cluster-specific Random Forest regressors whose outputs are combined by membership-weighted averaging. This section summarizes several analytical properties of this architecture, clarifying why it is well suited to heterogeneous, nonlinear time series without relying on explicit fuzzy rules. First, the FCM stage induces an overlapping partition of the lagged input space that minimizes the within-cluster dispersion in the Euclidean sense43,44. Because each row of the lag matrix encodes a short history window of the series, samples with similar recent trajectories tend to receive similar membership profiles. The resulting membership matrix therefore acts as a data-driven representation of local temporal regimes: different clusters emphasize different regions of the lag space, while the fuzziness parameter and the threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{cut}$$\end{document} jointly control the degree of overlap between these regions. Unlike hard clustering, this soft assignment allows each time point to influence multiple local models, which is important when regime boundaries are gradual rather than abrupt. Second, within each cluster the mapping from lag vectors to one-step-ahead responses is learned nonparametrically by a Random Forest regressor. RFs build an ensemble of decision trees on randomly selected subsets of features and average their predictions5,46. This random subspace mechanism creates diversity among trees even when bootstrap resampling is disabled, as in the best-performing configurations found by the random search (Table 6). Training a separate forest on each cluster-specific subset al.lows the local models to adapt to distinct nonlinear patterns that may appear in different regions of the lag space, while still preserving the relatively low tuning effort and piecewise-constant approximation behaviour characteristic of tree ensembles. Third, the final forecast at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} is obtained by aggregating the cluster-level predictions through the fuzzy memberships, as in (6). Provided that the memberships are non-negative and do not all vanish simultaneously, this yields a convex combination of the local forecasts. Consequently, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{t}$$\end{document} remains within the range spanned by the cluster-level outputs and changes smoothly as the membership vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:\left({\mathbf{d}}_{t}\right)$$\end{document} varies. This property distinguishes the proposed scheme from winner-takes-all or crisp switching models, where small perturbations in the input can trigger discontinuous jumps between regimes. These analytical observations do not constitute formal consistency proofs or generalization guarantees; rather, they explain how the individual components—soft clustering, cluster-wise tree ensembles, and membership-weighted aggregation—interact to produce a flexible yet structured forecasting mechanism. The empirical results in the Results and discussion section illustrate that, under the experimental conditions considered in this study, this architecture can effectively accommodate regime changes and local nonlinearities without relying on hand-crafted fuzzy rules. At the same time, the framework has clear limitations. The quality of the forecasts critically depends on the lagged representation and the FCM partition: if relevant lags are omitted or the series exhibits strongly non-spherical regimes, the induced clusters may not align with meaningful dynamics. Moreover, the method is essentially non-extrapolative: Random Forests, being tree-based, cannot reliably predict outside the range of training responses. Finally, the approach is designed for offline univariate forecasting; extending it to high-dimensional multivariate streams would require additional mechanisms for feature selection and concept drift adaptation, which are left for future work. Across the financial, industrial, and environmental datasets considered in this study, no domain-specific failure modes were observed. Instead, performance variations were closely aligned with structural characteristics of the individual series. Highly volatile and regime-shifting financial indices (e.g., TAIEX 2000–2002, BIST 2011) reduced the stability of fuzzy partitions and made local RF regressors more sensitive to abrupt transitions. In contrast, smoother and more regularly evolving series—such as the seasonal Australian beer production data and the low-amplitude Giresun temperature records—yielded more coherent cluster structures and more stable local models. These results indicate that the proposed framework is broadly domain-agnostic; its limitations arise primarily from volatility, irregular regime changes, and weak local stationarity rather than from the application domain itself. From an order-of-magnitude perspective, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n$$\end{document} denote the number of training observations, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p$$\end{document} the