A MF-ConvLSTM-XAI model integrating multi-feature and fuzzy control for financial time series forecasting Authors: Ruimin Liu, Shuihan Yi, Arman Ablikim, Xiaonan Song Published: 2025 ABSTRACT In financial time series prediction, existing methods struggle to fully capture the dynamic characteristics and nonlinear relationships of the data. Furthermore, they lack sufficient capability to integrate multiple features (such as prices, trading volumes, and technical indicators), particularly in highly volatile market conditions, where the prediction accuracy and stability of these models tend to be poor. This study proposes a financial time series forecasting (FTSF) model named Multi-Feature Convolutional Long Short-Term Memory Network–Explainable Artificial Intelligence (MF-ConvLSTM-XAI). By deeply integrating Convolutional Long Short-Term Memory Networks (ConvLSTM) with Explainable Artificial Intelligence (XAI) techniques, the model improves the accuracy and stability of financial time series predictions and enhances interpretability. Firstly, the sliding window technique is applied to segment the original financial time series, with data in each window transformed into multiple feature representations. The features analyzed encompass price fluctuations, trading volumes, and technical indicators, enabling the model to identify intricate patterns within the series. Subsequently, the Gramian Angular Difference Fields (GADF) technique is employed to transform the extracted sequential features into image formats. Regarding model design, the MF-ConvLSTM network leverages both convolutional neural networks (CNNs) and Long Short-Term Memory (LSTM) networks. Additionally, a Fuzzy Control Mechanism (FCM) is integrated into the input gate of the ConvLSTM to facilitate subspace reconstruction, enhancing the information flow and state update process in the memory units. To enhance the model’s interpretability, XAI technology is applied to feature importance analysis, decision path visualization, and model explanation reports. This helps researchers and practitioners better understand the model’s decision-making process, increasing the model’s transparency and credibility. To evaluate the effectiveness of the MF-ConvLSTM-XAI model, experiments are conducted using the New York Stock Exchange Composite Index dataset. The results indicate that MF-ConvLSTM-XAI achieves higher prediction accuracy compared with traditional LSTM models and other baseline methods. Specifically, the Mean Squared Error (MSE) on the test set decreases by approximately 15%, and the Mean Absolute Error (MAE) decreases by about 12%. Moreover, the model demonstrates greater robustness under highly volatile market conditions. These findings validate the effectiveness and advantages of MF-ConvLSTM-XAI in FTSF. Overall, by integrating multi-feature information and Fuzzy Cognitive Mechanisms (FCM), MF-ConvLSTM-XAI shows a positive impact on improving the accuracy and stability of financial time series predictions. Although the study achieves promising results on the New York Stock Exchange dataset, the model’s generalization capability still requires further validation across more diverse financial datasets, including cryptocurrency and emerging markets. Nonetheless, this approach provides a valuable tool for financial data analysis and offers new insights for research and applications in the field.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-27121-7. Introduction Financial time series forecasting (FTSF) is a core issue in financial market research, holding significant importance for investment decision-making, risk management, and market trend analysis. Financial time series data often exhibit high volatility, nonlinearity, and noise characteristics, posing substantial challenges for precise predictions1–3. Traditional statistical models, such as the Autoregressive Integrated Moving Average Model (ARIMA)4 and Support Vector Machine (SVM)5, perform well in capturing linear relationships but struggle to account for the complex nonlinear dynamics inherent in financial markets. Furthermore, the rapid growth of market data makes it difficult to rely solely on single data sources for comprehensive predictions, adding to the challenges of FTSF. Recently, the application of deep learning technologies in FTSF has garnered widespread attention. Models like convolutional neural networks (CNNs)6 and Long Short-Term Memory (LSTM) networks7 have been extensively used in financial data analysis due to their strengths in automatic feature extraction and time dependency handling8,9. However, existing deep learning models still have limitations in processing multiple features, often failing to effectively integrate multidimensional attributes such as price, trading volume, and technical indicators. Additionally, in the face of high market volatility and complex nonlinear relationships, traditional LSTM models often underperform in capturing long-term dependencies and global information, resulting in unstable predictions. To address these challenges, this study proposes a novel financial forecasting model, namely the Multi-Feature Convolutional Long Short-Term Memory Network–Explainable Artificial Intelligence (MF-ConvLSTM-XAI). The model uses a Convolutional Long Short-Term Memory Network (ConvLSTM) as its core architecture, capturing both spatial local patterns and long-term temporal dependencies in financial time series. To enhance predictive performance in complex market conditions, a Fuzzy Control Mechanism (FCM) is integrated into the network’s input gate, allowing intelligent regulation of information flow and filtering of market noise. The model also incorporates Explainable Artificial Intelligence (XAI) techniques, such as feature importance analysis and decision-path visualization, to improve transparency and trust in its predictions. The model’s key contributions and innovations are: Fusion of fuzzy control and deep learning: The FCM is deeply embedded in the ConvLSTM input gate. This enables adaptive handling of noise and uncertainty in financial data, significantly improving robustness and stability in volatile markets.End-to-end multidimensional feature engineering and spatiotemporal pattern extraction: A complete workflow transforms multidimensional financial features into two-dimensional images using Gramian Angular Difference Fields (GADF), preserving temporal dynamics. This allows the ConvLSTM to efficiently capture short-term local patterns (spatial dimension) and long-term dependencies (temporal dimension), providing a comprehensive representation of market dynamics. Fusion of fuzzy control and deep learning: The FCM is deeply embedded in the ConvLSTM input gate. This enables adaptive handling of noise and uncertainty in financial data, significantly improving robustness and stability in volatile markets. Fusion of fuzzy control and deep learning: The FCM is deeply embedded in the ConvLSTM input gate. This enables adaptive handling of noise and uncertainty in financial data, significantly improving robustness and stability in volatile markets. End-to-end multidimensional feature engineering and spatiotemporal