The pre-training of large language models (LLMs) relies on large, carefully curated corpora. However, common preprocessing pipelines typically apply document-level heuristic quality filters in isolation, neglecting corpus-level distributional effects and often resulting in topical imbalance and redundancy. To address this, we propose DaSE (Data Selection by corpus Entropy), a maximum entropy–based algorithm that explicitly optimizes corpus diversity by retaining documents that increase entropy and discarding those that reduce it, producing high-quality, balanced pre-training data. To scale to industrial corpora, DaSE employs first-order Taylor–based document pruning with distributed computation, efficiently handling hundreds of billions of tokens. Experiments on a 2B-parameter language model show consistent gains across OpenCompass benchmarks, with an average 2.64% improvement on Chinese corpora, confirming DaSE’s effectiveness for large-scale LLM pre-training.

The modern financial system has evolved into highly interconnected networks, where local risks can escalate into widespread risks through bidirectional interactions between topological contagion and temporal memory effects. Existing methods fail to model such bidirectional interactions due to their unidirectional graph learning -> time-series learning framework. Besides, new financial innovations continuously introduce concealed risks. To address these challenges, we propose GraphShield, a novel risk recognition and quantification framework with three core innovations: (a) A multi-layer fusion unit hierarchically models bidirectional interactions between network topology and time-varying risk patterns. (b) A risk detector employs deep spatiotemporal clustering to autonomously identify hidden risks such as gray rhino threats by leveraging their aggregation patterns across spatial and temporal domains. (c) A risk quantifier reveals risk pathways and critical nodes by analyzing impact coefficients, enabling targeted mitigation of key propagation nodes. Extensive experiments on two real-world and two open-source datasets demonstrate the superiority of our GraphShield in risk containment. It achieves average AUC improvements of 9.36%, 9.07%, and 9.38% at risk ratios of 1%, 5%, and 10%, respectively, outperforming stateof-the-art baseline models. Furthermore, in risk regulation simulations, removing the top 1% of critical nodes reduced risk by over 30% and 25%, respectively, on two real-world datasets, demonstrating the effectiveness of our method in identifying critical risk nodes and its potential to enhance financial stability.

Multivariate time series forecasting (MTSF) is crucial for decision-making in various domains but faces challenges due to the low signal-to-noise ratio (SNR) in real-world data. While frequency-domain methods have been employed to address this challenge, they often discard high-frequency components, assuming they are predominantly noise, thereby overlooking valuable short-term and event-driven information. To address this limitation, we propose a novel disentangled representation learning framework that separates high-frequency components into informative signals and noise using mutual information maximization and minimization strategies. We introduce the Frequency Distangle Network (FDNet), which integrates disentanglement with low- and high-frequency decomposition, gated neural networks, and variable relationship fusion to effectively preserve and utilize high-frequency signals. Extensive experiments on 12 real-world MTSF datasets demonstrate that FDNet significantly outperforms leading frequency-domain and time-domain baselines, highlighting the importance of leveraging rather than eliminating high-frequency information. The source code is publicly available at: https://github.com/aohu1105/FDNet.

The evolution of financial technology has been marked by successive waves of innovation, from rule-based systems to the widespread adoption of machine learning methods, including deep learning approaches, which have delivered remarkable performance in a range of financial tasks. Recently, the emergence of generative AI (GenAI) has marked a new era in financial evolution, fundamentally reshaping the industry by enabling the creation of data, content, and solutions. Unlike conventional discriminative AI models that primarily focus on prediction and classification, generative models—such as Variational Autoencoders, Generative Adversarial Networks, Normalizing Flow, Diffusion Models, and Large Language Models—offer broader capabilities, enhanced user interaction, and improved interpretability. These advances are driving a new wave of innovative applications across diverse financial domains. Despite the growing literature on GenAI, there remains a lack of comprehensive surveys dedicated to its applications and sectoral impact in finance. This paper systematically reviews the deployment of GenAI techniques across key financial sectors. In particular, we highlight representative applications in securities, investment, banking, accounting, and regulatory compliance, illustrating how GenAI is transforming financial services and operations. Furthermore, we discuss the challenges associated with integrating GenAI into financial applications. This work aspires to be a reference for researchers and practitioners navigating the new era of generative AI in finance.