input dimensionality (number of lags and nonlinear transforms), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} the number of clusters, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I$$\end{document} the number of FCM iterations, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:T$$\end{document} the number of trees in each forest, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m$$\end{document} the number of candidate features considered at each split, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d$$\end{document} the average tree depth. Constructing the lag matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D$$\end{document} costs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(np\right)$$\end{document}. Running fuzzy c-means once for a given \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D$$\end{document} has time complexity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(ncp\cdot\:I)$$\end{document}. Training a Random Forest with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:T$$\end{document} trees on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n$$\end{document} samples is approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(Tnlogn\cdot\:m)$$\end{document}. In the proposed pipeline, after a single FCM fit per candidate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document}, we train \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} cluster-specific forests on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left\{{n}_{k}\right\}}_{k=1}^{c}$$\end{document} samples with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{k}{n}_{k}=n$$\end{document}, leading to total training cost \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(n\:c\:p\cdot\:I+\sum\:_{k=1}^{c}T\:{n}_{k}\mathrm{log}{n}_{k}\cdot\:m\right)$$\end{document}. Memory complexity is dominated by storing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} forests with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:T$$\end{document} trees each, yielding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {c~\cdot~T~\cdot~number~of~nodes} \right),$$\end{document}, which remains manageable for the dataset sizes considered in this study.” Compared with classical type-1 fuzzy regression models—whose training cost is typically dominated by solving a single global linear or affine regression per rule—the proposed framework incurs additional computational cost due to the training of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} cluster-wise Random Forests. However, because RFs require no iterative parameter optimization and can be trained in parallel, the overall complexity remains tractable for the moderate-sized univariate datasets considered in this study. Results and discussion In this section, the empirical performance of the proposed model is evaluated through experiments on multiple real-world time series datasets drawn from finance, manufacturing, and climate domains. The evaluation focuses on forecasting accuracy, model adaptability, and parameter behaviour across structurally diverse temporal datasets. Experimental configuration and datasets To validate the proposed methodology, thirteen time series datasets were utilized. The financial datasets include daily observations from the Borsa Istanbul Stock Exchange (BIST100) across five different years between 2009 and 201348, where the first five months of each year were used for training and the final 15 observations were held out for testing. Similarly, six datasets from the Taiwan Stock Exchange (TAIEX) were selected, covering daily measurements from 1999 to 200449. For each TAIEX series, the last two months of data were used as the test segment. Beyond financial series, the model was also applied to an industrial dataset: the Australian beer production data (ABP50, , measured quarterly between March 1956 and June 1994. This series represents a well-known benchmark in forecasting literature due to its clear seasonality and gradual trend structure. The final 16 quarterly observations were excluded from training and used for forecasting evaluation. Lastly, the methodology was tested on a meteorological dataset consisting of daily average temperatures observed in Giresun Province (Türkiye) between January and April 2017, a period marked by abrupt short-term temperature variations. All experiments were conducted under a one-step-ahead forecasting horizon (h = 1). Multi-step forecasts were generated iteratively by recursively feeding each predicted value as an input for the subsequent time step. No direct multi-horizon models were trained in this study. This setup ensures consistency across datasets and facilitates fair comparison with existing univariate forecasting benchmarks. To visually illustrate the temporal structures and variability across the selected datasets, Fig. 3 presents a sample of representative time series plots. Fig. 3Representative time series used in the forecasting experiments: BIST100 (2010), Australian Beer Production (1956–1994), Daily Average Temperature in Giresun (Jan–Apr 