pattern extraction: A complete workflow transforms multidimensional financial features into two-dimensional images using Gramian Angular Difference Fields (GADF), preserving temporal dynamics. This allows the ConvLSTM to efficiently capture short-term local patterns (spatial dimension) and long-term dependencies (temporal dimension), providing a comprehensive representation of market dynamics. End-to-end multidimensional feature engineering and spatiotemporal pattern extraction: A complete workflow transforms multidimensional financial features into two-dimensional images using Gramian Angular Difference Fields (GADF), preserving temporal dynamics. This allows the ConvLSTM to efficiently capture short-term local patterns (spatial dimension) and long-term dependencies (temporal dimension), providing a comprehensive representation of market dynamics. Recent related work Current state of FTSF research FTSF has long been an important topic in financial market analysis, attracting considerable attention from researchers. Cheng et al. (2022)10 utilized a multimodal graph neural network for FTSF, improving prediction accuracy by integrating multimodal features from various financial data, thereby showcasing the potential of graph neural networks in the financial domain. Wu et al. (2022)11 proposed a price graph-based time series analysis method that leveraged the structural information of financial time series for stock prediction, effectively enhancing the model’s predictive capabilities. Dudukcu et al. (2023)12 employed a method combining temporal convolutional networks with recurrent neural networks for chaotic time series forecasting, providing an effective means to address the nonlinear features present in financial data. Edalatpanah et al. (2024)13 integrated fuzzy logic with hybrid time series forecasting methods, successfully applying them to financial predictions and demonstrating their applicability and accuracy in uncertain financial environments. To address the nonlinear characteristics of financial markets, an increasing number of studies in recent years have focused on machine learning-based nonlinear models. Babii et al. (2022)14 explored the application of machine learning time series regression models, particularly in the context of nowcasting in financial forecasting, demonstrating the potential and advantages of machine learning in improving prediction accuracy. Masini et al. (2023)15 reviewed recent advancements in machine learning for time series forecasting, discussing the application of various machine learning models in financial time series and their improvements in predictive performance. Zhang et al. (2024)16 analyzed and forecasted financial time series using Recurrent Neural Networks (RNNs), with results indicating that RNNs could effectively capture long-term dependencies in time series data, enhancing prediction accuracy. Song (2024)17 utilized a random forest model to analyze price fluctuations in China’s carbon market and explored international carbon credit financing mechanisms, illustrating the application potential of machine learning models in carbon market price forecasting. Applications of deep learning in time series forecasting With the advancement of deep learning technologies, an increasing number of researchers have turned their attention to neural network-based deep learning models to address the challenges of nonlinearity and multidimensional feature integration in time series forecasting. Wang and Yan (2023)18 applied machine learning models to develop quantitative trading strategies for the U.S. market. They systematically evaluated the effectiveness of these strategies across different time horizons. Their findings highlighted the practical potential of deep learning in financial market applications. Li and Jung (2023)19 reviewed methods and applications of deep learning in multivariate time series anomaly detection, discussing the challenges and prospects in processing complex data. Benidis et al. (2022)20 conducted a systematic review of the application of deep learning in time series forecasting, providing a literature overview that highlighted the rapid development and wide-ranging applications in this field. He et al. (2023)21 investigated deep learning ensemble models for financial time series, showing that these models could improve prediction accuracy and were suitable for complex financial data analysis. Zaheer et al. (2023)22 utilized LSTM and deep learning models for multi-parameter prediction of stock time series, with results indicating high accuracy in handling financial data. Bello et al. (2024)23 explored the application of deep learning in high-frequency trading, proposing solutions for real-time fraud detection and emphasizing the challenges of timeliness and accuracy. Some scholars have also explored the application of XAI techniques in time series forecasting. For example, Theissler et al. (2022)24 reviewed the use of XAI techniques in time series classification, highlighting the importance of interpretability and transparency in classification models for optimizing time series prediction. They proposed an explanatory framework for classification models, improving the credibility of model predictions and advancing the widespread application of XAI in time series analysis. Çelik et al. (2023)25 incorporated XAI techniques into FTSF, extending the predictive power of machine learning. By providing interpretability of model decisions, their research significantly improved prediction transparency and trust, offering a highly practical solution for time series analysis in the financial market. Al-Najjar et al. (2023)26 proposed a Shapley Additive Explanations-based XAI method for spatial landslide prediction in time-series SAR data. The results demonstrated that XAI technology enhanced the interpretability of time-series models, contributed to more accurate landslide risk identification, and provided important insights for geological disaster forecasting. Current state of FCM FCM is initially developed for control systems to handle uncertainty and imprecise information. In recent years, these mechanisms have begun to be applied in the input-output control of deep learning networks, particularly in time series models like LSTM. By implementing fuzzy control on the input and forget gates, the network’s ability to manage uncertainty in time series data can be enhanced27. Zhuang and Wei (2023)28 proposed a FCM that combined legal systems with regional financial risk warning models, effectively preventing regional financial risks and highlighting the integration of legal and financial systems. SS Júnior et al. (2023)29 reviewed the applications of deep fuzzy systems in regression problems, focusing on enhancing model interpretability and showcasing the potential of fuzzy control in complex systems. Kan et al. (2024)30 applied a fuzzy Deep Neural Network (DNN) framework to trading and portfolio rebalancing, demonstrating the practicality and interpretability of FCM in financial trading. Tan et al. (2024)31 introduced Fuzzy Embedded Recurrent Neural Network (FE-RNN), which improved the interpretability of neural networks through