As the real-world environment becomes increasingly complex, traditional neural network models that only consider pairwise (one-to-one) interaction and temporal dynamics exhibit some limitations. This paper investigates the spatiotemporal Hopf bifurcation dynamics in diffusion bidirectional associative memory (BAM) neural networks incorporating higher-order interactions. First, the classical BAM neural network model is extended by introducing diffusion-to describe the influence of non-uniform electromagnetic fields on voltage distribution-and higher-order interactions, which capture the cooperative activation of multiple neurons. Second, by choosing time delay as the bifurcation parameter, analytical conditions for local stability and the occurrence of Hopf bifurcation are established for an arbitrary number of neurons. Finally, numerical simulations are conducted to verify the theoretical results, demonstrating the rich spatiotemporal dynamic behaviors of the model. The results indicate that time delay, self-feedback strength, interaction strength, and network size significantly influence the bifurcation dynamics of the network. Compared to binary interaction, the incorporation of higher-order interactions effectively decreases the bifurcation threshold and increase the limit cycle amplitude, thereby reduce network stability.

The coexistence of multiple online travel agencies (OTAs) in the hotel booking market has led to information fragmentation and ambiguous sentiment expressions in user reviews, making objective hotel evaluation challenging. To address this problem, a cross-platform hotel evaluation method integrating intuitionistic fuzzy sets with sentiment analysis was proposed. User reviews were collected from three major OTAs, namely Ctrip, Tongcheng, and Qunar. The Latent Dirichlet Allocation (LDA) model was employed to identify core evaluation attributes from review texts. A hybrid sentiment analysis framework combining bidirectional long short-term memory (Bi-LSTM) and attention mechanism was constructed to accurately capture the sentiment tendency of each attribute, and the sentiment information was then transformed into intuitionistic fuzzy numbers to characterize users' positive, negative, and uncertain emotional perceptions. A dynamic comprehensive weighting model was established by integrating subjective LDA-based topic weights with objective intuitionistic fuzzy entropy weights. An improved score function was adopted to aggregate the intuitionistic fuzzy information and obtain the comprehensive evaluation scores and rankings of hotels. A case study based on cross-platform hotel review data demonstrated that the proposed method can effectively capture ambiguous emotional expressions in online reviews and objectively reflect the actual service quality of hotels. The method provides reliable decision support for consumers in hotel selection and practical guidance for hotel managers to improve service quality.

The human-robot collaborative order picking system represents a new paradigm aligned with the "human-centered" philosophy, wherein efficient collaboration between pickers and autonomous mobile robots (AMRs) in both time and space is critical. This paper addresses the scheduling problem of such systems under synchronization constraints, and a mixed 0-1 integer programming model is formulated to minimize the total picking cost incurred by both pickers and AMRs, while a Q-learning-based hyper-heuristic algorithm is also proposed to address the complex problem. The algorithm employs a greedy approach to obtain initial solutions, incorporates nine knowledge-driven low-level heuristics tailored to the problem's characteristics, and adopts the hybrid $\epsilon$-greedy and softmax strategy for adaptive heuristic selection. Additionally, the Metropolis criterion is employed to accept inferior solutions with a certain probability. Numerical experiment results demonstrate that the proposed algorithm outperforms exact method, genetic algorithm, variable neighborhood search algorithm, and random-search-based hyper-heuristic algorithm. The ablation experiments validate the effectiveness of multiple low-level heuristics. Sensitivity analysis results reveal the effect of economies of scale in AMR quantity. Moreover, the system achieves the optimal performance when the human-to-robot speed ratio is set to 1.5:1.