2017), and TAIEX Index (2002). Representative time series used in the forecasting experiments: BIST100 (2010), Australian Beer Production (1956–1994), Daily Average Temperature in Giresun (Jan–Apr 2017), and TAIEX Index (2002). For each dataset, the model was trained using varying lag orders (from 1 to 15), number of fuzzy clusters (between 2 and 7), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} membership thresholds (ranging from 0.0 to 0.2). Within each fuzzy region, RF regressors were constructed using randomized values for hyperparameters such as the number of estimators (between 10 and 200), tree depth (2 to 20), minimum sample split and leaf sizes, and bootstrap usage. A randomized hyperparameter search strategy was executed with 100 iterations to identify the best model configuration per dataset. Computationally, one FCM fit on n samples with input dimension p and c clusters costs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(n\:c\:p\cdot\:I)$$\end{document}, while a Random Forest with T trees costs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(T\:n\:\mathrm{log}n\cdot\:m)$$\end{document}. In our pipeline we run FCM once per candidate c and then train c cluster-specific forests on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left\{{n}_{k}\right\}}_{k=1}^{c}\:({\sum\:}_{k}{n}_{k}=n)$$\end{document}, yielding a total training cost \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(n\:c\:p\cdot\:I+\sum\:_{k=1}^{c}T\:{n}_{k}\mathrm{log}{n}_{k}\cdot\:m\right)$$\end{document} if clusters are balanced. At inference, computing memberships and traversing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:c$$\end{document} forests costs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(c\hspace{0.17em}p\:+\:c\hspace{0.17em}T\hspace{0.17em}d\right)$$\end{document} per time point. Because membership thresholding (αcut) reduces effective nk in practice, the sum \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{k}{n}_{k}\mathrm{log}{n}_{k}$$\end{document} often grows sublinearly with c. The cluster-wise structure is also amenable to parallelization across k=1, …, c but no wall-clock times are reported here since our focus is asymptotic complexity and predictive accuracy. We adopt randomized hyperparameter search since, in high-dimensional and non–axis-aligned spaces like (c, p , αcut and RF params), uniform random sampling explores more distinct configurations than grid search at the same budget and is supported by strong empirical/theoretical evidence51. Performance was assessed on test segments using standard forecasting metrics: root mean square error (RMSE) and mean absolute percentage error (MAPE). Each forecasting experiment was independently seeded to preserve stochastic reproducibility. Detailed descriptions of each dataset, including sampling frequency, forecast horizon, and original data characteristics, are available in the referenced studies from which the datasets were obtained. In this work, all series are used in their original univariate form without additional domain-specific preprocessing, ensuring consistency and fair comparability with prior benchmark evaluations. All three ablation baselines (Plain RF, Standard Type-1 Fuzzy Functions, and Global RF with membership features) were fully re-implemented by the author(s) and evaluated under identical preprocessing, lag construction, hyperparameter search protocol, and chronological train–test splits as the proposed model. Other comparison results cited from the literature are explicitly indicated in the corresponding tables. All preprocessing parameters were estimated exclusively on the training data to prevent information leakage into the test set. All implementations were conducted using Python 3.10.6. The random forest models were constructed via the Random Forest Regressor class from the scikit-learn library52, a widely adopted machine learning toolkit. Apart from this external module, all algorithmic components—including fuzzy partitioning, membership-based transformations, and model orchestration—were custom-developed by the author to enable full control over the proposed methodology and its evaluation pipeline. Forecasting results and observations The results across all datasets consistently reflect the proposed model’s ability to adapt to complex time series dynamics through soft clustering and ensemble-based learning. To assess the forecasting performance of the proposed model, we first applied it to five distinct subsets of the BIST100 index spanning from 2009 to 2013. Table 2 presents the MAPE results across these periods in comparison with classical and hybrid forecasting benchmarks. The effectiveness of the proposed method for the BIST100 datasets is first evaluated against classical and fuzzy-based forecasting models. The benchmark methods used for comparison were selected based on their relevance in the literature, including probabilistic methods53,54, non-linear machine learning models55, and advanced fuzzy-based forecasting models4,30,32,56. These methods have been discussed and applied in recent