AI-Driven Sensing for Cross-Lingual Risk Prediction via Semantic Alignment and Multimodal Temporal Fusion

Abstract

In the context of highly interconnected global markets and the rapid dissemination of multilingual information, traditional risk prediction methods that rely on single numerical sequences or monolingual text are insufficient for achieving early perception of cross-market risks. To address this issue, a cross-market risk early warning framework based on multilingual large language models and multimodal sensing fusion is proposed. The proposed approach is centered on a unified risk semantic space, where cross-lingual semantic alignment is employed to reduce semantic discrepancies across languages. Furthermore, a semantic–volatility coupling attention mechanism is introduced to capture the dynamic relationship between textual semantic evolution and market fluctuations. In addition, cross-market knowledge transfer and low-resource enhancement strategies are incorporated to improve the model’s generalization capability across multilingual and multi-market environments, thereby establishing an intelligent perception and early warning system for complex sensing scenarios. Experimental results demonstrate that the proposed method significantly outperforms multiple baseline models in multilingual cross-market risk prediction tasks. In the main experiment, the model achieves a root mean squared error (RMSE) of $0.1127$, an mean absolute error (MAE) of $0.0846$, and an area under the curve (AUC) of $0.8879$, while the early warning gain is improved to $5.2$ days, which is substantially better than the Transformer model (RMSE $0.1365$, AUC $0.8042$) and the multilingual BERT-based fusion model (AUC $0.8395$). In terms of classification performance, higher accuracy, precision, and recall are consistently achieved, with overall accuracy exceeding $0.88$, and both precision and recall are maintained above $0.85$, indicating strong discriminative capability in risk identification tasks. Cross-lingual generalization experiments further verify the robustness of the proposed framework. When trained solely on the English market, the model achieves AUC values of $0.8624$ and $0.8471$ on the Chinese and European markets, respectively, with RMSE reduced to $0.1185$, significantly outperforming competing methods. Overall, the proposed approach achieves substantial improvements in prediction accuracy, cross-lingual generalization, and early warning performance, providing an effective solution for artificial intelligence-driven sensing and risk early warning.

1. Introduction

In the context of highly integrated global economies and rapid information dissemination, the interconnectivity among financial markets has been continuously strengthened, and the formation and propagation of risks exhibit pronounced cross-market, cross-regional, and even cross-lingual characteristics [1]. Policy changes, macroeconomic indicators, and multilingual news sentiments from different countries and regions are often transmitted through complex information networks to global capital markets, becoming key driving factors of market fluctuations [2]. Therefore, how to achieve early perception and accurate forecasting of financial risks in a multi-source heterogeneous information environment has become a critical research problem in financial engineering and intelligent computing [3]. In particular, under the realistic scenario where multilingual information coexists, constructing risk modeling approaches capable of unified semantic understanding across different languages and capturing their impact mechanisms on markets is of great theoretical significance and practical value for enhancing cross-market risk management capabilities [4].

Traditional financial risk prediction methods primarily rely on structured time-series data, such as prices, returns, trading volumes, and macroeconomic indicators, and perform analysis through statistical modeling or classical machine learning approaches [5]. For instance, models based on autoregressive integrated moving average (ARIMA) or generalized autoregressive conditional heteroskedasticity (GARCH) are able to characterize temporal dependencies of market volatility to a certain extent [6]. However, such methods are generally built upon linear assumptions, making them insufficient for capturing complex nonlinear relationships, and they exhibit limited responsiveness to sudden risk events [7]. With the development of machine learning, methods such as support vector machines and random forests have been introduced into financial prediction tasks, achieving certain progress in nonlinear modeling; nevertheless, they still mainly depend on structured data and ignore the potential risk signals embedded in textual information [8]. Moreover, these approaches generally lack the capability to model cross-market information propagation mechanisms, making them inadequate for addressing the interconnected nature of global financial systems [9].

In recent years, deep learning methods have demonstrated significant advantages in financial risk prediction [10]. Models based on recurrent neural networks (RNN) and their variants, such as long short-term memory (LSTM), can effectively capture long-term dependencies in time series, while Transformer architectures further enhance the modeling of complex temporal patterns through self-attention mechanisms [11]. In terms of textual analysis, approaches based on pretrained language models such as BERT are capable of extracting sentiment and semantic features from unstructured data, including news and announcements, thereby providing new information sources for financial prediction [12]. However, existing deep learning methods still suffer from several limitations [13]. On the one hand, most studies are confined to monolingual textual environments, making it difficult to handle semantic discrepancies and alignment across different languages. On the other hand, textual semantic signals and numerical market data evolve at different temporal scales, and current methods typically adopt simple concatenation or coarse-grained fusion strategies, lacking in-depth modeling of cross-modal dynamic relationships. In addition, under low-resource language scenarios, data scarcity often leads to poor generalization performance [14].

With the rapid development of multilingual large language models (Multilingual Large Language Models, Multilingual LLM), their capabilities in cross-lingual semantic representation and knowledge transfer provide new technical pathways for financial risk modeling [15]. Such models, pretrained on large-scale multilingual corpora, are able to map different languages into a unified semantic space, thereby laying the foundation for cross-lingual information fusion [16]. However, directly applying multilingual LLMs to financial risk prediction still faces several critical challenges. First, semantic shifts exist in financial terminology and expressions across different languages, making it difficult to directly align risk information. Second, textual semantic signals are inherently dynamic, and their temporal evolution is not synchronized with high-frequency market fluctuations, increasing the difficulty of cross-modal modeling. Finally, in certain language markets, limited data availability raises challenges for effective cross-market knowledge transfer and model robustness [17]. Therefore, constructing a unified cross-lingual semantic representation and achieving effective coupling between semantic information and market dynamics remain key scientific challenges in intelligent financial modeling [18]. Yue Dazhi et al. [19] proposed a Dynamic Weighted Multimodal Financial Forecasting Models framework, which integrates multimodal financial data through dynamic weighting strategies, improving prediction accuracy and multi-source information utilization. Hsiu-Min Chuang et al. [20] proposed a framework for Chinese financial news analysis for sentiment and stock prediction, integrating multiple textual representations and model categories to achieve systematic comparison and improved performance. Zhi Yang et al. [21] introduced the UniFinEval multimodal financial evaluation benchmark, integrating text, image, and video data to evaluate large language models in high-density financial information environments. Georgios Kostopoulos et al. [22] presented a comprehensive review of deep learning in e-commerce, analyzing core models, application challenges, and future directions. Despite recent progress in multilingual prediction studies, several critical limitations remain. First, most existing approaches rely on static semantic representations, which are insufficient to capture the dynamic evolution of semantics across languages, leading to suboptimal cross-lingual consistency. Second, current multimodal methods often adopt simple fusion strategies and lack effective modeling of temporal dependencies and inter-modal interactions. In addition, the generalization capability of these models remains limited in complex cross-market environments.

To address the aforementioned issues, a cross-market risk early warning framework based on multilingual large language models and multimodal financial sensor fusion is proposed. The framework is centered on a unified risk semantic space, where multilingual semantic encoding is employed to extract latent risk information from texts in different languages and is jointly modeled with market sensor data, including prices, volatility, and trading volume. On this basis, a semantic–volatility coupling attention mechanism is designed to dynamically capture the relationship between semantic variations and market fluctuations, thereby enhancing risk perception capability and prediction accuracy. Meanwhile, cross-lingual alignment and low-resource enhancement strategies are introduced to enable knowledge transfer across different language markets, improving model generalization in diverse market environments. The main contributions of this study are summarized as follows:

• A unified cross-lingual semantic representation space for financial risk is constructed, where semantic alignment mechanisms are employed to reduce expression discrepancies across languages, achieving language-invariant and risk-consistent modeling;
• A semantic–volatility coupling attention mechanism is proposed to dynamically model the relationship between textual semantic changes and market fluctuations, thereby enhancing cross-modal fusion effectiveness;
• A cross-market generalization and low-resource enhancement strategy is designed to improve model stability and robustness under data-scarce scenarios through cross-lingual knowledge transfer;
• A multilingual and multi-market risk early warning experimental framework is established to validate the effectiveness of the proposed method from both prediction accuracy and early warning capability perspectives.