With the evolution of the “dual carbon” goals and the climate policy system, climate policy uncertainty has become a crucial factor affecting the credit allocation of commercial banks. Using a sample of 42 A-share listed commercial banks in China from 2012 to 2024, this paper investigates the impact of climate policy uncertainty on green credit in commercial banks. The study shows that climate policy uncertainty has a significant positive impact on green credit in commercial banks, and this effect is more pronounced in banks with smaller scale, higher management attention to climate risks, as well as in banks with higher levels of green innovation and higher government attention to climate risks. At the same time, climate policy uncertainty enhances green credit in commercial banks by strengthening environmental regulations and improving banks' ESG performance. Additionally, enhancing the degree of digital transformation of banks and the development level of regional green finance can help strengthen the promoting effect of climate policy uncertainty on green credit of commercial banks. This study provides an important reference for policymakers to construct a reasonable climate policy system that guides commercial banks in optimizing green credit allocation.

Operational and decision-making optimization in hospital management can be effectively formulated as a nonlinear programming problem characterized by a fractional structure. This paper reconstructs the classic Mathis model, explicitly providing the variable substitutions, feasible region reconstruction, and parameter mapping relationships from the original hospital management model to the sum of linear ratios problem, thereby transforming the hospital management optimization problem into an equivalent sum of linear ratios problem. To address the inherent non-convexity of the problem, a linear relaxation programming framework is constructed. By embedding image-space partitioning and domain reduction technique, a novel branch-and-bound algorithm is developed to solve the hospital management optimization problem globally. Theoretical analysis demonstrates that the proposed algorithm guarantees global convergence within a given error tolerance. Finally, numerical experiments are conducted to validate the computational efficiency and effectiveness of the algorithm.

The construction of uniform designs, a popular class of experimental designs used in both physical and computer experiments, is often complex and challenging. This study systematically investigates the construction of uniform designs under the newly proposed absolute discrepancy. We introduce a stochastic optimization algorithm for generating designs with flexible sizes and provide a comprehensive exposition of the level permutation method. We display analytical expressions for the average absolute discrepancy and establish its theoretical connections with other design screening criteria. Comparisons between the two construction methods reveal that uniform designs constructed by the level permutation method exhibit better properties, including lower absolute discrepancy values and less aberration. Simulation studies and a practical application demonstrate the superior performance of uniform designs under absolute discrepancy.

Community-embedded elderly care (CEEC) is an emerging service model that brings professional care resources into communities and supports older adults in ageing in place. The quality of CEEC services directly determines whether this model can be translated into actual well-being for older adults in everyday community settings. Drawing on the classical structure–process–outcome (SPO) framework, this study considers CEEC service quality as a comprehensive state co-generated through continuous interactions between the supply and demand sides. Moving beyond the one-way logic of traditional SPO models, in which demand-side outcomes are mainly explained by supply-side structure and process, this study extends the SPO framework into a dual structure composed of a supply support chain and a demand response chain, and develops a supply–demand dual-SPO deep coupling framework for CEEC service quality evaluation. Furthermore, to address the limitation of traditional coupling methods that rely on the coordination of aggregate system scores and fail to reveal complex interactions within CEEC service quality, this study proposes a deep coupling method integrating hesitant fuzzy Total Correlation, differentiated Copula, and linguistic-quantifier-guided POWA. The proposed method captures cognitive differences in multi-stakeholder evaluation and the overall dependence among supply–demand multi-indicator sets, identifies high-level advantageous synergy and low-level disadvantageous clustering, and reflects the priority constraint of "ensuring basic services before promoting improvement." An empirical analysis of six embedded elderly care communities shows that the proposed method can identify supply–demand coupling structures, sources of advantageous synergy, paths of shortcoming aggregation, and priority directions for improvement across different cases. The findings provide quantitative support for CEEC service quality diagnosis and classified improvement.