studies, particularly in the elastic net-based fuzzy regression framework by Tak and Inan31. Baseline selection follows the conventions of each benchmark family. For BIST100 (2009–2013), we report classical statistical (ARIMA/ES), neural (MLP-ANN) and fuzzy functions baselines commonly used in the Turkish market literature (Table 2). For TAIEX (1999–2004), we adopt the widely used set of fuzzy and neural models reported in prior work on this benchmark (Table 3). For the ABP industrial series, we include seasonal fuzzy functions baselines standard in this domain (Table 4), and for the Giresun temperature series we mirror the deep intuitionistic fuzzy LSTM comparison from Kocak et al.13 to ensure methodological consistency (Table 5). This yields like-for-like comparisons within each dataset family. In addition to classical statistical, neural, and fuzzy benchmarks, we also report results for two tree-based baselines to contextualize the contribution of the proposed fuzzy–ensemble coupling. Specifically, a global RF (Global RF) trained on the full lagged input space without fuzzy partitioning is included to assess the benefit of cluster-wise modeling. Moreover, we report results obtained using a standard RF regressor implemented via the scikit-learn library (Pedregosa et al.52, which corresponds to a plain RF model without fuzzy memberships. These baselines help disentangle the effect of fuzzy soft partitioning from that of ensemble learning alone. Table 2MAPE for the BIST100 series (2009–2013) across different methods.BIST100Box et al.53Brown54Wilamowski et al.55Turksen4Global RFPedregosa et al.52Bas et al.32Tak30Proposed2009/150.01200.01200.01140.01220.03330.03350.01120.0108 0.0096* 2010/150.02200.02200.02200.02640.02130.02230.02130.0200 0.0144* 2011/150.01500.01500.01460.01560.01260.01450.01430.0135 0.0115* 2012/150.00880.00880.01170.01610.01420.01330.01230.0116 0.0075* 2013/150.01090.01090.01070.01080.01350.01340.01030.0102 0.0088* Mean0.01370.01370.01410.01620.01900.01940.01390.0132 0.0104* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). Table 2MAPE for the BIST100 series (2009–2013) across different methods.BIST100Box et al.53Brown54Wilamowski et al.55Turksen4Global RFPedregosa et al.52Bas et al.32Tak30Proposed2009/150.01200.01200.01140.01220.03330.03350.01120.0108 0.0096* 2010/150.02200.02200.02200.02640.02130.02230.02130.0200 0.0144* 2011/150.01500.01500.01460.01560.01260.01450.01430.0135 0.0115* 2012/150.00880.00880.01170.01610.01420.01330.01230.0116 0.0075* 2013/150.01090.01090.01070.01080.01350.01340.01030.0102 0.0088* Mean0.01370.01370.01410.01620.01900.01940.01390.0132 0.0104* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). MAPE for the BIST100 series (2009–2013) across different methods. The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). On the BIST100 datasets, the model achieved the lowest MAPE among the compared methods in all five annual evaluations, with the 2010 and 2012 series showing particularly improved accuracy compared to benchmark models such as ARIMA, T1FFs, and ridge-regression-based fuzzy models. The relatively weaker performance of the Global RF baseline suggests that a single global nonlinear model is insufficient to capture regime-dependent dynamics in highly volatile financial series. Similarly, the Pedregosa et al.52 reference highlights that performance gains are not solely attributable to the RF learner itself, but rather to its integration within the proposed fuzzy partitioning and aggregation framework. The capability to account for local nonlinearities using cluster-specific regressors allowed the model to outperform even state-of-the-art fuzzy regression methods. Note that our goal is not to evaluate cluster validity per se; c is selected as part of the forecasting model to optimize out-of-sample accuracy (Table 1 provides the search ranges). Similar to the evaluation on the BIST100 index, the forecasting results on the TAIEX datasets are compared with a set of established benchmark methods. These comparison models, including both classical and fuzzy-based techniques, were selected from recent publications to ensure methodological relevance. Specifically, the reference values for all competing models in Table 3 were sourced from Tak and Inan31, who applied an elastic net-regularized fuzzy regression framework to the same dataset configuration. Table 3RMSE values for the TAIEX dataset (1999–2004) comparing the proposed model with conventional and fuzzy-based forecasting methods.TAIEX199920002001200220032004MeanTurksen4115.76178.30254.6684.4654.9684.09128.71Pedregosa et al.52125.96316.77119.7967.2260.2366.76126.12Global RF115.33336.03117.2066.4361.4465.52126.99Chen57120.00176.32147.84101.1874.4684.28117.35Wright and Ziegler42101.82128.95110.6660.4150.6552.8684.23Tak3097.81122.23106.8164.2451.5052.7982.56Yu and Huarng58112.00131.00130.0080.0058.0067.0096.33Chen and Phuong5999.97126.59110.1761.6253.0153.2884.11Tak et al.6098.33128.18106.4865.1452.3853.7884.05Cheng et al.61100.74125.62113.0462.9451.4654.2584.68Ye et al.62101.29125.42113.2263.9952.9952.4084.89Bas et