fuzzy control, applying it to complex time series forecasting tasks. Research gaps and innovations in this study The literature review reveals that existing FTSF models, particularly those based on deep learning, have made notable progress in addressing temporal dependencies and nonlinear relationships. However, there remain deficiencies in multi-feature fusion, making it challenging to fully integrate multidimensional information such as price, trading volume, and technical indicators. Additionally, the performance of existing models in high-volatility market environments is still not robust enough, with the stability of predictions requiring further enhancement. Although FCM has demonstrated their capability in addressing complex nonlinear problems within certain neural networks, their application in multi-feature fusion deep learning models has not been extensively explored. To fill this gap, this study effectively combines the strengths of CNNs and LSTM to extract both local features and long-term dependencies from time series data. Furthermore, it innovatively introduces a FCM to optimize the input gates of the LSTM network, thereby enhancing the model’s ability to process uncertain information and improve prediction robustness. Additionally, the model innovatively integrates XAI techniques, enhancing transparency through feature importance analysis and decision-path visualization. By integrating multi-feature fusion with FCM, the proposed model not only enhances prediction accuracy but also exhibits greater stability and adaptability under high volatility conditions, offering a new solution for FTSF. FTSF methods Overall architecture analysis of deep learning-based FTSF models In FTSF, the complexity and multidimensionality of the data impose significant demands on the model. Data in financial markets not only exhibit substantial temporal dependencies but also encompass characteristics such as nonlinearity, noise, and sudden fluctuations32. Moreover, market data typically includes various feature dimensions, such as price, trading volume, and technical indicators, making it insufficient for a single feature to comprehensively capture market changes33,34. Therefore, constructing a model capable of fully integrating multi-feature information while simultaneously addressing short-term local fluctuations and long-term trends represents a key challenge in FTSF. To tackle these issues, this study proposes the MF-ConvLSTM-XAI, as illustrated in Fig. 1, to address problems related to multi-feature processing, complex dependency capturing, and robustness enhancement in FTSF. Fig. 1FTSF model architecture based on MF-ConvLSTM-XAI. FTSF model architecture based on MF-ConvLSTM-XAI. In Fig. 1, the feature input layer first utilizes a sliding window technique to segment the raw time series into several subsequences. These subsequences encompass multidimensional features such as price, trading volume, and technical indicators, ensuring that complex patterns occurring in the market can be captured. During the feature extraction process, the input features are initially processed through the CNNs layer. After local feature extraction by the CNNs, the feature maps are passed to the LSTM layer. The LSTM layer serves as the core component for handling temporal dependencies within the model, enabling it to capture long-term dependencies in financial time series data. In the LSTM layer, XAI techniques were also introduced to enhance the model’s transparency and interpretability. XAI techniques help users understand how the model derives predictions from input data by visualizing the internal states of the LSTM units and decision paths. To further optimize the information transfer process, the MF-ConvLSTM model introduced in this study incorporates a FCM to adjust the input gates of the LSTM. This FCM dynamically regulates the opening and closing of the network’s input gates based on the uncertainty and fuzziness of the input data, ensuring that important information is effectively transmitted while irrelevant or noisy information is filtered out. Consequently, the model can make more robust predictions in uncertain market environments, particularly under high-volatility conditions, where the FCM significantly enhances prediction stability. Feature extraction and GADF transformation analysis in financial time series In the feature extraction process of financial time series, this study employs a sliding window technique to segment the time series into multiple local segments, each containing data across several feature dimensions, including price, trading volume, and technical indicators (such as moving averages and the Relative Strength Index)35. These features collectively form a complete input sequence aimed at capturing the complex dynamic changes in the financial market. Based on the sliding window technique, MF-ConvLSTM-XAI can divide the time series into multiple subsequences without losing temporal information, processing each subsequence individually. First, the data from each segment is replaced with the original values, first-order differences, growth entropy, and standard deviation. The first-order difference series represents the degree of change between local elements, while growth entropy reflects the complexity of the financial time series, and standard deviation indicates the volatility of the original data. For financial window data of size sn, the first-order difference series \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d_i}$$\end{document} can be expressed as shown in Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d_i}={x_{i+1}} - {x_i},1 \leqslant i \leqslant sn$$\end{document} Growth entropy can evaluate the complexity of data fluctuations, considering both the rank order and the magnitude of the time series fluctuations. The generation process is shown in Fig. 2. Fig. 2Flowchart of the growth entropy sequence generation process. Flowchart of the growth entropy sequence generation process. In Fig. 2, when calculating the growth entropy sequence, a financial time series with a length of sn is first given. Using the sliding window technique, the financial time series is divided into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {\frac{{sn - sn{\prime}}}{{sd{\prime}}}+1} \right)$$\end{document} sub-windows of equal size, with a window size of sn′ and a step size of sd′. Each sub-window is denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U=\left\{ {{u_1},{u_2},{u_3}, \ldots ,{u_{sn{\prime}}}} \right\}$$\end{document}. According to the definition of growth entropy, it is necessary to define an embedding dimension m and a precision g to represent the fluctuation amplitude of the financial time series. For shorter sequences, the embedding dimension m is typically set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2 \leqslant m \leqslant 6$$\end{document}, and the precision for the fluctuation amplitude is set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2 \leqslant g \leqslant 8$$\end{document}. The first-order difference sequence of the sub-window data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v\left( i \right)$$\end{document} is