Stock price prediction is a challenging problem in quantitative investment, as financial markets generate complex, noisy, and dynamic time series containing heterogeneous signals. Short-term fluctuations usually exhibit greater uncertainty and stronger local variation, whereas long-term trends convey relatively stable and persistent information shaped by market and macroeconomic conditions. However, most existing methods struggle to distinguish these two components effectively, often leading to interference between short-term fluctuations and longer-term trends. In addition, they fail to capture dynamic temporal dependencies and cross-stock information propagation while preserving the causal structure of financial time series. To tackle these issues, we propose the Wavelet-Decoupled Spatiotemporal Network (WaveDSTN). It leverages wavelet transformation to decompose stock returns into high-frequency and low-frequency information, corresponding to short-term fluctuations and long-term trends, respectively. This decomposition enables the model to learn complementary predictive patterns more effectively. Furthermore, WaveDSTN incorporates a Dual-Path Spatiotemporal Encoder to capture complex temporal dependencies and evolving cross-stock information propagation while preserving temporal order and causal consistency. Extensive experiments demonstrate that WaveDSTN achieves significant improvements over existing methods, showing that explicitly modeling trend and fluctuation components can enhance predictive accuracy and reduce uncertainty in stock return forecasting.

Accurate modeling of correlations among variables (also referred to as channels) is essential for achieving precise and reliable multivariate time series forecasting. However, challenges persist in multivariate time series analysis due to anomalies such as outliers, missing values, and noise, which obscure true relationships between variables and reduce forecasting accuracy. To address this, we propose a Multivariate Relational Decomposition Network (MRDNet) that separates variable relationships into global and local components. The global component captures shared seasonal patterns among variables using learnable latent vectors, establishing stable long-term correlations. The local component focuses on window-specific relationships by analyzing trends in smoothed time series data. Additionally, we design an adaptive neural network to refine these relationships and mitigate the effects of anomalies. By integrating these components, MRDNet effectively handles anomalies while capturing both stable global dependencies and dynamic local patterns. Extensive experiments on benchmark datasets demonstrate that MRDNet achieves superior forecasting performance, addressing challenges posed by multivariate time series anomalies.

Traditional financial studies focused on the fundamental indicators data and financial news to track stock movements. Recent studies have revealed that firm relationships, measuring spillovers from related firms, are also important for stock forecasting. However, previous studies via machine learning mostly relied on incomplete market information and failed to capture the interactions and heterogeneity within multimodal data. In this study, a multi-graph neural network framework (MGNN) is proposed to fuse heterogeneous market information with different data structures, i.e., numerical, textual and graphical representations. With such a design, the unique and heterogeneous nature of multimodal data is well addressed. Specially, a fine-grained event extraction methodology is employed to identify valuable events within a substantial corpus of financial news, thereby ensuring the reliability and quality of the extracted events. Then, an explicit and an implicit graph are constructed to integrate firm relationships with fundamental and news information, respectively. Finally, the distinct spillover effects of fundamental and news information are further integrated using a multi-scale cross-attention dual-branch transformer to capture the joint impact on stock movements. Experiments performed on real market data demonstrate the effectiveness of the proposed approach over state-of-the-art algorithms.