al.3299.12119.73113.1762.5548.7351.6682.49Tak and Inan3196.01118.14103.6163.9350.8851.6080.69Proposed 93.85* 107.14* 97.38* 56.54* 43.44* 47.58* 74.32* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). Table 3RMSE values for the TAIEX dataset (1999–2004) comparing the proposed model with conventional and fuzzy-based forecasting methods.TAIEX199920002001200220032004MeanTurksen4115.76178.30254.6684.4654.9684.09128.71Pedregosa et al.52125.96316.77119.7967.2260.2366.76126.12Global RF115.33336.03117.2066.4361.4465.52126.99Chen57120.00176.32147.84101.1874.4684.28117.35Wright and Ziegler42101.82128.95110.6660.4150.6552.8684.23Tak3097.81122.23106.8164.2451.5052.7982.56Yu and Huarng58112.00131.00130.0080.0058.0067.0096.33Chen and Phuong5999.97126.59110.1761.6253.0153.2884.11Tak et al.6098.33128.18106.4865.1452.3853.7884.05Cheng et al.61100.74125.62113.0462.9451.4654.2584.68Ye et al.62101.29125.42113.2263.9952.9952.4084.89Bas et al.3299.12119.73113.1762.5548.7351.6682.49Tak and Inan3196.01118.14103.6163.9350.8851.6080.69Proposed 93.85* 107.14* 97.38* 56.54* 43.44* 47.58* 74.32* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). RMSE values for the TAIEX dataset (1999–2004) comparing the proposed model with conventional and fuzzy-based forecasting methods. The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). In the TAIEX datasets, where structural volatility and regime shifts are pronounced, the model exhibited stable empirical performance, often producing lower RMSE values than previously published fuzzy and neural forecasting models. In addition to fuzzy and neural benchmarks, we also report results for a plain global random forest (Global RF) and a standard RF regressor implemented via the scikit-learn framework52. While these models benefit from ensemble averaging, their performance remains consistently inferior to the proposed approach across all years. This highlights that the performance gains cannot be attributed solely to the use of RFs, but rather to their integration within a fuzzy partitioned structure. In highly volatile regimes such as TAIEX, a single global RF struggles to adapt to regime-dependent dynamics, whereas cluster-specific forests combined through membership-weighted aggregation provide a more stable and responsive forecasting mechanism. A comparable evaluation was conducted on the ABP dataset, a standard benchmark used in fuzzy forecasting literature due to its seasonal dynamics and industrial relevance. Forecasting results for this dataset, shown in Table 4, are aligned with prior literature, particularly the study by Bas and Egrioglu20, providing a consistent basis for comparison. Table 4Forecasting results for the ABP time series: comparison of actual and predicted values across alternative forecasting models.MethodsAladag63Yolcu et al.64Turksen4Bas et al.32Tak and Inan31Bas and Egrioglu20Global RFPedregosa et al.52ProposedRMSE29.140018.790017.390017.080016.823014.978022.877323.5068 11.4604* MAPE0.05000.04000.03000.03000.03200.02600.04280.0450 0.0192* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). Table 4Forecasting results for the ABP time series: comparison of actual and predicted values across alternative forecasting models.MethodsAladag63Yolcu et al.64Turksen4Bas et al.32Tak and Inan31Bas and Egrioglu20Global RFPedregosa et al.52ProposedRMSE29.140018.790017.390017.080016.823014.978022.877323.5068 11.4604* MAPE0.05000.04000.03000.03000.03200.02600.04280.0450 0.0192* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). Forecasting results for the ABP time series: comparison of actual and predicted values across alternative forecasting models. The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). The proposed method outperforms all compared approaches, achieving the lowest RMSE and MAPE values, as seen in the final rows of Table 4. The ABP dataset, often referred to in econometric literature as a seasonal benchmark, revealed another important characteristic of the proposed approach: its ability to handle both seasonal regularity and long-term trend shifts. Both the Global RF and the standard RF baseline52 underperform classical seasonal fuzzy models and the proposed method. This suggests that, when applied globally, tree-based ensembles may fail to exploit repetitive seasonal structures effectively. In contrast, the proposed fuzzy-partitioned regression forest framework benefits from cluster-wise specialization, allowing different forests to focus on distinct seasonal phases and trend regimes. As a result, the proposed model achieves substantially lower RMSE and MAPE values, confirming its suitability for periodic industrial time series. Finally, to assess the model’s applicability to environmental datasets, it was tested on a meteorological time series capturing daily average temperatures observed in Giresun, Türkiye. The forecasting performance for this series is summarized in Table 5. Benchmark values for comparison were obtained from Kocak et al.13, who applied a deep intuitionistic fuzzy time series model enhanced with LSTM components. This parallel provides methodological