then calculated, which can be expressed as Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v\left( i \right)={u_{i+1}} - {u_i},1 \leqslant i \leqslant sn^{\prime}$$\end{document} Based on the embedding dimension m, the sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ {v\left( i \right)} \right\}$$\end{document} is divided into vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V\left( i \right)=\left\{ {v\left( i \right),v\left( {i+1} \right), \ldots ,v\left( {i+m - 1} \right)} \right\}$$\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}$$1 \leqslant i \leqslant \left( {sn^{\prime} - m} \right)$$\end{document}. Next, for the above subsequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V\left( i \right)$$\end{document}, the corresponding pattern vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w\left( i \right)=\left[ {\left( {s\left( 1 \right),q\left( 1 \right)} \right),\left( {s\left( 2 \right),q\left( 2 \right)} \right), \cdots ,\left( {s\left( m \right),q\left( m \right)} \right)} \right]$$\end{document} is obtained, where s(i) and q(i) can be expressed as Eqs. (3) and (4):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\left( i \right)=\operatorname{sgn} \left( {v\left( i \right)} \right)$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q\left( i \right)=\hbox{min} \left\{ {g,\left| {\frac{{v\left( i \right) \times g}}{{std\left( {\left\{ {v\left( i \right)} \right\}} \right)}}} \right|} \right\}$$\end{document} When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$std\left( {\left\{ {v\left( i \right)} \right\}} \right)=0$$\end{document}, the magnitude of each element is represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\left( i \right)=0,q\left( i \right)=0$$\end{document}. The frequency of occurrence of the pattern vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w\left( i \right)$$\end{document} is calculated, represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q\left( {w\left( i \right)} \right)$$\end{document}. From this, the occurrence frequency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P\left( {w\left( i \right)} \right)$$\end{document} of the pattern vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w\left( i \right)$$\end{document} can be obtained, and its calculation is given in Eq. (5):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P\left( {w\left( i \right)} \right)=\frac{{Q\left( {w\left( i \right)} \right)}}{{sn^{\prime} - m}}$$\end{document} Finally, the growth entropy for a specific window can be calculated as Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H\left( m \right)= - \frac{1}{{m - 1}}\sum\limits_{{i=1}}^{{\left( {2g+1} \right){}^{m}}} {P\left( {{w_i}} \right)\log P\left( {{w_i}} \right)}$$\end{document} For the standard deviation sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ {std} \right\}$$\end{document}, the standard deviation is calculated as shown in Eq. (7):7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$st{d_i}=\sqrt {\frac{1}{{sn^{\prime} - 1}}\sum\limits_{{i=1}}^{{sn{\prime}}} {{{\left( {{u_i} - \bar {U}} \right)}^2}} }$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar {U}$$\end{document} refers to the mean of the data in the sub-window sequence. To fully leverage multidimensional features and enable CNNs to capture deep patterns in time series, this study employs GADF for feature transformation. This method projects time series data into polar coordinates and computes the angular differences between different time points to generate two-dimensional image matrices36. The key advantage of this image-based transformation is that it encodes one-dimensional temporal dependencies into two-dimensional spatial correlations. This allows CNNs to effectively extract local patterns and structures that are difficult to detect directly from the raw sequence. Additionally, because this transformation is insensitive to the absolute values of timestamps, it demonstrates a degree of robustness when handling irregularly spaced sequences. However, applying GADF involves inherent trade-offs. First, the process significantly increases data dimensionality and computational complexity; for long sequences, the resulting image matrices can become very large. Second, although GADF aims to preserve temporal dynamics, there is a theoretical risk of information loss during coordinate mapping, which may prevent capturing all subtle fluctuations of the original sequence. Therefore, it is crucial to carefully select an appropriate window size in practice to balance feature richness and computational cost, making this a key consideration in the design of the present model. During the feature extraction process, XAI techniques are further applied to analyze the image features after GADF transformation. By conducting interpretive analysis of the image features, the model is able to identify which time points or features have a key influence on the prediction results. This, in turn, enhances the model’s understanding of complex financial time series. The financial time series image processing based on GADF is shown in Fig. 3. Fig. 3Schematic diagram of financial time series image processing based on GADF. Schematic diagram of financial time series image processing based on GADF. In Fig. 3, during this image processing, a standard deviation sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S=\left\{ {{s_1},{s_2},{s_3}, \ldots ,{s_{sn}}} \right\}$$\end{document} is first scaled to the range of [− 1, 1] for each element in the sequence, as expressed in Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde {S}_{{ - 1}}^{i}=\frac{{\left( {{s_i} - \hbox{max} \left( S \right)} \right)+\left( {{s_i} - \hbox{min} \left( S \right)} \right)}}{{\hbox{max} \left( S \right) - \hbox{min} \left( S \right)}}$$\end{document} The normalized standard deviation in Cartesian coordinates is then converted to the normalized standard deviation in polar coordinates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {{\theta _i},r} \right)$$\end{document}, as shown in Eqs. (9) and (10):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _i}=\arccos \left( {{{\tilde {s}}_i}} \right), - 1 \leqslant {\tilde {s}_i} \leqslant 1,{\tilde {s}_i} \in \tilde {S}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r=\frac{{{t_i}}}{{sn}},{t_i} \leqslant sn$$\end{document} The GADF image is obtained by considering the angular difference relationships between points in the polar coordinate system, as expressed in Eq. (11):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} GAD{F_{matrix}}= & \left[ {\sin \left( {{\theta _i}+{\theta _j}} \right)} \right] \\ = & \left( {\begin{array}{*{20}{c}} {\sin \left( {{\theta _1}+{\theta _1}} \right)}&{\sin \left( {{\theta _1}+{\theta _2}} \right)}& \cdots &{\sin \left( {{\theta _1}+{\theta _n}} \right)} \\ {\sin \left( {{\theta _2}+{\theta _1}} \right)}&{\sin \left( {{\theta _2}+{\theta _2}} \right)}& \cdots &{\sin \left( {{\theta _2}+{\theta _n}} \right)} \\ \vdots & \vdots & \ddots & \vdots \\ {\sin \left( {{\theta _n}+{\theta _1}} \right)}&{\sin \left( {{\theta _n}+{\theta _2}} \right)}& \cdots &{\sin \left( {{\theta _n}+{\theta _n}} \right)} \end{array}} \right) \\ = & {\sqrt {I - {{\tilde {S}}^2}} ^{\prime}} \cdot \tilde {S} - {{\tilde {S}}^{\prime}} \cdot \sqrt {I - {{\tilde {S}}^2}} \\ \end{aligned}$$\end{document} I refers to the unit row vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {1,1, \cdots ,1} \right]$$\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}$${\tilde {S}^{\prime}}$$\end{document} refers to the transposed vector of the standard deviation sequence in polar coordinates. Analysis of FCM in the financial time series prediction model In FTSF tasks, market data often exhibit high levels of noise and uncertainty. To address these issues, the MF-ConvLSTM model proposed in this study not only combines the advantages of CNN and LSTM but also introduces FCM37 to optimize the input gates of LSTM. Additionally, the model further enhances interpretability through XAI techniques, forming the MF-ConvLSTM-XAI model. By combining multi-feature fusion and FCM, this model seems to handle both short-term patterns and long-term dependencies in time series data. Additionally, it provides a transparent decision-making process through XAI techniques. In the MF-ConvLSTM-XAI model, FCM is used to dynamically adjust the opening and closing of the LSTM input gates to address the uncertainty and fuzziness in the input data. Specifically, FCM filters out high-volatility and noisy data by calculating the fuzzy information of the input, ensuring that important information is certainly transmitted. In the MF-ConvLSTM-XAI model, growth entropy and standard deviation are used to measure the complexity and volatility of the financial time series, respectively. By combining these two metrics, the input gate of the ConvLSTM unit filters the data to eliminate outliers and points with high volatility. The FCM is applied in the input gate of the ConvLSTM unit. This mechanism not only strengthens the model’s ability to handle uncertain data but also provides an explanatory analysis of the fuzzy control process through XAI techniques. Firstly, the fuzzy information \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$fuz{z_i}$$\end{document} is calculated, as expressed in Eqs. (12) and (13):12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Gs=\frac{1}{{2\pi {{\bar {z}}^2}}}{e^{ - \frac{{\left( {{o^2}+{p^2}} \right)}}{{2{{\bar {z}}^2}}}}}$$\end{document}13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$fuz{z_i}=Gs \cdot x_{t}^{{in\_std}}$$\end{document} In this context, o and p represent the coordinates of the central pixel, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar {z}$$\end{document} refers to the standard deviation of the Gaussian filter kernel, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{t}^{{in\_std}}$$\end{document} denotes the input tensor formed by growth entropy and standard deviation at time t. The hidden state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_i}$$\end{document} at the current time step can be expressed as Eq. (14), while the hidden state HF after processing by the FCM can be expressed as Eqs. (15),14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_i}={W_{hi}} \cdot {h_{t - 1}}$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_{fi}}=\tanh \left( {{h_i} \circ {W_{fuzz\_i}} * fuz{z_i}} \right)$$\end{document} *denotes the convolution operator and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\circ$$\end{document} denotes the Hadamard operator. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_{t - 1}}$$\end{document} refers to the output of the unit at the previous time step, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W_{hi}}$$\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}$${W_{fuzz\_i}}$$\end{document} are the learnable parameters of the model. The hyperparameters within the FCM, such as the standard deviation of the Gaussian kernel in Eq. (12), are not obtained through end-to-end training. Instead, they are empirically set based on preliminary experiments on the validation set. The goal is to optimize the filtering of input uncertainty, with the chosen values designed to balance noise suppression and the retention of useful information. To integrate the information from the input gate, it can be expressed as Eq. (16).16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_{t\_i}}={h_{t - 1}} - {h_i}+{h_{fi}}$$\end{document} Finally, after the FCM processes the uncertainty in the input data, the input gate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${i_t}$$\end{document}), forget gate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_t}$$\end{document}), and output gate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${o_t}$$\end{document}) of the MF-ConvLSTM-XAI gating mechanism at time t are calculated, represented by Eqs. (17) to (21).17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${i_t}=\sigma \left( {{W_{xi}} * x_{t}^{{ori\_diff}}+{h_{t\_i}}+{W_{ci}} \circ {c_{t - 1}}+{b_i}} \right)$$\end{document}18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_t}=\sigma \left( {{W_{xf}} * x_{t}^{{ori\_diff}}+{W_{hf}} * {h_{t - 1}}+{W_{cf}} \circ {c_{t - 1}}+{b_f}} \right)$$\end{document}19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${c_t}={f_t} \circ {c_{t - 1}}+{i_t} \circ tanh\left( {{W_{xc}} * x_{t}^{{ori\_diff}}+{W_{hc}} * {h_{t - 1}}+{b_c}} \right)$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${o_t}=\sigma \left( {{W_{xo}} * x_{t}^{{ori\_diff}}+{W_{ho}} * {h_{t - 1}}+{W_{co}} \circ {c_{t - 1}}+{b_o}} \right)$$\end{document}21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_t}={o_t} \circ tanh\left( {{c_t}} \right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{t}^{{ori\_diff}}$$\end{document} indicates the current input formed by the original value sequence and the difference value sequence, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_{t - 1}}$$\end{document} represents the output of the unit at time t-1, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${c_t}$$\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}$${c_{t - 1}}$$\end{document} denote the cell states at times t and t-1, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W_{xi}},{W_{ci}},{W_{xf}},{W_{hf}},{W_{xc}},{W_{xo}},{W_{ho}},{W_{co}}$$\end{document} are the weight matrices, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b_i},{b_f},{b_c},{b_o}$$\end{document} is the bias term. Lastly, a fully connected layer is used as the output layer of the model, as shown in Eqs. (22),22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$output=\rho \left( {{W_e}{h_t}+{b_e}} \right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho$$\end{document} denotes the activation function and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W_e}$$\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}$${b_e}$$\end{document} are the learnable parameters of the model. Building on FCM, XAI techniques are used to analyze how the FCM impacts the model’s decision-making process. Through feature importance analysis, XAI techniques can identify which input features are prioritized for retention or filtering during the fuzzy control process. This helps users understand how the model handles noisy data