From the perspectives of cross-lingual semantic alignment and multimodal dynamic fusion, a novel financial risk perception approach for global market environments is presented. The core innovation lies in mapping multilingual semantic representations and financial sensor signals into a unified latent space, enabling structured representation of cross-lingual risk information, introducing a semantic change-driven coupling mechanism to reveal the potential triggering role of textual information in market fluctuations, and enhancing model generalization under low-resource language environments through cross-market knowledge transfer strategies. This approach provides an important theoretical foundation and technical support for the development of next-generation cross-market risk early warning systems.

2. Related Work

This section provides a systematic review of the theoretical foundations and recent advances closely related to this study. First, existing approaches to multilingual representation learning and semantic alignment are examined to highlight their capabilities and limitations in cross-lingual contexts. Second, recent developments in multimodal data fusion for complex temporal prediction tasks are discussed, with an emphasis on their applicability to heterogeneous data environments. Finally, the limitations of current methods in cross-market risk modeling are summarized. These observations reveal that challenges remain in achieving consistent cross-lingual semantic representation and effective dynamic coupling of multimodal information, which motivates the proposed framework in this study.

2.1. Financial Risk Forecasting and Multimodal Market Modeling

The core task of financial risk forecasting lies in characterizing the dynamic evolution of market states over time and subsequently modeling and predicting potential future risks [23]. Traditional approaches are primarily grounded in time-series analysis theory, where mathematical models are constructed based on historical observations to capture autocorrelation and evolutionary patterns among variables [24]. For instance, autoregressive integrated moving average models achieve linear forecasting of future trends through stationarity transformation and lag modeling, while conditional heteroskedasticity models describe the time-varying variance characteristics of financial time series from a volatility perspective [25]. Although such methods exhibit strong interpretability, their modeling capacity is constrained by linear assumptions, making them insufficient for capturing the complex nonlinear structures and abrupt fluctuations commonly observed in financial markets [26].

With the advancement of deep learning, neural network-based time-series modeling methods have gradually become mainstream [27]. Recurrent neural networks model temporal dependencies by recursively propagating hidden states, while long short-term memory networks alleviate the gradient vanishing problem through gating mechanisms, enabling the capture of longer-term dependencies [28]. Furthermore, Transformer models based on self-attention mechanisms break the limitations of traditional sequential models by modeling global dependencies, demonstrating stronger representational capabilities in complex time-series forecasting tasks [29]. These methods share the common characteristic of learning latent representations through high-dimensional nonlinear mappings, thereby improving predictive performance [30].

Building upon these developments, multimodal market modeling has emerged as a research hotspot. Its core idea is to integrate information from diverse sources, such as price sequences, trading volume, macroeconomic indicators, and textual data, to achieve a more comprehensive representation of market dynamics [31]. Typical multimodal fusion strategies include feature-level fusion, decision-level fusion, and attention-based dynamic fusion mechanisms. For example, textual sentiment features can be concatenated with price sequences as inputs to neural networks, or cross-modal attention can be employed to learn interactions between different modalities [32]. Such approaches have improved the adaptability of models to complex market environments to a certain extent. However, existing multimodal financial modeling methods still exhibit notable limitations. First, most studies remain dominated by structured numerical data, with textual information treated merely as auxiliary features, thereby failing to fully exploit its leading role in risk propagation. Second, in cross-market scenarios, information sources from different regions exhibit significant heterogeneity, and current methods lack a unified semantic space to integrate multilingual information. Finally, existing fusion strategies are often static or shallow, making it difficult to capture the dynamic coupling relationship between semantic information and market fluctuations [33]. Therefore, incorporating cross-lingual semantic modeling into multimodal frameworks and achieving deep interaction between semantic and market signals remains a critical challenge in financial risk forecasting.

2.2. Financial Text Analysis and Sentiment Perception Methods

Financial text analysis aims to extract predictive information from unstructured textual data, based on the principle of transforming textual content into quantifiable feature representations and modeling their relationships with market behavior to predict price movements or risk states [34]. Early approaches were primarily based on sentiment lexicons, where sentiment indicators were constructed by counting the frequency of positive and negative words in text [35]. Although such methods rely on manually constructed lexicons and offer certain interpretability, their expressive capacity is limited, making them inadequate for handling complex semantic structures and contextual dependencies. With the development of natural language processing, text modeling methods based on distributed representations have gained increasing attention [36]. Word embedding models map words into continuous vector spaces, allowing semantically similar words to be located closer to each other, thereby providing a foundation for text modeling [37]. Building on this, deep learning-based language models, including convolutional neural networks and recurrent neural networks, have been widely applied in sentiment classification and text prediction tasks. Furthermore, pretrained language models, which are trained on large-scale corpora through unsupervised learning, can generate context-aware representations with rich semantic information. Among them, models such as BERT, which employ bidirectional encoding mechanisms, have significantly improved semantic understanding and achieved promising results in financial text analysis tasks [38]. Nevertheless, existing financial text analysis methods still face several critical challenges [39]. First, most studies are conducted in monolingual environments, with model training and application restricted to specific languages, making them unsuitable for handling multilingual inputs commonly found in global markets. Second, significant differences exist in expression patterns and semantic structures across languages, and direct translation or simple alignment often introduces information loss or semantic shifts. Third, current approaches typically rely on simple concatenation or independent modeling of textual features and market data, lacking in-depth modeling of the relationship between dynamic semantic changes and market fluctuations [40]. Therefore, overcoming the limitations of monolingual settings, achieving unified cross-lingual semantic modeling, and deeply integrating textual information with market sensor signals are crucial for improving financial risk forecasting performance [41].

2.3. Multilingual Large Language Models and Cross-Lingual Semantic Alignment

The development of multilingual large language models provides a new technical paradigm for cross-lingual semantic modeling. The fundamental principle is to perform joint pretraining on large-scale multilingual corpora, enabling the model to learn shared semantic structures across different languages and map them into a unified representation space. During this process, tasks such as masked language modeling, translation alignment, and cross-lingual contrastive learning are typically employed to establish implicit semantic correspondences between languages. This capability allows multilingual models to process information from multiple languages in a unified manner without requiring explicit translation [42]. In practical applications, multilingual models have been widely used in cross-lingual information retrieval, machine translation, and multilingual question answering tasks [43]. For example, in cross-lingual retrieval, queries and documents in different languages can be mapped into a shared vector space to enable semantic matching, while in machine translation, models learn correspondences between languages to achieve high-quality text transformation. These applications demonstrate the advantages of multilingual models in semantic alignment and knowledge transfer. However, directly applying multilingual large language models to financial risk modeling still presents several challenges [44]. First, the financial domain involves highly specialized terminology, and expressions across languages may differ significantly, leading to potential semantic misalignment even within a unified representation space. Second, existing multilingual models mainly focus on static semantic representations, whereas financial risk is closely associated with temporal semantic evolution, and effective modeling of such dynamics remains insufficient. Moreover, multilingual models are primarily designed for text processing, while financial risk prediction requires the integration of market sensor data, making cross-modal fusion an open problem [45]. Therefore, although multilingual large language models provide a solid foundation for cross-lingual semantic alignment, their potential in financial risk forecasting has not yet been fully explored. Particularly in cross-market scenarios, constructing a unified risk semantic space and achieving effective coupling between semantic information and market dynamics remain key research directions [46]. Under this background, the present study attempts to deeply integrate multilingual semantic representations with multimodal financial data to achieve intelligent perception and early warning of cross-market financial risks [19]. To provide a clearer comparison of existing methods and their limitations, a structured summary of representative studies is presented in

Table 1. Comparison of existing methods and research gaps.

Method	Modality	Cross-Lingual	Fusion	Limitations
Statistical (ARIMA, GARCH)	Time series	No	No	Weak nonlinear modeling
Deep learning (LSTM, Trans.)	Time + text	Partial	Limited	Weak alignment
PLMs (BERT, mBERT)	Text	Yes	No	No temporal modeling
Multimodal (CNN+Text)	Multi	Partial	Yes	Weak dynamics
Cross-lingual multi (XLM)	Multi	Yes	Yes	Limited generalization
Proposed	Multi + temporal	Yes	Yes	Better alignment

. From multiple perspectives, including data modality, cross-lingual modeling capability, and multimodal fusion strategies, it can be observed that existing approaches have achieved progress in specific aspects but still exhibit notable limitations in cross-lingual semantic consistency and dynamic multimodal coupling. In particular, statistical models struggle to capture complex nonlinear relationships, pretrained language models lack temporal modeling capability, and existing multimodal approaches remain insufficient in cross-modal alignment and cross-lingual generalization. These limitations highlight the need for a unified framework that can effectively integrate cross-lingual semantic representation with dynamic multimodal fusion, thereby enabling more robust modeling in complex real-world scenarios.