consistency and strengthens the basis for performance evaluation. Table 5Forecasting performance metrics for the Giresun temperature dataset using various competing models.Wilamowski et al.55Shin and Ghosh65Chen57Turksen4Song and Chissom8Bas et al.66RMSE2.43942.34382.31485.63792.56822.2254Aladag63Bas et al.67Egrioglu et al.68Kocak et al.13Global RFPedregosa et al.52ProposedRMSE2.30172.30912.17281.95002.24932.3407 1.3802* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). Table 5Forecasting performance metrics for the Giresun temperature dataset using various competing models.Wilamowski et al.55Shin and Ghosh65Chen57Turksen4Song and Chissom8Bas et al.66RMSE2.43942.34382.31485.63792.56822.2254Aladag63Bas et al.67Egrioglu et al.68Kocak et al.13Global RFPedregosa et al.52ProposedRMSE2.30172.30912.17281.95002.24932.3407 1.3802* The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). Forecasting performance metrics for the Giresun temperature dataset using various competing models. The lowest error value in the table is denoted in bold and accompanied by an asterisk (*). Regarding the Giresun temperature dataset, the method succeeded in capturing short-term shifts. Its ensemble structure offered resilience against temperature noise, while fuzzy transitions between regimes enabled smooth adaptation across days with sharp changes. In periods such as April 17–20, 2017, where temperatures fluctuated by more than 10 °C in under 72 h, the model maintained prediction errors within narrow bounds. These results demonstrate not only the effectiveness of the model across domains, but also its ability to self-adjust through the interaction of fuzzy partitioning and ensemble prediction. In order to provide a visual comparison of the proposed method’s forecasting ability, Fig. 4 presents representative plots depicting the alignment between actual and predicted values for four selected time series. Complete forecast–observation comparisons for all datasets and test periods are provided in the supplementary material. Fig. 4Actual vs. forecast values produced by the proposed hybrid fuzzy–ensemble model for selected datasets: BIST100 (May 2010), Australian Beer Production (1990–1994), Daily Temperature in Giresun (Apr 2017), and TAIEX Index (Nov–Dec 2002). Actual vs. forecast values produced by the proposed hybrid fuzzy–ensemble model for selected datasets: BIST100 (May 2010), Australian Beer Production (1990–1994), Daily Temperature in Giresun (Apr 2017), and TAIEX Index (Nov–Dec 2002). As seen in Fig. 4, the model successfully captures underlying trends and fluctuations across different types of data. Forecast results for all datasets are provided in the supplementary material. The next section investigates the parameter configurations that influenced the model’s forecasting accuracy. Model behaviour and parameter insights Hyperparameter search results revealed a tendency toward compact fuzzy structures: most optimal models selected 2 to 4 clusters and low to zero \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} thresholds, supporting the hypothesis that soft overlaps among partitions enhance generalization. Decision trees typically converged to moderate depths (5 to 10), suggesting that deep trees were not necessary to capture localized behaviors due to the already segmented input space. Bootstrap sampling was consistently disabled in all optimal configurations, with no dataset—regardless of its domain or temporal structure—favoring bootstrapped resampling. This uniform tendency likely reflects the relatively limited sample sizes observed across the majority of datasets, where training lengths typically ranged between 80 and 130 instances. In such settings, the increased variance introduced by bootstrap resampling may outweigh its potential ensemble benefits, as previously reported by Buhlmann and Yu69. To provide a clearer picture of the parameter tuning outcomes, the best-performing hyperparameter combinations selected through random search for each dataset are summarized in Table 6. These configurations represent the optimal settings found across the joint fuzzy partitioning and ensemble learning framework. In line with our forecasting-first objective, the number of clusters is treated as a tunable hyperparameter (search range as in Table 1) and chosen by random search purely on the training split to maximize out-of-sample performance; we do not target a standalone cluster-validity study. We prefer random search over exhaustive grid search because the joint space (c, lagn, αcut and RF params) is high-dimensional and interactions are non-separable; sampling-based search identifies good regions with far fewer evaluations, a well-established result in the hyperparameter optimization literature51. Table 6Optimal hyperparameter configurations selected for each dataset.Series n \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}}$$\end{document} p c \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} N \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document} BootstrapBIST100/200910315820.116574FalseBIST100/20101041522030639FalseBIST100/201110615430.110None86FalseBIST100/201210615133043323FalseBIST100/2013106151120.138493FalseTAIEX/1999266451060.187542FalseTAIEX/200027147130129992FalseTAIEX/200124443820891789FalseTAIEX/200224843450.21391345FalseTAIEX/2003249431060.1168542FalseTAIEX/200425043730.11041126FalseABP15416970.1641623FalseTemp. of Giresun12014114070771False Table 6Optimal hyperparameter configurations selected for each dataset.Series n \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}}$$\end{document} p c \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} N \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document} BootstrapBIST100/200910315820.116574FalseBIST100/20101041522030639FalseBIST100/201110615430.110None86FalseBIST100/201210615133043323FalseBIST100/2013106151120.138493FalseTAIEX/1999266451060.187542FalseTAIEX/200027147130129992FalseTAIEX/200124443820891789FalseTAIEX/200224843450.21391345FalseTAIEX/2003249431060.1168542FalseTAIEX/200425043730.11041126FalseABP15416970.1641623FalseTemp. of Giresun12014114070771False Optimal hyperparameter configurations selected for each dataset. As presented in Table 6, the optimal hyperparameter combinations obtained through random search exhibit notable variability across the datasets, reflecting the flexibility and adaptability of the proposed model framework. Particularly, the number of estimators (N) varies considerably, ranging from as low as 10 for BIST100/2011 to as high as 168 for TAIEX/2003, indicating that deeper ensemble learning was necessary for certain series with more volatile or non-linear dynamics. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} parameter reveals that shallower trees (e.g., depth 3–5) were sufficient for several BIST100 and TAIEX subsets, whereas deeper structures (up to depth 17) were preferred for TAIEX / 2001 and ABP. The presence of None for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} in BIST100/2011 suggests that unrestricted growth led to better generalization in that case, potentially due to underlying pattern complexity. Interestingly, all configurations consistently favoured disabling bootstrap sampling. This trend might be attributed to the relatively limited number of observations in most datasets, where bootstrapping could reduce data diversity and impair the ensemble’s performance. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{t}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f}}$$\end{document} values also demonstrate adaptation to data-specific structures. For instance, smaller values (e.g., 2) were often chosen when capturing sharp transitions or local patterns was crucial, as seen in BIST100/2012 and ABP. In contrast, larger values in BIST100 / 2010 and the temperature time series from Giresun suggest a preference for smoother partitions in those contexts. The lag order (p) and cluster count (c) further confirm the method’s sensitivity to the temporal characteristics of each series. Longer lag structures were selected for the temperature time series from Giresun and BIST100/2012, possibly to capture extended autocorrelation patterns, while fewer lags sufficed for datasets like TAIEX/2000. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} parameter, which determines the membership threshold, predominantly took values of 0.1 or 0.0, implying that retaining soft partitioning—even at low thresholds—was generally beneficial. However, the presence of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{\mathrm{c}\mathrm{u}\mathrm{t}}$$\end{document} = 0.2 in TAIEX/2002 suggests that for some series, stronger membership discrimination yielded more stable model behaviour. Because the experiments follow an iterative one-step-ahead protocol, error accumulation may occur in longer forecasting sequences. Investigating direct multi-horizon extensions constitutes a direction for future research. Overall, these results emphasize that the success of the proposed model lies not only in the integration of fuzzy logic and ensemble learning but also in its ability to finely tune model settings according to the underlying structure of the input data. The experimental results suggest that the proposed hybrid fuzzy–ensemble model offers a competitive and, in several cases, superior alternative to established baselines within the class of univariate time series examined in this study. We do not claim universal superiority across all forecasting domains. In particular, the current analysis does not cover multivariate settings, high-frequency tick data, or streaming scenarios with real-time constraints. Extending the method to such regimes, and assessing its robustness under more extreme distributional shifts, are important directions for future work. Although the proposed model does not incorporate explicit concept-drift detection or online adaptation mechanisms, its soft partitioning structure allows partial accommodation of regime-like behaviors through cluster-specific local models. However, in the presence of abrupt structural breaks or rapidly evolving drift patterns, performance may