in high-volatility markets. Furthermore, XAI techniques also enhance the model’s interpretability by visualizing the decision paths, showing how FCM adjusts the states of the input gates under different market conditions. The pseudocode for the financial time series prediction model based on MF-ConvLSTM-XAI is illustrated in Fig. 4. Fig. 4Pseudocode flow of the MF-ConvLSTM-XAI algorithm applied to financial time series prediction. Pseudocode flow of the MF-ConvLSTM-XAI algorithm applied to financial time series prediction. Explainability analysis and visualization To interpret the MF-ConvLSTM-XAI model, this study uses the game-theory-based SHapley Additive exPlanations (SHAP) method. SHAP is a model-agnostic technique that explains predictions by calculating the marginal contribution of each feature to an individual prediction. It has a strong theoretical foundation and consistent results. In this study, SHAP is applied at two levels: Global Feature Importance Analysis: SHAP values are calculated for all test set samples and then aggregated to produce a global feature importance map. In this context, “features” correspond to the pixels in the input images generated using GADF. By identifying pixel regions with the highest mean absolute SHAP values, the critical temporal patterns that most influence model predictions are determined.Local Prediction (Decision-Path) Visualization: For an individual prediction (e.g., forecasting the next day’s stock price), a SHAP Force Plot is generated. This plot visually shows which features (specific regions of the input image) push the predicted value higher (bullish factors) and which push it lower (bearish factors). It provides a clear visual explanation of the decision path for that specific prediction. Global Feature Importance Analysis: SHAP values are calculated for all test set samples and then aggregated to produce a global feature importance map. In this context, “features” correspond to the pixels in the input images generated using GADF. By identifying pixel regions with the highest mean absolute SHAP values, the critical temporal patterns that most influence model predictions are determined. Global Feature Importance Analysis: SHAP values are calculated for all test set samples and then aggregated to produce a global feature importance map. In this context, “features” correspond to the pixels in the input images generated using GADF. By identifying pixel regions with the highest mean absolute SHAP values, the critical temporal patterns that most influence model predictions are determined. Local Prediction (Decision-Path) Visualization: For an individual prediction (e.g., forecasting the next day’s stock price), a SHAP Force Plot is generated. This plot visually shows which features (specific regions of the input image) push the predicted value higher (bullish factors) and which push it lower (bearish factors). It provides a clear visual explanation of the decision path for that specific prediction. Local Prediction (Decision-Path) Visualization: For an individual prediction (e.g., forecasting the next day’s stock price), a SHAP Force Plot is generated. This plot visually shows which features (specific regions of the input image) push the predicted value higher (bullish factors) and which push it lower (bearish factors). It provides a clear visual explanation of the decision path for that specific prediction. A pseudocode example is provided in Fig. 5, demonstrating how to use the SHAP library to analyze a trained MF-ConvLSTM-XAI model. Fig. 5Pseudocode example for implementing XAI using the SHAP library. Pseudocode example for implementing XAI using the SHAP library. Using the above approach and the SHAP library, a high-accuracy predictive model is constructed with strong interpretability. This transforms the model’s internal decision-making process from a “black box” into a transparent system, significantly enhancing its reliability and trustworthiness in high-risk domains such as finance. Experimental evaluation This study aims to verify the performance of the proposed financial time series prediction model based on a multi-feature fusion convolutional LSTM network, using the New York Stock Exchange Composite Index (NYA) dataset (https://www.kaggle.com/datasets/borismarjanovic/price-volume-data-for-all-us-stocks-etfs). The dataset is selected based on several key considerations. First, the New York Stock Exchange (NYSE) is one of the largest and most representative mature financial markets worldwide. Its high-quality data and complex noise patterns offer a challenging yet fair benchmark for evaluating the model’s robustness and effectiveness. Second, the primary goal of this study is to validate the novelty and performance of the MF-ConvLSTM-XAI model architecture. Using a widely studied benchmark market allows the improvements of the model over existing methods to be demonstrated clearly while minimizing confounding factors caused by market heterogeneity. It is acknowledged that the model’s generalizability needs further evaluation on financial markets with different characteristics, such as highly volatile cryptocurrency markets or emerging markets with distinct data distributions. However, this falls outside the scope of the current study, which focuses on validating the core model architecture, and is highlighted as an important direction for future research. After acquiring the dataset, it first normalizes using MinMaxScaler to scale the data to the range of [− 1, 1], ensuring that different features (such as stock price, trading volume, and technical indicators) have the same dimensionality, thereby improving the stability of model training. Next, a sliding window technique is employed to segment the time series data into multiple subsequences of length 50, with a step size of 5 between each subsequence. These subsequences serve as the model input to capture dynamic changes in the financial market. During the feature extraction phase, first-order differences, growth entropy, and standard deviation are computed to extract key features influencing price fluctuations. Subsequently, the GADF technique is utilized to convert the standard deviation sequence into a two-dimensional image representation, suitable for the ConvLSTM model input format. The obtained data are finally categorized into a training set (70%) and a testing set (30%). In terms of the experimental environment, a TensorFlow-based deep learning framework is used for model construction and training. In terms of computational resources, all experiments in this study are conducted on a server equipped with an NVIDIA RTX 3090 GPU. For the NYSE dataset, completing 100 training epochs for the full model takes approximately 75 min. The main computational overhead arises from the GADF image transformation process and the spatiotemporal feature extraction performed by the ConvLSTM layers. While the training time is acceptable, computational efficiency will be a critical consideration for future deployment of the model on larger-scale or higher-frequency datasets in real-world trading scenarios. To ensure the reproducibility of the experiments, a random seed is set. Furthermore, Jupyter Notebook is used as the development and experimental environment, facilitating visualization and data processing. The hyperparameter settings include reasonable configurations for the model’s key parameters. The sliding window size is set to 50 with a step size of 5 to adequately capture the dynamic features of the time series. The MF-ConvLSTM-XAI model utilizes 64 convolutional kernels with a kernel size of 3 × 3 to capture local patterns in the time series. During training, the Adam optimizer is selected, with the Mean Squared Error (MSE) as the loss function, a batch size of 32, a learning rate of 1e-4, and 100 training epochs to ensure rapid convergence and enhance the model’s learning capability. Specific parameter settings are shown in Table 1. Table 1Environmental and hyperparameter settings.ItemValueDescriptionHardwareNVIDIA GPUUsed to accelerate the training and inference of deep learning modelsSoftware frameworkTensorFlow 2.0Deep learning framework for model building and trainingDevelopment environmentPython3. 