Stock markets are complex dynamical systems shaped by macro fundamentals, policy, and investor behavior. Their evolution is non-stationary with recurrent market states and abrupt shocks, which challenges models that assume stable distributions. We tackle this problem with modern Koopman theory, which portrays complex dynamical systems and considers such nonstationary dynamics. Specifically, we propose REAKA, a Residual-Enhanced Adaptive Koopman Autoencoder for modeling stock-return dynamics in the latent space, with a residual path to capture higher-order nonlinearities. REAKA introduces an Adaptive Koopman Selector to adaptively choose the appropriate operator for different market conditions and a diffusion-based residual corrector to handle noise, abrupt shocks, and nonlinear effects beyond the limits of finite-dimensional Koopman linearization. By embedding these modules within an autoencoder, REAKA learns Koopman-invariant coordinates and advances dynamics in latent space. Experiments on real stock market data demonstrate that REAKA outperforms existing methods, significantly improving prediction accuracy and robustness in complex financial environments.

The detection of abnormal fluctuations can provide listed companies with an early warning of potential risks, which is critical to ensuring the stability and sustainable operation of companies. Previous studies, however, tended to consider only the inherent risks of a firm, ignoring the contagion risk and macro state to which the firm is exposed. This study is the first work to incorporate spillover effects and global information in the supply chain into enterprise risk detection. Specifically, we first propose a direction-sensitive messaging method to simulate the contagion effects in supply chain networks. This is achieved by adopting a heterogeneous mapping mechanism to distinguish spillovers generated and suffered by firms in different business flows. Then, we design a self-supervised representation learning module to dynamically adjust the representation of firm nodes based on the situation of both the firm and the supply chain it belongs. Finally, an end-to-end framework is employed to simultaneously capture inherent risk, contagion risk, and global risk. This study paves the way for addressing supply chain information in enterprise risk detection. Experiments on the real-world dataset demonstrate the advantages of the proposed framework over four classic and three state-of-the-art algorithms.

The rapid advancement of generative AI (GenAI) has introduced transformative methodologies to the financial sector, enabling the creation of novel data and solutions to longstanding challenges such as data scarcity, privacy, and domain adaptation. Despite the proliferation of GenAI research in finance, there is a lack of comprehensive surveys that systematically review the core generative techniques and the unique research problems posed by different financial data modalities. This paper addresses this gap by providing an in-depth overview of foundational GenAI methodologies, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Normalizing Flow, Diffusion Models, and Large Language Models (LLMs), as well as their adaptations for financial tasks. We categorize and analyze research challenges according to data modalities: textual, time series, tabular, and graph data. For each modality, we introduce representative tasks such as classification, forecasting, question answering, and synthetic data generation, and further discuss current limitations and future research directions. This survey aims to serve as a technical reference for researchers and practitioners seeking to understand and advance GenAI techniques in the financial domain.

Graph learning is a prevalent domain that endeavors to learn the intricate relationships among nodes and the topological structure of graphs. Over the years, graph learning has transcended from graph theory to graph data mining. With the advent of representation learning, it has attained remarkable performance in diverse scenarios. Owing to its extensive application prospects, graph learning attracts copious attention. While some researchers have accomplished impressive surveys on graph learning, they failed to connect related objectives, methods, and applications in a more coherent way. As a result, they did not encompass current ample scenarios and challenging problems due to the rapid expansion of graph learning. Particularly, large language models have recently had a disruptive effect on human life, but they also show relative weakness in structured scenarios. The question of how to make these models more powerful with graph learning remains open. Our survey focuses on the most recent advancements in integrating graph learning with pre-trained language models, specifically emphasizing their application within the domain of large language models. Different from previous surveys on graph learning, we provide a holistic review that analyzes current works from the perspective of graph structure, and discusses the latest applications, trends, and challenges in graph learning. Specifically, we commence by proposing a taxonomy and then summarize the methods employed in graph learning. We then provide a detailed elucidation of mainstream applications. Finally, we propose future directions.

Multimodal sarcasm detection (MSD) must recognize ironic intent that often arises from semantic mismatch between image and text. Multimodal large language models (MLLMs) can verbalize such mismatches as conflict explanations, offering useful high-level cues, but their outputs are prone to noise and hallucination. To reconcile these strengths and weaknesses, we propose KLGate. It leverages LLM-generated conflict explanations and employs a Kullback-Leibler (KL) divergence–based gate to align them with the original modalities, preserving reliable signals while suppressing noise. The gated features are fused with visual and textual representations to generate the final prediction. KLGate sets a new state of the art on standard benchmarks, achieving a binary F1 of 90.35 with 1.42 points gain over the best baseline.