3. Materials and Method

3.1. Data Collection

The cross-market multilingual risk prediction dataset constructed in this study is derived from multiple heterogeneous data sources, aiming to characterize market dynamics and risk evolution under different linguistic environments. The dataset spans the period from 2018 to 2024 and is uniformly resampled to a daily temporal granularity. The overall data system consists of two main components, namely market sensor signals and multilingual textual information, as summarized in

Table 2. Statistics of multimodal multilingual dataset before and after preprocessing and augmentation.

Data Type	Data Source	Raw Volume	After Preprocessing	After Augmentation
High-frequency price data	NYSE, NASDAQ, SSE, SZSE APIs	$5.2\times {10}^8$	$3.9\times {10}^8$	$4.5\times {10}^8$
Return and volatility series	Wind, Refinitiv, Bloomberg	$1.8\times {10}^7$	$1.4\times {10}^7$	$1.6\times {10}^7$
Trading volume and order book data	Exchange APIs	$1.3\times {10}^8$	$1.0\times {10}^8$	$1.2\times {10}^8$
Macroeconomic indicators	World Bank, IMF, Central Banks	$2.5\times {10}^5$	$2.1\times {10}^5$	$2.3\times {10}^5$
Chinese financial news	Sina Finance, Eastmoney, CSRC	1,250,000	980,000	1,350,000
English financial reports	Bloomberg, Reuters, SEC filings	980,000	760,000	1,120,000
European policy documents	ECB, EU Commission	420,000	330,000	470,000
Social financial texts	Twitter, financial forums	2,300,000	1,850,000	2,600,000
Total textual data	Aggregated multilingual sources	4,950,000	3,920,000	5,540,000

. Market sensor data cover major global markets, including the United States (NYSE, NASDAQ), China (SSE, SZSE), and Europe, and are collected from public exchange APIs and professional financial data platforms. To address inconsistencies in data frequency across markets, unified resampling and missing value imputation strategies are applied to ensure temporal consistency and comparability.

The construction of multilingual textual data emphasizes risk relevance rather than mere scale. A multilingual keyword system is developed, including asset categories, policy-related terms, sentiment expressions, and risk trigger phrases, combined with an event-driven filtering mechanism to remove irrelevant noise. The dataset includes Chinese, English, and European language texts, accounting for approximately 40%, 40%, and 20%, respectively. Original language forms are preserved to avoid semantic distortion. Key metadata such as timestamps, sources, and topic labels are consistently extracted and retained. These textual data are then aligned with market sensor signals through precise timestamp matching.

In the sample construction stage, a sliding window strategy with a window length of $T=10$ is employed to segment and reorganize multi-source data into aligned semantic–market samples. Risk labels are generated through a hybrid approach combining statistical methods and event-driven verification. Specifically, high-risk intervals are identified based on return and volatility quantiles, while major market shock events are incorporated for manual correction. Through these procedures, a multimodal dataset with strong consistency across temporal, semantic, and market dimensions is established, providing a reliable foundation for model training and cross-market generalization analysis.

3.2. Data Preprocessing and Augmentation Strategy

In cross-lingual multimodal financial risk modeling tasks, the primary objective of data preprocessing and data augmentation is to construct unified, alignable, and temporally consistent input representations, thereby providing high-quality feature foundations for subsequent model learning. Since the raw data originate from multilingual textual sources and multidimensional market sensor signals, significant discrepancies exist in terms of distribution forms, temporal scales, and semantic expressions. Therefore, a systematic preprocessing pipeline is required to achieve unified modeling across languages and modalities. In general, this process can be decomposed into four interrelated components, namely textual semantic normalization, multilingual embedding extraction, temporal sample construction, and low-resource language augmentation. The essence of this pipeline lies in the coordinated optimization of “semantic consistency” and “temporal alignment” within a unified semantic space. First, during multilingual text processing, raw texts typically contain noise such as HTML tags, special symbols, and irrelevant tokens, while different languages exhibit substantial differences in lexical structures and grammatical rules. Consequently, standardized text cleaning is performed, followed by language-aware tokenization to map text sequences into fundamental semantic units. Let the raw text corresponding to language l at time t be denoted as $T_t^{\left(l\right)}$. After cleaning and tokenization, it can be represented as a word sequence $W_t^{\left(l\right)}=\{w_{t,1}^{\left(l\right)},w_{t,2}^{\left(l\right)},\dots ,w_{t,n}^{\left(l\right)}\}$. Based on this representation, a multilingual semantic encoding function $f_{\theta }(·)$ is introduced to map discrete word sequences into continuous semantic vector representations, thereby achieving cross-lingual unification. This process can be formalized as:

$${\mathbf{e}}_t^{\left(l\right)}=f_{\theta }\left(W_t^{\left(l\right)}\right),$$

(1)

where ${\mathbf{e}}_t^{\left(l\right)}\in {\mathbb{R}}^d$ denotes the text embedding vector in the unified semantic space, and d represents the embedding dimension. This mapping process relies on the pretraining capability of multilingual large language models on large-scale corpora, enabling semantic alignment across different languages within a shared vector space and providing a foundation for cross-lingual risk modeling. After obtaining semantic embeddings, the temporal alignment between textual information and market sensor data must be addressed. Financial market data, such as prices and volatility, are typically represented as high-frequency time series, whereas textual information follows irregular temporal distributions. Therefore, a sliding time window mechanism is adopted to organize these heterogeneous data sources into a unified structure. Let the market sensor data be denoted as ${\mathbf{x}}_t\in {\mathbb{R}}^m$, where m represents the dimensionality of sensor features. Given a window length k, the temporal sample sequences can be constructed as:

$${\mathbf{X}}_t=\{{\mathbf{x}}_{t-k+1},{\mathbf{x}}_{t-k+2},\dots ,{\mathbf{x}}_t\},{\mathbf{E}}_t=\{{\mathbf{e}}_{t-k+1},{\mathbf{e}}_{t-k+2},\dots ,{\mathbf{e}}_t\},$$

(2)

where ${\mathbf{E}}_t$ denotes the sequence of semantic embeddings within the corresponding time window. Through this sliding window construction, asynchronous raw data are transformed into a unified temporal input structure, enabling the model to simultaneously capture semantic evolution and market dynamics. Furthermore, to characterize the dynamic nature of semantic information, a semantic drift variable is defined to measure the difference between semantic representations at adjacent time steps:

$$\Delta {\mathbf{e}}_t={\mathbf{e}}_t-{\mathbf{e}}_{t-1}.$$

(3)

This variable reflects the degree of semantic change and often exhibits significant variation prior to risk events, thereby providing critical signals for subsequent risk prediction. To address the issue of data scarcity in low-resource language scenarios, a semantic consistency-driven data augmentation strategy is further introduced. The core idea is to generate diversified samples through cross-lingual mapping and semantic perturbation while preserving semantic invariance, thereby improving model generalization capability. Specifically, for semantically equivalent content expressed in different languages, a multilingual alignment mechanism is employed to constrain their embedding representations to be consistent in the semantic space. Let ${\mathbf{e}}_t^{\left(l_1\right)}$ and ${\mathbf{e}}_t^{\left(l_2\right)}$ denote the semantic representations in two different languages; their alignment can be achieved by minimizing the distance between them:

$${\mathcal{L}}_{align}={∥{\mathbf{e}}_t^{\left(l_1\right)}-{\mathbf{e}}_t^{\left(l_2\right)}∥}_2^2.$$

(4)