6Easy to interactively develop and visualize data processingSliding window size50Window size for time series segmentationSliding step size5Step size of sliding windowNumber of convolution kernels64Number of convolution kernels in ConvLSTM layerConvolution kernel size(3, 3)Size of convolution kernels in ConvLSTM layerBatch size32Batch size of data used in each trainingNumber of training rounds100Total number of rounds for model trainingOptimizerAdamUsed to optimize the training process of the modelLoss functionMSEUsed to evaluate the accuracy of model predictions Table 1Environmental and hyperparameter settings.ItemValueDescriptionHardwareNVIDIA GPUUsed to accelerate the training and inference of deep learning modelsSoftware frameworkTensorFlow 2.0Deep learning framework for model building and trainingDevelopment environmentPython3. 6Easy to interactively develop and visualize data processingSliding window size50Window size for time series segmentationSliding step size5Step size of sliding windowNumber of convolution kernels64Number of convolution kernels in ConvLSTM layerConvolution kernel size(3, 3)Size of convolution kernels in ConvLSTM layerBatch size32Batch size of data used in each trainingNumber of training rounds100Total number of rounds for model trainingOptimizerAdamUsed to optimize the training process of the modelLoss functionMSEUsed to evaluate the accuracy of model predictions Environmental and hyperparameter settings. The hyperparameters listed in Table 1 are not selected arbitrarily. They are determined through a systematic combination of grid search and manual tuning on a validation set, created by further splitting 20% from the 70% training data. For instance, different combinations of sliding window sizes ([30, 50, 70]), learning rates ([1e-3, 1e-4, 1e-5]), and numbers of convolutional kernels ([32, 64, 128]) are tested. The final values (e.g., sliding window size of 50, learning rate of 1e-4) achieve the lowest MSE on the validation set, balancing model performance and computational efficiency. The model’s performance is evaluated against several algorithms, including those in38, standard LSTM, ConvLSTM39, and the model proposed by Kan et al. (2024). Evaluation uses metrics such as Accuracy, Precision, Recall, F1 score, MSE, and Mean Absolute Error (MAE). Stock prices serve as the evaluation metric. To benchmark the model against current state-of-the-art (SOTA) architectures, the Informer model40 is included as a comparison. Informer, a Transformer-based model designed for long-sequence time series forecasting, employs a ProbSparse self-attention mechanism and a generative decoder to address the computational complexity and memory limitations of traditional Transformers. The official open-source implementation is used, and optimal hyperparameters (e.g., encoder layers = 2, attention heads = 8, d_model = 512) are configured via grid search on the validation set to ensure a fair comparison. Results and discussion Analysis of convergence performance of different algorithms Firstly, the convergence results of different algorithms on the training set are shown in Fig. 6. Fig. 6Convergence results of different algorithms. Convergence results of different algorithms. In Fig. 6, MF-ConvLSTM-XAI demonstrates a significant advantage in convergence compared to other models based on the loss function values. Experimental results indicate that MF-ConvLSTM-XAI rapidly reduces the loss function value in the early stages of training, achieving a loss of 4.19 in the first training round, which is substantially lower than the initial loss values of the other models: 4.20 for Kan et al., 4.30 for ConvLSTM, 4.32 for BiLSTM, and 4.50 for LSTM. This trend becomes increasingly evident in subsequent training: by the 30th round, the loss value for MF-ConvLSTM-XAI decreases to 1.06, outperforming 1.54 for Kan et al., 2.35 for ConvLSTM, 2.34 for BiLSTM, and 2.82 for LSTM. More notably, in the later stages of training, MF-ConvLSTM-XAI maintains a low loss value, ultimately dropping to 0.24 by the 100th round, while the loss values for other models are generally above 0.58. This convergence speed and loss level not only reflect the efficiency of MF-ConvLSTM-XAI in the optimization process but also indicate its effectiveness and stability in handling financial time series prediction tasks, further reinforcing the superiority of the model presented in this study. Further analysis of the results regarding Accuracy, Precision, Recall, and F1 score for each algorithm is illustrated in Fig. 7. Fig. 7Comparison of prediction accuracy for financial time series across different algorithms. Comparison of prediction accuracy for financial time series across different algorithms. In Fig. 7, a comparison of MF-ConvLSTM-XAI with other models (including BiLSTM, LSTM, ConvLSTM, and the model proposed by Kan et al. (2024)) clearly demonstrates that MF-ConvLSTM-XAI excels across all metrics. Specifically, MF-ConvLSTM-XAI achieves an Accuracy of 97.19%, significantly higher than 85.13% for Kan et al., 82.07% for ConvLSTM, 80.11% for BiLSTM, and 67.85% for LSTM. Additionally, in terms of Precision, Recall, and F1 score, MF-ConvLSTM-XAI achieves results of 88.76%, 72.23%, and 75.40%, respectively, markedly surpassing other models. These results indicate that MF-ConvLSTM-XAI demonstrates higher accuracy, stability, and robustness in FTSF, likely due to its design tailored for handling complex financial data. Additionally, an analysis of the normalized stock prices under different algorithms over a period of 16 trading days is provided in Fig. 8. Fig. 8Fitting results of normalized stock prices under various algorithms. Fitting results of normalized stock prices under various algorithms. In Fig. 8, in the comparison of normalized price fitting performance, the MF-ConvLSTM-XAI model demonstrates outstanding predictive ability. According to the experimental results, MF-ConvLSTM-XAI’s predicted values at several time points are significantly closer to the actual values compared to other models. For instance, on Day 1, MF-ConvLSTM-XAI’s predicted value is − 0.63, outperforming Kan et al.