Accurately forecasting the impact of salient financial events on markets is critical for investors and policymakers. However, existing multimodal time-series models typically fuse text and prices symmetrically, without an explicit way to decide when event text is truly predictive, and thus struggle to exploit the directional event-to-price structure and the heterogeneous roles of textual and price signals. In this work, we propose GS-Fuse, a multimodal event-based forecasting framework that employs (i) a Granger-supervised, causal-aware gated fusion module, which learns to open toward event text only when it provides incremental predictive value beyond historical prices, and (ii) a multi-granularity alignment mechanism that jointly aligns high-level event representations and fine-grained textual cues with future market trajectories. Built as a flexible, plug-and-play adapter on top of off-the-shelf large language models and time-series foundation models, GS-Fuse can be instantiated across diverse backbones and market settings. Extensive experiments on real-world financial datasets show that GS-Fuse consistently outperforms state-of-the-art time-series and multimodal baselines across multiple assets and forecasting horizons.

The exponential growth of textual content on the internet, alongside vast archives of news articles, scientific papers, legal documents, and other domains, has made Automatic Text Summarization (ATS) increasingly important. ATS aims to create concise and accurate summaries, significantly reducing the effort required to process large volumes of text. Originating in the 1950s, ATS has evolved through several technical shifts, moving from statistical models to machine learning and deep learning approaches, and more recently to pre-trained models. Previous surveys have focused on conventional ATS methods, which are often constrained by the predefined generative paradigms. However, the advent of Large Language Models (LLMs) has introduced a paradigm-flexible approach to summarization. With their superior generative capabilities, in-context learning, and few-shot learning abilities, LLMs have demonstrated remarkable improvements in coherence, fluency, and overall summarization quality. In this survey, we provide a comprehensive review of both conventional ATS approaches and the latest advancements in LLM-based methods. Our contributions include: (1) offering an up-to-date survey of ATS; (2) reviewing the latest LLM-based summarization methods.

Stock crash risk often propagates through various interconnected relationships between firms, amplifying its impact across financial markets. Few studies predicted the crash risk of one firm in terms of its relevant firms. A common strategy is to adopt graph neural networks (GNNs) with some predefined firm relations. However, many relations remain undetected or evolve over time. Restricting to several predefined relations inevitably makes noise and thus misleads stock crash predictions. In addition, these relationships are not independent during the process of propagating information and interacting with each other. This study proposes the multi-relation reconstruction and fusion graph neural network (MRRFGNN) to predict stock crash risk by capturing complex relations among listed companies. First, the model employs self-supervised learning and contrastive learning to reconstruct and infer implicit relationships between companies. Second, the model incorporates a relation self-attention mechanism to integrate various types of relationships, enabling a more nuanced understanding of the multiple spillover effects. Empirical evidence from a series of experiments demonstrates the superiority of the proposed method, which achieves the best performance with improvements of at least 2.14% in area under the curve (AUC) and 2.64% in Matthews correlation coefficient (MCC), highlighting its potential for practical application in financial markets.

The influence of the digital economy on organizational resilience remains understudied, especially in emerging markets. Thus, this study examines the impact of the digital economy on organizational resilience in China, with specific focus on the moderating role of customer stability. Using panel data from 27,342 Chinese A-share listed companies (2012-2022), we find that the digital economy significantly enhances organizational resilience, and this effect is enhanced by customer stability. Moreover, nonstate-owned enterprises in China's more digitally developed eastern regions tend to experience greater resilience. These findings extend the resource-based view and dynamic capabilities perspective by highlighting the synergistic contribution of external digital resources and stable customer relationships to the concept of organizational resilience. Overall, our research provides valuable insights for managers and policymakers aiming to navigate digital transformation and build resilient organizations.