This alignment loss effectively reduces cross-lingual semantic shifts, enabling the model to learn language-invariant risk representations. In addition, to enhance robustness against semantic perturbations, a semantic-preserving reconstruction mechanism is introduced. By adding a small perturbation $\mathit{ϵ}$ to the original semantic vector, an augmented sample ${\tilde{\mathbf{e}}}_t={\mathbf{e}}_t+\mathit{ϵ}$ is generated, and consistency constraints are imposed to maintain stable predictions:

$${\mathcal{L}}_{cons}={∥f_{\varphi }\left({\mathbf{e}}_t\right)-f_{\varphi }\left({\tilde{\mathbf{e}}}_t\right)∥}_2^2,$$

(5)

where $f_{\varphi }(·)$ denotes the risk prediction function. This strategy improves model stability under input perturbations and enhances generalization performance in low-resource language environments. Overall, the objective of data preprocessing and augmentation can be formulated as a joint optimization problem, where robust cross-lingual representations are learned under the constraints of semantic alignment and temporal consistency:

$$\mathcal{L}={\mathcal{L}}_{task}+{\lambda }_1{\mathcal{L}}_{align}+{\lambda }_2{\mathcal{L}}_{cons},$$

(6)

where ${\mathcal{L}}_{task}$ denotes the task-specific risk prediction loss, and ${\lambda }_1$ and ${\lambda }_2$ are weighting parameters. Through this joint optimization framework, collaborative modeling of multilingual textual information and market sensor data is achieved within a unified semantic space, thereby providing high-quality inputs for the subsequent semantic–volatility coupling mechanism. To provide a clearer and more intuitive understanding of the overall data processing pipeline, a schematic illustration is presented in Figure 1. This figure systematically summarizes the complete workflow, including multi-source data inputs (i.e., market data, macroeconomic data, and multilingual textual information), preprocessing operations, temporal alignment, and feature construction. Such a structured visualization helps clarify the logical relationships among different processing stages and establishes a coherent data flow foundation for the subsequent modeling framework.

3.3. Proposed Method

This section presents the overall framework and key methodological components proposed in this study. An overview of the model architecture is first provided, clarifying the functional roles of each component and the data flow across modules. Subsequently, detailed descriptions of the cross-lingual semantic alignment mechanism, multimodal temporal fusion strategy, and risk prediction module are introduced. In addition, the interactions among these components are analyzed to illustrate how the proposed framework effectively adapts to complex cross-market environments. This structured presentation is intended to provide a clear and comprehensive understanding of the methodological design and its underlying rationale.

3.3.1. Overall

At the model design level, the proposed method takes processed multilingual textual semantic representations and synchronously aligned market sensor temporal features as unified inputs, and follows a progressive modeling pipeline consisting of “cross-lingual semantic encoding–semantic–volatility coupling fusion–risk early warning output”. First, in the cross-lingual semantic encoding layer, news, announcements, and commentary texts in different languages are fed into a multilingual large language model and mapped into semantic vectors located in the same latent space. Subsequently, a unified risk semantic mapping mechanism is introduced to further constrain text representations with similar risk meanings across different languages to remain close in the embedding space, thereby producing window-level semantic representations with the characteristics of “language invariance” and “risk consistency”. Meanwhile, market sensor signals within the same time window, including prices, returns, volatility, trading volume, and macroeconomic indicators, are encoded into temporal feature sequences while preserving their local volatility intensity and dynamic variation trends. Thereafter, semantic representations and market features are jointly input into the semantic–volatility coupling fusion layer. In this layer, the magnitude of semantic changes between adjacent time windows is computed to capture whether abrupt risk-related semantic shifts have occurred, while volatility-enhanced signals are extracted from the market sensor sequence, and both are jointly used to generate dynamic coupling weights. Based on these weights, the contribution of textual semantics and market volatility in the fusion process can be adaptively adjusted, allowing semantic shifts to receive greater attention during high-volatility periods and thus forming a temporally sensitive cross-modal fused representation. After the unified fused risk features are obtained, they are fed into the risk early warning prediction layer, where future-window risk probabilities or continuous risk intensity values are generated through fully connected mapping and temporal dependency modeling, thereby enabling early identification of high-volatility periods, market downturn intervals, and other risk states. Furthermore, in order to improve transferability across markets with different languages, cross-lingual alignment constraints and knowledge transfer mechanisms are incorporated into the overall framework during training so that risk discrimination knowledge learned from high-resource markets can be transferred to low-resource language markets. As a result, the entire method forms a complete closed loop from unified semantic modeling to cross-modal dynamic fusion and finally to cross-market risk early warning output, enabling the model to both understand deep risk signals in multilingual financial texts and achieve more robust cross-market early warning by integrating market volatility structures.

3.3.2. Cross-Lingual Unified Risk Semantic Mapping Module

In the cross-lingual unified risk semantic mapping module, the problem of multilingual semantic alignment is formulated as a directional consistency learning problem in the latent space. Specifically, let the hidden representation obtained at the T-th layer of a pretrained multilingual large language model for multilingual text be denoted as ${\mathbf{h}}^{\left(l\right)}\left(t\right)\in {\mathbb{R}}^d$, where l denotes the language type. For text pairs within the same time window that are semantically equivalent but expressed in different languages, the corresponding representations are denoted as ${\mathbf{h}}^{\left(l_1\right)}\left(t\right)$ and ${\mathbf{h}}^{\left(l_2\right)}\left(t\right)$, respectively. Their semantic offset can then be represented by the difference vector $\Delta {\mathbf{h}}_i\left(t\right)={\mathbf{h}}_i^{\left(l_1\right)}\left(t\right)-{\mathbf{h}}_i^{\left(l_2\right)}\left(t\right)$. Based on this, N semantically matched samples are aggregated to construct a global semantic alignment direction $\mathbf{v}\left(t\right)={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N\Delta {\mathbf{h}}_i\left(t\right)$, which can be regarded as an “offset compensation direction” in the cross-lingual semantic space and is used to characterize systematic differences in risk semantics across different language expressions.

As shown in Figure 2, the semantic encoding layer adopts a multilingual encoder $f_{\theta }(·)$ with shared parameters.

Its output is transformed through a linear projection $W_s\in {\mathbb{R}}^{d\times d}$ followed by normalization to obtain a unified semantic representation $\mathbf{s}\left(t\right)=\mathrm{Norm}\left(W_s\mathbf{h}\left(t\right)\right)$, where d denotes the latent space dimension, which is set to $d=768$ or 1024 in the experiments. To further strengthen semantic alignment capability, a contrastive learning loss is introduced during training so that the similarity of positive cross-lingual sample pairs is maximized while the similarity to irrelevant samples is suppressed, which can be expressed as ${\mathcal{L}}_{align}=-\log {\displaystyle \frac{\exp (\cos ({\mathbf{s}}_i,{\mathbf{s}}_j)/\tau )}{{\sum }_k\exp (\cos ({\mathbf{s}}_i,{\mathbf{s}}_k)/\tau )}}$. Meanwhile, a directional constraint based on the difference vector is incorporated by injecting $\mathbf{v}\left(t\right)$ into the semantic representation, yielding the corrected representation ${\mathbf{s}}^{\prime }\left(t\right)=\mathbf{s}\left(t\right)+\alpha ·\mathbf{v}\left(t\right)$, where $\alpha$ is a learnable coefficient used to control the strength of cross-lingual alignment. From a mathematical perspective, semantic space alignment is achieved through two mechanisms in this module. The first is a similarity-based local alignment constraint, which encourages different languages to form compact clusters under the same risk semantics. The second is the establishment of a linear mapping relationship between cross-lingual distributions through the global offset vector $\mathbf{v}\left(t\right)$, which is equivalent to learning a low-rank alignment transformation in the embedding space, thereby effectively reducing the systematic offset introduced by distributional differences between languages. Furthermore, since $\mathbf{v}\left(t\right)$ is derived from the statistical average of sample-level differences, it exhibits favorable stability under noise perturbation and can provide an additional alignment prior when low-resource language samples are insufficient. This design offers significant advantages for the task addressed in this study. First, by explicitly modeling the direction of semantic offset, semantic loss caused by direct reliance on machine translation is avoided, allowing alignment to be completed in the original language space. Second, the difference aggregation mechanism captures common structural patterns in cross-lingual risk expressions, making risk signals in different languages comparable within a unified space. Finally, by injecting the alignment vector into semantic representations, the subsequent cross-modal fusion module can directly exploit aligned semantic features, thereby reducing inter-modal noise interference and improving the stability and generalization capability of risk early warning. Overall, this module achieves a structurally unified representation of cross-lingual risk information while preserving semantic expressiveness, thus providing a solid foundation for subsequent coupling modeling.