degrade, and online extensions would be required. While the proposed methodology demonstrates strong forecasting performance, it inherently carries additional computational cost due to its dual-phase structure: fuzzy partitioning and per-cluster ensemble training. The application of random search over a multidimensional hyperparameter space further increases runtime. However, these costs are mitigated by the parallelizable nature of both the clustering and random forest training steps. In practice, this allows the model to scale well with modern multi-core computing architectures. Future work may focus on integrating reinforcement learning for adaptive hyperparameter tuning or applying the proposed method to multivariate and streaming time series scenarios. Notably, these parameter trends were consistent across datasets with markedly different temporal structures, suggesting that the proposed framework adapts its complexity primarily through data-driven partitioning rather than manual tuning. Conclusion This study introduced a hybrid forecasting methodology that integrates type-1 fuzzy function modeling with ensemble regression via random forests. By combining soft clustering of lagged inputs with localized ensemble learning, the proposed framework removes the need for manually defined fuzzy rules and offers a structured way to capture both global and local nonlinearities in univariate, offline time series. This architecture provides a data-driven alternative to conventional fuzzy inference systems while maintaining interpretability through membership-weighted aggregation. The inclusion of RF–based baselines in the empirical evaluation further clarifies the contribution of the proposed architecture. While plain and global RF models benefit from nonlinear approximation and ensemble averaging, their performance was consistently inferior to the proposed approach across datasets with pronounced regime changes. This observation suggests that the performance gains are not solely attributable to the use of tree ensembles, but rather to their integration with fuzzy soft partitioning, which enables localized learning and smooth regime transitions that global models fail to capture. Empirical evaluations across thirteen real-world datasets—including financial indices, industrial production data, and meteorological measurements—showed that the method can adapt to heterogeneous temporal structures. The model achieved competitive performance on series with varying degrees of volatility, smoothness, and seasonal regularity, illustrating the complementary strengths of fuzzy partitioning and tree-based ensemble regression. The randomized hyperparameter search further enabled automatic adjustment of lag order, cluster number, and RF configurations without domain-specific tuning, supporting the model’s applicability across diverse datasets. The main contribution of this work lies in establishing a flexible and reproducible learning mechanism that unifies soft computing principles with ensemble modeling under a rule-free structure. While the model leverages only standard components—FCM clustering, nonlinear membership transformations, and Random Forest regressors—their integration into a coherent forecasting pipeline enables a level of adaptability and interpretability not typically present in purely statistical or purely machine-learning approaches. Future extensions may investigate multivariate forecasting settings, interval-valued or type-2 fuzzy formulations to better represent uncertainty, and more efficient ensemble variants, such as Extremely Randomized Trees41, to further enhance model diversity and reduce computational cost. Despite its promising results, the framework exhibits several limitations. Forecast quality may degrade on very short or highly unstable series, where fuzzy partitions become poorly defined and cluster-specific models lack sufficient support. Membership distributions that are excessively diffuse or imbalanced can also reduce the reliability of local regressors. In addition, the current empirical evaluation does not include intermittent or zero-inflated time series. In such settings, where long stretches of zero values dominate the signal, both fuzzy partitioning and tree-based regression may require structural modifications (e.g., demand-occurrence modeling or alternative loss functions). Although random search provides a scalable optimization strategy, it may still miss subtle parameter interactions in high-dimensional search spaces. Furthermore, as with most clustering-based techniques, extreme outliers can distort membership patterns; while the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{cut}$$\end{document} threshold mitigates such effects, robust or possibilistic clustering (e.g., GK or PFF) variants may yield improved stability under heavy-tailed noise. Addressing these challenges, particularly through adaptive clustering mechanisms and hybrid optimization strategies, constitutes a promising direction for future research. Supplementary Information Supplementary Material 1 Supplementary Material 1