‘s − 0.83, ConvLSTM’s − 0.90, BiLSTM’s − 0.54, and LSTM’s − 0.54. In the subsequent days, MF-ConvLSTM-XAI maintains a consistently high level of accuracy, especially on Day 4 and Day 8, where the predicted values are 0.32 and 0.40, respectively, reflecting its strong ability to capture market trends. In contrast, other models show more fluctuations in their predictions and larger deviations from the actual values, particularly ConvLSTM and LSTM, whose predictions on Day 11 have a gap of more than 0.50 from the actual values. These results indicate that MF-ConvLSTM-XAI not only excels in capturing the overall trend but also demonstrates strong performance in detailed fitting, effectively reducing prediction errors. Overall, MF-ConvLSTM-XAI significantly outperforms BiLSTM, LSTM, ConvLSTM, and the Kan et al. model in financial time series prediction tasks, further proving its effectiveness and feasibility in real-world applications. Analysis of MAE and MSE results for different algorithms A further comparison of MSE and MAE results for each algorithm is presented in Figs. 9 and 10. Fig. 9MSE comparison of various algorithms. MSE comparison of various algorithms. Fig. 10MAE comparison of various algorithms. MAE comparison of various algorithms. In Figs. 9 and 10, MF-ConvLSTM-XAI demonstrates clear advantages in both MSE and MAE comparisons. According to the data in the tables, MF-ConvLSTM-XAI consistently exhibits lower error values at various stages of training. Notably, at the 100th epoch, its MSE is 5.37, while the MSE values of other algorithms are all above 20, with the highest reaching 25.23. This indicates that MF-ConvLSTM-XAI is significantly more accurate in fitting the actual data compared to traditional methods such as BiLSTM, LSTM, and ConvLSTM. Similarly, in terms of MAE, MF-ConvLSTM-XAI has a value of 6.79 at the 100th epoch, whereas the MAE for other models exceed 20. This further validates MF-ConvLSTM-XAI’s effectiveness in reducing prediction errors. These results demonstrate that MF-ConvLSTM-XAI better captures the dynamic characteristics of financial time series, providing more accurate predictions and offering higher reliability in practical applications. Table 2 presents a comparative evaluation of the proposed MF-ConvLSTM-XAI model against other baseline models, including LSTM, BiLSTM, ConvLSTM, the model by Kan et al., and the state-of-the-art (SOTA) Transformer-based Informer. Table 2Comprehensive performance evaluation of different models on the test set.ModelAccuracy (%)Precision (%)Recall (%)F1 Score (%)MSEMAELSTM67.8558.1350.8853.4125.2322.15BiLSTM80.1170.7154.4957.1723.421.06ConvLSTM82.0773.0756.5460.6821.8820.34Kan et al.85.1375.0659.5562.4418.5216.77Informer96.1584.7871.0277.297.128.54MF-ConvLSTM-XAI97.1988.7672.2375.45.376.79 Table 2Comprehensive performance evaluation of different models on the test set.ModelAccuracy (%)Precision (%)Recall (%)F1 Score (%)MSEMAELSTM67.8558.1350.8853.4125.2322.15BiLSTM80.1170.7154.4957.1723.421.06ConvLSTM82.0773.0756.5460.6821.8820.34Kan et al.85.1375.0659.5562.4418.5216.77Informer96.1584.7871.0277.297.128.54MF-ConvLSTM-XAI97.1988.7672.2375.45.376.79 Comprehensive performance evaluation of different models on the test set. The results indicate that traditional LSTM and BiLSTM models perform relatively poorly across all metrics, struggling to capture the complex dynamics of financial time series. Performance improves with ConvLSTM and the model by Kan et al., which leverage convolutional structures. The Transformer-based Informer, a state-of-the-art model, achieves strong predictive capabilities, with accuracy reaching 96.15% and MSE decreasing to 7.12. This demonstrates the advantage of attention mechanisms in capturing long-range dependencies and establishes a highly competitive benchmark. Despite this, MF-ConvLSTM-XAI outperforms Informer on key metrics. It attains the highest accuracy (97.19%), precision (88.76%), MSE (5.37), and MAE (6.79). While Informer achieves a slightly higher F1 score due to its balanced precision-recall trade-off, the very high precision of MF-ConvLSTM-XAI indicates more reliable predictions with fewer false alarms. This advantage arises from the combination of GADF and ConvLSTM, which enables effective extraction of unique local spatiotemporal patterns in financial data. The integration of the FCM further filters noise and uncertainty, enhancing prediction accuracy. These findings suggest that general SOTA architectures like Informer are indeed powerful. However, domain-specific hybrid models that integrate multi-feature fusion and customized modules can achieve superior performance on key metrics. Such models offer a more robust and practical solution for financial forecasting. Conclusion The proposed MF-ConvLSTM-XAI model demonstrates promising results in FTSF. By integrating multi-feature fusion, the FCM, and XAI techniques, the model enhances both prediction accuracy and stability. Experimental results show that, compared to traditional LSTM models and other baseline methods, the MF-ConvLSTM-XAI model reduced the MSE by approximately 15% and the MAE by about 12% on the test set. Particularly in high-volatility market environments, the model demonstrated stronger robustness, further validating its effectiveness and reliability in real-world applications. This helps researchers and practitioners better understand the model’s decision-making process. Although the MF-ConvLSTM-XAI model performs excellently in FTSF, it still has some limitations. First, the experimental validation in this study is conducted solely on the NYSE Composite Index dataset. Whether the model maintains robust performance across other markets—such as highly volatile cryptocurrency markets or emerging markets with distinct data characteristics—remains a key issue for future research to address, in order to comprehensively assess its generalizability. Second, the model’s performance is highly dependent on the selection of hyperparameters, such as the sliding window size and step size, which may need to be dynamically adjusted according to changes in market conditions. Future research could explore adaptive adjustment strategies for these hyperparameters to better respond to market fluctuations and changes in data characteristics. Third, while FCM enhances the model’s robustness, its design and parameter tuning may require extensive domain knowledge, which could limit the model’s applicability across different financial scenarios. Future work could further optimize the design of FCM to improve its adaptability in various market environments. Additionally, the application of the MF-ConvLSTM-XAI model is not limited to the financial sector. It could also be extended to other time series forecasting tasks, such as healthcare and environmental monitoring. This would help explore the potential of deep learning in solving cross-domain problems. Overall, the MF-ConvLSTM-XAI model provides a new solution for FTSF, and significantly improves prediction accuracy and stability, advancing the further development of the financial market forecasting field. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1