From a theoretical perspective, the proposed cross-lingual semantic alignment and multimodal temporal fusion mechanisms offer clear advantages over conventional approaches. First, traditional cross-lingual methods typically rely on static embedding alignment, which is insufficient to capture dynamic semantic variations across languages. In contrast, the proposed dynamic alignment mechanism enables semantic representations to adapt to temporal and contextual changes, thereby mitigating cross-lingual distribution discrepancies. Second, existing multimodal fusion strategies often adopt simple concatenation or static weighting schemes, which fail to capture complex inter-modal dependencies and temporal dynamics. The proposed approach explicitly models these interactions and temporal relationships, leading to more efficient information integration and improved prediction stability. Therefore, the proposed framework provides a more consistent and robust representation learning paradigm from a theoretical standpoint.

3.3.3. Semantic–Volatility Coupling Attention Mechanism

In the semantic–volatility coupling attention mechanism, a cross-modal attention network with structured coupling relationships is introduced on the basis of cross-lingual semantic mapping. Its core idea is to explicitly model the collaborative interaction between semantic change intensity and market volatility structure, thereby achieving dynamic enhancement of risk-triggering signals. First, multi-layer feature encoding channels are constructed separately on the semantic side and the market side. The semantic feature ${\mathbf{s}}_t\in {\mathbb{R}}^d$ is mapped into a high-order representation ${\tilde{\mathbf{s}}}_t\in {\mathbb{R}}^d$ through a two-layer feed-forward network, whose structure consists of a linear layer with input dimension d and output dimension $2d$, followed by GELU activation, and then another linear layer that projects the feature back to dimension d. The market sensor sequence ${\mathbf{x}}_t\in {\mathbb{R}}^{T\times F}$ is processed by a one-dimensional convolutional network to extract local volatility features, where the convolution kernel size is set to 3, and the number of channels is increased from F to $4F$. Stable representations ${\tilde{\mathbf{x}}}_t\in {\mathbb{R}}^{T\times 4F}$ are then formed through layer normalization and residual connections.

As shown in Figure 3, in the coupling stage, a semantic change operator is first constructed, and semantic mutation features are extracted through a difference operation and a gating mechanism:

$${\mathbf{g}}_t=\sigma \left(W_g·({\mathbf{s}}_t\odot {\mathbf{s}}_{t-1})\right),{\mathbf{z}}_t={\mathbf{g}}_t\odot ({\mathbf{s}}_t-{\mathbf{s}}_{t-1}),$$

(7)

where $\sigma (·)$ denotes the Sigmoid function, ⊙ denotes element-wise multiplication, and ${\mathbf{g}}_t$ is used to control the validity of semantic changes. Meanwhile, a volatility structure tensor is extracted from market features:

$${\mathbf{v}}_t=\mathrm{Conv}1\mathrm{D}\left({\mathbf{x}}_t\right)+\mathrm{Pool}\left({\mathbf{x}}_t\right).$$

(8)

This structure integrates local volatility and global trend information. Subsequently, the semantic change feature ${\mathbf{z}}_t$ is projected into a Query vector, while the market feature ${\mathbf{v}}_t$ is projected into Key and Value vectors, thereby forming cross-modal attention:

$${\mathbf{Q}}_t=W_Q{\mathbf{z}}_t,{\mathbf{K}}_t=W_K{\mathbf{v}}_t,{\mathbf{V}}_t=W_V{\mathbf{v}}_t,$$

(9)

$${\mathbf{A}}_t=\mathrm{Softmax}\left({\displaystyle \frac{{\mathbf{Q}}_t{\mathbf{K}}_t^{\top }}{\sqrt{d}}}\right),{\mathbf{h}}_t={\mathbf{A}}_t{\mathbf{V}}_t.$$

(10)

To further enhance the representation of critical risk signals, an activation reconstruction mechanism based on a sparse autoencoder (SAE) is introduced to sparsely encode the attention output:

$${\mathbf{u}}_t=\varphi \left(W_e{\mathbf{h}}_t\right),{\widehat{\mathbf{h}}}_t=W_d{\mathbf{u}}_t,$$

(11)

where $\varphi (·)$ denotes a sparse activation function, and $W_e$ and $W_d$ denote the encoder and decoder parameters, respectively. By minimizing the reconstruction error and the sparse regularization term,

$${\mathcal{L}}_{\mathrm{sparse}}=\parallel {\mathbf{h}}_t-{\widehat{\mathbf{h}}}_t{\parallel }_2^2+\beta {\parallel {\mathbf{u}}_t\parallel }_1,$$

(12)

this process is equivalent to identifying a set of low-dimensional sparse bases in the feature space such that important risk patterns are explicitly activated, thereby improving model interpretability and noise resistance. Furthermore, according to the definition of the attention weight matrix, its essence lies in a similarity mapping between semantic changes and market volatility. When ${\mathbf{z}}_t$ is highly correlated with certain volatility patterns, the corresponding weights are significantly increased, thus allowing the output feature ${\mathbf{h}}_t$ to exhibit stronger responses along those directions. This guarantees that the model can generate an amplification effect when semantic mutations and market anomalies co-occur. From a theoretical perspective, this mechanism transforms simple weighted fusion into structured coupling modeling. By introducing semantic change gating, cross-modal attention, and sparse activation reconstruction, the model is enabled to simultaneously capture the bidirectional relationships of “semantic driving” and “volatility response”. For the task considered in this study, such a design effectively improves sensitivity to early risk signals, particularly when textual semantics have not yet fully diffused but local semantic shifts have already emerged, allowing potential risks to be captured in advance through the coupling mechanism and thus enabling more forward-looking cross-market early warning. In addition, the introduction of sparse representations endows key risk patterns with stronger transferability across markets, thereby contributing to improved generalization performance of the overall model in multilingual and multi-market environments.

3.3.4. Cross-Market Generalization and Low-Resource Enhancement Module

In the cross-market generalization and low-resource enhancement module, a unified modeling structure integrating “distribution correction–cross-market relational embedding–knowledge transfer enhancement” is designed to address the issues of imbalanced sample sizes, inconsistent statistical distributions, and substantial differences in risk-triggering patterns across markets with different languages. The input to this module is the joint representation sequence $\mathbf{H}\in {\mathbb{R}}^{T\times d}$ obtained from the semantic–volatility coupling fusion layer, where T denotes the time window length, and d denotes the dimension of the fused features. First, to alleviate the interference caused by scale differences and volatility amplitude shifts across markets, instance normalization is independently applied to each market sample, resulting in the standardized representation

$$\tilde{\mathbf{H}}={\displaystyle \frac{\mathbf{H}-\mu \left(\mathbf{H}\right)}{\sigma \left(\mathbf{H}\right)+ϵ}},$$

(13)

where $\mu \left(\mathbf{H}\right)$ and $\sigma \left(\mathbf{H}\right)$ denote the mean and standard deviation of the current sample along the temporal dimension, respectively, and $ϵ$ is a stability term. This design prevents the model from over-relying on the inherent numerical scale of a particular market and instead encourages greater attention to the relative structural characteristics of risk evolution.

As shown in Figure 4, at the feature modeling stage, the module adopts a two-level structure consisting of “cross-market relational embedding + patch representation learning”.