Financial risks can propagate across both tightly coupled temporal and spatial dimensions, posing significant threats to financial stability. Moreover, risks embedded in unlabeled data are often difficult to detect. To address these challenges, we introduce GraphShield, a novel approach with three key innovations: Enhanced Cross-Domain Infor mation Learning: We propose a dynamic graph learning module to improve information learning across temporal and spatial domains. Advanced Risk Recognition: By leveraging the clustering characteristics of risks, we construct a risk recognizing module to enhance the identification of hidden threats. Risk Propagation Visualization: We provide a visualization tool for quantifying and validating nodes that trigger widespread cascading risks. Extensive experiments on two real-world and two open-source datasets demonstrate the robust performance of our framework. Our approach represents a significant advancement in leveraging artificial intelligence to enhance financial stability, offering a powerful solution to mitigate the spread of risks within financial networks.

Large Language Models (LLMs) have been well-researched in many long-context tasks. However, due to high annotation costs, high-quality long-context summary datasets for training or evaluation are scarce, limiting further research. In this work, we introduce CNNSum, a new multi-scale Chinese long-context novel summarization benchmark, including four subsets, length covering 16k to 128k, 695 samples in total, the annotations are human-driven. We evaluate commercial and open-source models on CNNSum and conduct a detailed analysis. Based on the observations, we further conduct fine-tuning exploration with short-context summary data. In our study: (1) GPT-4o underperformed, due to excessive subjective commentary. (2) Currently, long-context summarization mainly relies on memory ability, small LLMs with stable longer context lengths are the most cost-effective. Using long data concatenated from short-context summaries makes a significant improvement. (3) Prompt templates may cause a large performance gap but can be mitigated through fine-tuning. (4) Fine-tuned Chat or Instruction versions may harm the Base model and further fine-tuning cannot bridge performance gap. (5) while models with RoPE base scaling exhibit strong extrapolation potential, their performance may vary significantly when combined with other interpolation methods and need careful selection. (6) CNNSum provides more reliable and insightful evaluation results than other benchmarks. We release CNNSum to advance research in this field (https://github.com/CxsGhost/CNNSum).

This paper aims to investigate the impact of non-interest income diversification on the performance of commercial banks in China. We conduct an empirical analysis using the panel data of 159 Chinese commercial banks from 2006 to 2020. The results indicate that there is a positive relationship between the degree of non-interest income diversification and bank performance. We suggest the commercial banks to actively develop diversified non-interest business and to balance the proportion of interest and non-interest income. Furthermore, we conduct heterogeneity analysis and find remarkable differences existing among different types of commercial banks. The results show that the effect of non-interest income diversification on bank performance is more significant for the small-sized or local commercial banks whose target customers are local residents. We also take a further analysis with interaction terms and find that the positive effect of non-interest income diversification on bank performance is stronger for those with higher market competitiveness. Our findings have implications for bank managers about expanding banks' income resources through diversified business strategies and can provide experience for small or local banks seeking to advance the performance.

Stock movement prediction is a difficult task in the field of financial technology due to non-stationary dynamics and complex market interdependencies. Most of the existing research is based on deep neural networks, which lack interpretability. An interpretable prediction method helps uncover the mystery of the underlying operating mechanism of the securities market. In this work, we propose a model-guided method with interpretable homogeneous-heterogeneous processing for stock movement prediction. Specifically, based on that the correlations among the entities in the market are homogeneous within a short period, we unroll the iterative algorithm for solving the tensor robust principal component analysis (TRPCA) to separate the homogeneous and heterogeneous patterns from multiview data. Then, a specialized tensor-based attention for homogeneous and heterogeneous feature extraction is designed, and embedded in long short-term memory (LSTM) for better prediction. Experiments on real datasets show our model's superiority over state-of-the-art stock forecast methods.