For the normalized input $\tilde{\mathbf{H}}$, local temporal dependencies and cross-variable correlations are first extracted through one-dimensional convolution. The convolution layer uses temporal convolution with kernel size 3, stride 1, and padding 1, with the input channel number equal to d and the output channel number expanded to $2d$, thereby generating local interaction features $\mathbf{Z}\in {\mathbb{R}}^{T\times 2d}$. Subsequently, the original normalized features and the convolution output are fused in a residual weighted manner to construct the cross-market relational embedding:

$$\mathbf{E}=\alpha ·P\left(\tilde{\mathbf{H}}\right)+(1-\alpha )·\mathbf{Z},$$

(14)

where $P(·)$ denotes a linear projection operator used to map $\tilde{\mathbf{H}}$ into the same dimensional space as $\mathbf{Z}$, and $\alpha$ is a learnable gating coefficient. This embedding preserves both the original dynamic risk patterns and cross-variable correlation structures, thereby helping to improve representational consistency across different markets. To further enhance temporal generalization in low-resource markets, the concept of patch embedding is adopted, and $\mathbf{E}$ is segmented into multiple local temporal blocks of fixed length, which are then mapped into higher-level token representations. Let the patch-divided sequence be denoted as $\{{\mathbf{e}}_1,{\mathbf{e}}_2,\dots ,{\mathbf{e}}_M\}$. Token representations ${\mathbf{p}}_m$ are obtained through linear projection, and learnable positional encodings ${\mathbf{r}}_m$ are further added:

$${\mathbf{u}}_m=W_p{\mathbf{e}}_m+\beta {\mathbf{r}}_m,$$

(15)

where $W_p$ denotes the patch projection matrix, and $\beta$ is the scaling factor for positional encoding. The core role of this design is to compress fine-grained fluctuations over long time series into several structured subsegments, allowing the model to capture the repetitiveness and commonality of local risk patterns even in markets with limited data. For cross-market transfer learning, a teacher–student knowledge distillation mechanism is further introduced. Let the output of the teacher network trained on high-resource markets be denoted as ${\mathbf{q}}^{tea}$, and let the output of the student network for low-resource markets be denoted as ${\mathbf{q}}^{stu}$. The distillation constraint is then written as

$${\mathcal{L}}_{\mathrm{kd}}={\displaystyle \sum _c}{\mathbf{q}}_c^{tea}\log {\displaystyle \frac{{\mathbf{q}}_c^{tea}}{{\mathbf{q}}_c^{stu}}}.$$

(16)

At the same time, to reduce the overall feature distribution shift across markets, a distribution matching term is introduced at the joint representation level:

$${\mathcal{L}}_{\mathrm{dist}}={∥{\overline{\mathbf{u}}}^{\left(h\right)}-{\overline{\mathbf{u}}}^{\left(l\right)}∥}_2^2+{∥{\mathbf{C}}^{\left(h\right)}-{\mathbf{C}}^{\left(l\right)}∥}_F^2,$$

(17)

where ${\overline{\mathbf{u}}}^{\left(h\right)}$ and ${\overline{\mathbf{u}}}^{\left(l\right)}$ denote the mean token representations of high-resource and low-resource markets, respectively, and ${\mathbf{C}}^{\left(h\right)}$ and ${\mathbf{C}}^{\left(l\right)}$ represent their covariance structures. This term constrains cross-market representation distributions to become consistent from the perspectives of both mean and correlation, thereby theoretically reducing the upper bound of generalization error in the target market. This design exhibits clear advantages for the task considered in this study. On the one hand, instance normalization removes amplitude differences caused by institutional settings, currencies, and trading activity across countries and exchanges, enabling the model to learn more stable relative risk patterns. On the other hand, cross-market relational embedding preserves local correlation structures in risk propagation through convolution extraction and residual fusion, making similar risk phases across different markets easier to align in the feature space. Furthermore, patch-based representation partitions long sequences into transferable local risk units, which is particularly suitable for low-resource language markets where data are limited but patterns remain reusable. Finally, the combined use of distillation and distribution matching ensures that risk discrimination knowledge learned in high-resource markets can be smoothly transferred to low-resource markets, thereby significantly improving the stability, robustness, and generalization capability of the cross-market risk early warning model under small-sample conditions.

4. Results and Discussion

4.1. Experimental Configuration

4.1.1. Hardware and Software Platform

In terms of hardware, the experiments were conducted on a high-performance computing environment equipped with NVIDIA RTX 4090 GPUs (24 GB memory; NVIDIA Corporation, Santa Clara, CA, USA) and CUDA 12.1. Each computing node was configured with Intel Xeon multi-core processors (approximately 2.7 GHz; Intel Corporation, Santa Clara, CA, USA), 128 GB system memory, and NVMe solid-state drives to ensure efficient data processing and I/O performance. From the software perspective, the experiments were implemented on Ubuntu 20.04 using the PyTorch 2.0 deep learning framework. Model development and data processing were carried out in Python 3.14, with supporting libraries including NumPy 2.2.5, Pandas 3.0, and scikit-learn 1.7.2. Multilingual semantic representations were obtained using pretrained language models (e.g., mBERT), which were further fine-tuned on the task-specific dataset. TensorBoard 2.19 was employed to monitor training dynamics and evaluation metrics. Regarding model configurations, the cross-lingual encoder was implemented as a 12-layer Transformer with a hidden dimension of $d=768$ and 12 attention heads. The multimodal fusion module adopted an attention-based fusion mechanism to model inter-modal interactions. The dataset was split into training, validation, and test sets in chronological order with proportions of 70%, 15%, and 15%, respectively. The Adam optimizer was used with an initial learning rate of $1\times {10}^{-4}$ and a cosine annealing schedule, along with a weight decay of $1\times {10}^{-5}$. The batch size was set to 32, and the maximum number of training epochs was 100. Early stopping was applied if no improvement was observed on the validation set for 10 consecutive epochs. In addition, five-fold cross-validation was conducted to enhance the robustness and reliability of the results.

4.1.2. Baseline Models and Evaluation Metrics

The baseline models include ARIMA [47], SVR [48], LSTM [49], GRU [50], Transformer [51], BERT + LSTM (monolingual) [52], Multilingual BERT + LSTM [53], and Text + Numeric Fusion (Concat) [54]. The ARIMA model is based on autoregressive and moving average principles of time series, modeling historical sequences $\left\{x_t\right\}$ to achieve linear prediction of future risks, with advantages in simplicity and interpretability under stationary conditions. SVR introduces kernel functions to map inputs into high-dimensional spaces and constructs optimal margin regression models, enabling nonlinear relationship modeling with strong generalization performance for small-sample data. LSTM models temporal dependencies through gating mechanisms, including input, forget, and output gates, effectively capturing long-term dependencies and demonstrating stability in modeling long-sequence risk evolution. GRU, as a simplified variant of LSTM, reduces parameter complexity by merging gating structures, thereby improving computational efficiency while maintaining comparable performance. The Transformer model leverages self-attention mechanisms to model dependencies between arbitrary positions in a sequence, where weights ${\alpha }_{ij}$ capture relationships between different time steps, providing strong parallelization capability and effectiveness in modeling complex temporal structures. The BERT + LSTM (monolingual) approach first extracts semantic representations from single-language texts using BERT and subsequently models temporal dynamics via LSTM, thereby integrating semantic and temporal features to improve prediction accuracy. The Multilingual BERT + LSTM method extends this framework by introducing a multilingual semantic space, mapping different languages into unified representations before applying LSTM for temporal modeling, enabling cross-lingual information fusion suitable for global market analysis. The Text + Numeric Fusion (Concat) method concatenates textual feature vectors ${\mathbf{e}}_t$ with numerical features ${\mathbf{x}}_t$ to form a joint representation $[{\mathbf{e}}_t;{\mathbf{x}}_t]$, which is then fed into the prediction model, offering a simple yet effective way to utilize multimodal information for risk perception.

The evaluation metrics employed in this study include root mean square error (RMSE) to measure the overall deviation between predicted and true risk values, mean absolute error (MAE) to quantify the average magnitude of prediction errors, area under the curve (AUC) to evaluate the discriminative capability of risk classification models under different thresholds, and early warning gain to measure the advance time of risk detection. The computation of these metrics can be uniformly expressed as follows:

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{(y_i-{\widehat{y}}_i)}^2},\mathrm{MAE}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}|y_i-{\widehat{y}}_i|.$$

(18)

$$\mathrm{AUC}={\int }_0^{1}TPR\left(FP{R}^{-1}\left(x\right)\right)dx,$$

(19)

$$\Delta T={\displaystyle \frac{1}{M}}{\displaystyle \sum _{j=1}^M}(t_j^{pred}-t_j^{true}),$$

(20)

where N denotes the number of samples, $y_i$ and ${\widehat{y}}_i$ represent the ground-truth and predicted risk values of the i-th sample, respectively; $TPR$ and $FPR$ denote the true positive rate and false positive rate; $\Delta T$ represents the average early warning time, M denotes the number of risk events, $t_j^{pred}$ indicates the predicted occurrence time of the j-th risk event, and $t_j^{true}$ denotes the corresponding ground-truth occurrence time.

4.2. Main Experimental Results

This experiment aims to systematically evaluate the performance differences among various models in cross-lingual and cross-market risk forecasting tasks, with a particular focus on the contributions of semantic information, multimodal fusion, and cross-lingual modeling. To ensure the reliability of the results, each experiment is repeated five times, and both the mean and standard deviation are reported. Statistical significance is evaluated using a t-test ($p<0.05$). In addition, five-fold cross-validation is employed to assess model robustness and stability. Further analysis of error distribution and performance variability is also conducted to provide a more comprehensive and objective evaluation.

It can be observed that traditional statistical models such as ARIMA and machine learning models such as SVR exhibit the weakest performance across all metrics, with higher RMSE and MAE and significantly lower AUC and early warning gains, as shown in

Table 3. Main experimental results (mean $\pm$ std) with statistical significance (^† indicates $p<0.05$ compared to the best baseline).

Method	RMSE ↓	MAE ↓	AUC ↑	Early Warning Gain ↑
ARIMA	0.1568$\pm$0.0041	0.1194$\pm$0.0032	0.7421$\pm$0.0065	1.2$\pm$0.3
SVR	0.1493$\pm$0.0038	0.1142$\pm$0.0030	0.7567$\pm$0.0059	1.5$\pm$0.4
LSTM	0.1432$\pm$0.0035	0.1087$\pm$0.0028	0.7814$\pm$0.0051	1.8$\pm$0.4
GRU	0.1405$\pm$0.0033	0.1069$\pm$0.0026	0.7923$\pm$0.0048	2.0$\pm$0.3
Transformer	0.1365$\pm$0.0031	0.1031$\pm$0.0024	0.8042$\pm$0.0045	2.3$\pm$0.3
BERT + LSTM (Monolingual)	0.1298$\pm$0.0029	0.0974$\pm$0.0022	0.8267$\pm$0.0041	2.9$\pm$0.3
Multilingual BERT + LSTM	0.1256$\pm$0.0027	0.0948$\pm$0.0021	0.8395$\pm$0.0038	3.3$\pm$0.3
Text + Numeric Fusion (Concat)	0.1229$\pm$0.0026	0.0926$\pm$0.0020	0.8462$\pm$0.0036	3.6$\pm$0.2
XLM-R + Temporal Fusion	0.1204$\pm$0.0024	0.0908$\pm$0.0019	0.8547$\pm$0.0034	3.9$\pm$0.2
Multimodal Transformer (Cross-Attention)	0.1186$\pm$0.0023	0.0895$\pm$0.0018	0.8612$\pm$0.0032	4.2$\pm$0.2
Proposed Method	0.1127$\pm$0.0021 †	0.0846$\pm$0.0017 †	0.8879$\pm$0.0028 †	5.2$\pm$0.2 †

and Figure 5. This indicates that relying solely on historical numerical sequences is insufficient to capture complex nonlinear risk propagation mechanisms in financial markets. With the introduction of deep temporal models, the performance of LSTM, GRU, and Transformer gradually improves, where the Transformer demonstrates advantages in modeling long-range dependencies and thus achieves better AUC and early warning performance compared to recurrent architectures. After incorporating textual information, the BERT + LSTM model significantly reduces prediction errors and improves classification performance. The use of multilingual BERT further enhances generalization under cross-lingual scenarios, while the simple multimodal fusion model based on feature concatenation provides limited improvement due to the lack of deep interaction mechanisms. In contrast, the proposed method achieves the best performance across all metrics, particularly demonstrating a substantial advantage in early warning time, indicating its ability to capture potential risk signals at earlier stages. Furthermore, the incorporation of XLM-R leads to additional performance gains over multilingual BERT, highlighting the effectiveness of more advanced cross-lingual pretraining in capturing richer semantic alignment across languages. The Multimodal Transformer with cross-attention further improves the results by modeling deeper interactions between textual semantics and numerical market signals; however, it still falls short of the proposed method, indicating that simply enhancing fusion architecture is insufficient without explicitly modeling dynamic semantic–market coupling.

From a theoretical perspective, the performance differences among models stem from their varying capacities to model temporal dependencies, semantic information, and cross-modal relationships. ARIMA is inherently based on linear stationarity assumptions and thus fails to capture abrupt and non-stationary changes in financial markets. Although SVR introduces kernel functions to enhance nonlinear modeling, it still lacks the ability to model temporal structures and semantic information. LSTM and GRU capture temporal dependencies through gating mechanisms; however, their sequential modeling capability is constrained by recursive structures, leading to attenuation in long-range information propagation. The Transformer alleviates this limitation through global attention mechanisms, yet it still operates primarily on numerical sequences. With the introduction of BERT, semantic representations can be extracted from textual data, thereby supplementing latent risk information sources, and the multilingual variant further reduces information loss caused by language differences through a shared semantic space. However, simple fusion methods essentially perform feature-level concatenation and fail to model the coupling relationship between semantic information and market volatility. The proposed method constructs a unified semantic space, introduces a dynamic coupling mechanism between semantic changes and volatility intensity, and incorporates cross-market knowledge transfer, thereby achieving higher discriminative power and stability in feature representation, which leads to significant improvements in error reduction, risk identification, and early warning capability.

4.3. Cross-Lingual Generalization Performance and Robustness Capability

This experiment aims to evaluate the robustness and generalization capability of the model under cross-lingual and cross-market distribution shifts. Specifically, the model is trained solely on English market data and directly tested on Chinese and European markets, simulating realistic deployment scenarios in low-resource languages and unseen market environments. To ensure the reliability of the results, all experiments are repeated five times, and the mean and standard deviation are reported. In addition, the performance variation under cross-region transfer is analyzed to assess the robustness of the model.

The results indicate that the traditional temporal model LSTM exhibits a significant performance decline in cross-market testing, with relatively low AUC values in both Chinese and European markets and higher prediction errors, demonstrating its sensitivity to distribution shifts and limited cross-domain generalization capability, as shown in

Table 4. Cross-lingual generalization performance (mean $\pm$ std) under distribution shift (train on English, test on other markets).

Method	Chinese Market (AUC)	European Market (AUC)	RMSE ↓	Early Warning Gain ↑
LSTM	0.7523$\pm$0.0052	0.7411$\pm$0.0056	0.1487$\pm$0.0034	1.4$\pm$0.3
Transformer	0.7715$\pm$0.0048	0.7598$\pm$0.0051	0.1412$\pm$0.0031	1.8$\pm$0.3
Multilingual BERT + LSTM	0.8027$\pm$0.0042	0.7884$\pm$0.0046	0.1326$\pm$0.0028	2.5$\pm$0.2
Text + Numeric Fusion	0.8189$\pm$0.0039	0.8035$\pm$0.0042	0.1283$\pm$0.0026	2.9$\pm$0.2
Proposed Method	0.8624$\pm$0.0035	0.8471$\pm$0.0038	0.1185$\pm$0.0023	4.6$\pm$0.2

and Figure 6. The Transformer shows slight improvement due to its global dependency modeling capability, yet it remains insufficient to handle the dual heterogeneity of language and market structures. With the incorporation of multilingual textual information, the Multilingual BERT + LSTM model shows substantial improvement across both target markets, indicating that cross-lingual semantic representations can effectively mitigate language-induced discrepancies. The multimodal fusion model that integrates numerical features further improves performance, suggesting that multimodal information enhances model stability. However, the proposed method achieves the best results across all metrics, particularly in AUC and early warning capability, demonstrating strong risk recognition performance under cross-lingual conditions. From a theoretical perspective, these results reflect fundamental differences in feature representation and distribution alignment capabilities among models. LSTM and Transformer primarily rely on the distributional structure of training data, and the temporal patterns learned from one market are difficult to directly transfer to others, resulting in limited generalization. Multilingual BERT reduces semantic discrepancies by mapping texts from different languages into a shared representation space, yet it lacks deep interaction with market numerical signals, limiting its performance improvement. Simple multimodal fusion methods combine textual and numerical features but do not explicitly model cross-market distribution differences, thus constraining generalization performance. In contrast, the proposed method reduces cross-lingual representation shifts through semantic alignment at the representation level, enhances key risk signal expression through semantic–volatility coupling at the fusion level, and introduces cross-market distribution alignment and knowledge transfer mechanisms during training. This dual constraint at both the representation and distribution levels enables the model to maintain consistent discriminative capability across different languages and markets, thereby achieving superior cross-lingual generalization performance.

4.4. Ablation Study

This experiment aims to systematically analyze the contributions of each core component in the proposed framework by progressively removing key modules and evaluating their impact on model performance, thereby validating the role of each design in risk prediction and early warning.

The results demonstrate that the full model achieves the best performance across all metrics, indicating strong synergy among the modules, as shown in

Table 5. Ablation study of key components in the proposed framework.

Model Variant	RMSE ↓	MAE ↓	AUC ↑	Early Warning Gain ↑
Full Model	0.1127	0.0846	0.8879	5.2
w/o Cross-lingual Alignment	0.1213	0.0915	0.8534	3.9
w/o Semantic–Volatility Coupling	0.1258	0.0942	0.8417	3.5
w/o Cross-market Transfer Module	0.1236	0.0928	0.8479	3.7
Text Only	0.1276	0.0959	0.8348	3.1
Numeric Only	0.1339	0.1012	0.8115	2.1

and Figure 7. When the cross-lingual semantic alignment module is removed, performance degrades significantly across all metrics, particularly in AUC and early warning capability, highlighting the importance of semantic consistency for cross-lingual risk identification. Further removal of the semantic–volatility coupling mechanism leads to the largest performance decline, indicating that this module plays a central role in capturing the critical relationship between semantic changes and market volatility. Removing the cross-market transfer module also results in performance degradation, although to a lesser extent, suggesting that this component primarily contributes to improving generalization stability. In contrast, models using only textual or only numerical signals perform substantially worse, confirming the necessity of multimodal information fusion in risk modeling. From a theoretical perspective, the ablation results reflect the distinct roles of different modules in feature representation and information interaction. The cross-lingual semantic alignment module constrains representations from different languages to be consistent in the latent space, enabling the model to learn language-invariant risk semantics and reducing errors caused by input distribution shifts. When this module is removed, semantic representations become misaligned across languages, making it difficult for the model to establish a unified decision boundary. The semantic–volatility coupling mechanism introduces dynamic weighting, amplifying semantic changes during high-volatility periods and enhancing the model’s responsiveness to risk-triggering signals; without this mechanism, the model degenerates into a static fusion structure that cannot effectively capture critical cross-modal dependencies. The cross-market transfer module aligns feature distributions and incorporates knowledge transfer constraints, enabling consistent representation across markets and improving stability and generalization. When only a single modality is used, the feature space becomes limited and fails to jointly capture market behavior and semantic drivers, leading to significant performance degradation. Overall, these modules enhance model capability from three aspects: semantic consistency, cross-modal interaction, and cross-market distribution alignment, and their combination enables stronger representation power and robustness in complex financial environments.

4.5. Discussion

The proposed method demonstrates strong applicability in real-world financial scenarios, particularly in environments characterized by rapid cross-market information propagation and significant risk interdependence. Under the highly interconnected global capital market system, policy changes or sudden events are often disseminated rapidly through multilingual information channels. For example, interest rate decisions by the Federal Reserve, energy policy adjustments in Europe, or regulatory announcements in Asian markets may appear in different languages with varying expressions yet produce cascading effects across global markets. Traditional risk monitoring systems that rely on a single market or a single language are typically reactive, responding only after price fluctuations occur, and thus fail to capture early-stage signals in a timely manner. By integrating multilingual textual information with market sensor data, the proposed method enables early identification of potential risks at the stage of news release or sentiment shifts, thereby providing more forward-looking early warning support for investment institutions, risk management departments, and regulatory authorities. In practical applications, such as cross-border portfolio management, fund managers are required to monitor multiple markets simultaneously. When contractionary signals are released by the European Central Bank, relevant information may initially appear in European languages and subsequently propagate through English media to global markets. The proposed method can extract unified risk semantics from multilingual texts and combine them with corresponding market volatility dynamics to identify potential asset price adjustment trends in advance, thereby assisting investors in adjusting their positions. In another scenario, emerging markets or regions with strong information asymmetry often exhibit fragmented and limited textual data across different languages, making it difficult for traditional models to operate reliably. Through the cross-market knowledge transfer mechanism, risk patterns learned from high-resource markets can be transferred to low-resource markets, thereby enhancing overall risk identification capability. Furthermore, at the regulatory level, this method can be employed to construct multilingual financial risk monitoring platforms, enabling unified analysis of global policy changes, sentiment fluctuations, and market anomalies, and thereby improving the efficiency of systemic risk detection and response. Overall, the proposed method not only achieves improved model performance but also demonstrates strong interpretability and practical potential in real financial environments.

4.6. Limitation and Future Work

Although the proposed method achieves promising results in cross-lingual financial risk prediction, several limitations remain. First, the model relies on high-quality alignment between multilingual textual data and market signals, while in real-world scenarios, data from different sources may suffer from inaccurate timestamps, delayed information, or noise interference, which can affect the alignment accuracy between semantic and market signals. Second, despite the introduction of cross-market transfer mechanisms, the generalization capability of the model may still be limited in extremely low-resource language environments or markets with severe data scarcity. In addition, the current approach mainly focuses on time-series and textual semantic modeling, while structured relational information, such as capital flow networks and institutional behaviors, has not been fully exploited, which limits the ability to capture complex risk propagation pathways. Future research can be extended in several directions. On the one hand, graph-based modeling approaches can be introduced to incorporate cross-market capital flow relationships and industry correlation structures into a unified framework, thereby enhancing the modeling of systemic risk propagation mechanisms. On the other hand, more refined event-level modeling and causal inference methods can be integrated to improve the interpretability of risk-triggering mechanisms. In terms of data, additional multimodal information, such as sentiment propagation paths and trading behavior data, can be explored to construct a more comprehensive risk perception system. Finally, for real-world deployment, privacy-preserving and secure computation techniques can be incorporated to enable safe model operation under data-constrained environments, thereby further expanding its applicability in practical financial systems.

5. Conclusions

In the context of highly interconnected global financial markets and the rapid dissemination of multilingual information, a cross-market financial risk early warning framework based on multilingual large language models and multimodal financial sensing fusion is proposed to address the limitations of traditional risk prediction methods in integrating cross-lingual semantic information and heterogeneous market signals. The proposed approach is centered on a unified risk semantic space, where cross-lingual semantic alignment is introduced to mitigate semantic shifts across languages. Furthermore, a semantic–volatility coupling attention mechanism is designed to capture the dynamic interactions between textual semantic evolution and market fluctuations. In addition, cross-market knowledge transfer and low-resource enhancement strategies are incorporated to improve the model’s generalization capability across different linguistic and market environments. The overall framework enhances risk perception from three complementary perspectives, namely semantic modeling, cross-modal fusion, and cross-market distribution alignment, thereby demonstrating the potential of artificial intelligence-driven sensing techniques in complex financial sensing scenarios. Experimental results demonstrate that the proposed method significantly outperforms various baseline models in the multilingual cross-market financial risk forecasting task. In the main experiment, the model achieves an RMSE of $0.1127$, an MAE of $0.0846$, and an AUC of $0.8879$, surpassing the performance of the Transformer model (RMSE $0.1365$, AUC $0.8042$) and the multilingual BERT-based fusion model (AUC $0.8395$). Meanwhile, the early warning gain is improved to $5.2$ days, compared to approximately $1.2$–$2.3$ days achieved by conventional methods. In terms of classification performance, higher accuracy, precision, and recall are consistently observed, indicating more reliable discrimination between risk and non-risk states. Furthermore, cross-lingual generalization experiments verify the robustness of the proposed framework. When trained solely on the English market, the model achieves AUC values of $0.8624$ and $0.8471$ on the Chinese and European markets, respectively, with an RMSE reduced to $0.1185$, outperforming all competing methods. These results indicate that the proposed framework effectively handles cross-lingual and cross-market distribution discrepancies. Overall, the proposed method achieves substantial improvements in prediction accuracy, risk identification capability, and early warning performance, providing a solid theoretical foundation and technical support for the development of multilingual intelligent risk sensing and early warning systems in global financial markets.