The artificial intelligence (AI) industry is characterized by a long cycle, high uncertainty, and massive capital requirements. Against this backdrop, patient capital——which focuses on long-term value——has emerged as a critical force to alleviate the financing predicament of AI startups. This paper constructs a bilateral investment and financing game model covering AI enterprises and venture capital institutions, and endogenizes the subscription demand in the secondary market. The study compares the incomplete information signaling game without patient capital intervention and the complete information sequential game with patient capital intervention, and further explores dynamic contract design by introducing the moral hazard of hidden action. The findings are as follows: (1) Under the traditional capital regime, the equilibrium outcome is closely related to the marginal costs of different types of AI projects; (2) Through ex-ante information screening and ex-post value co-creation, patient capital effectively transforms incomplete information into complete information, and realizes a multi-party win-win outcome under the optimal empowerment intensity; (3) The realization of value for Patient Capital depends heavily on the initial market conditions; (4) To address the moral hazard associated with firms' ex post performance, a multi-round investment model can create credible intertemporal penalties and incentives, thereby compelling firms to fulfill their obligations in full. Finally, this paper conducts a numerical analysis to intuitively verify the governance effectiveness of patient capital under different market prior beliefs and contract performance costs. This study provides a solid theoretical basis for optimizing the supply structure of sci-tech financial capital and formulating long-term oriented incentive-compatible policies.

This paper proposes a generalized linearized gradient descent symmetric alternating direction method of multipliers for a class of nonconvex and nonsmooth optimization problems with non-separable structures. The proposed method integrates the PR splitting (Peaceman-Rachford Splitting) technique, gradient descent, linearization, and the generalized ADMM framework. Firstly, under the Kurdyka-Łojasiewicz property and appropriate parameter assumptions, the global convergence and strong convergence of the generated sequence are proved. Finally, the algorithm is applied to the $l_1$-logistic regression problem and the SCAD-$l_2$ (Smoothly Clipped Absolute Deviation-$l_2$) problem, and comparative experimental results verify its effectiveness.

In this paper, we use Bayesian regularization method to study the estimation and selection of fixed effects and random effects in linear mixed effects quantile regression models. Firstly, by utilizing the modified Cholesky decomposition of the covariance matrix of the random effects distribution, the functions of covariance parameters are introduced into the model as regression coefficients to obtain the reparameterized form of the model. Secondly, under the assumption that the error terms follow the asymmetric Laplace distribution, $L_\xi(0<\xi<1)$ bridge penalties are added to the fixed effects and the diagonal elements of the decomposed random effect covariance matrix in the reparameterized model. By adding prior information to the bridge penalty parameters, a Bayesian quantile regression method based on bridge-randomized penalties is proposed. We further combine the partially collapsed Gibbs sampler and the adaptive random walk Metropolis algorithm to develop an adaptive Markov chain Monte Carlo (MCMC) algorithm for the proposed model, in order to improve the sampling efficiency of traditional MCMC algorithms. We demonstrate the performance of the proposed method through extensive simulation studies. The simulation results show that when fixed effects and random effects are sparse, the method based on the bridge-randomized penalty performs better in parameter estimation and variable selection than the method based on the lasso penalty. Finally, the proposed method is applied to a real longitudinal dataset in the medical field. Under the sparse linear mixed effects quantile regression framework, the proposed bridge-randomized penalization method can achieve accurate and robust parameter estimation and variable selection, providing an effective regularization approach for longitudinal data modeling.

Interacting Multiple Model (IMM) filtering is a classic Bayesian target tracking algorithm that exhibits good performance when dealing with maneuvering targets that frequently switch motion modes. However, its performance heavily depends on the reasonable configuration of the internal Transition Probability Matrix (TPM). Traditional methods typically rely on prior experience or offline training to obtain the TPM, which not only increases labor costs but also easily leads to degraded filtering performance in complex dynamic environments. To address this issue, this paper proposes an IMM filtering algorithm based on Online Adaptive Gradient Descent (OAGD), termed IMM_OAGD. This method dynamically estimates the TPM during the IMM filtering recursion process, thereby achieving online adaptive updates of the model transition probabilities. Firstly, the overall framework of IMM_OAGD is constructed. Secondly, a weighted loss function that relies solely on the observation sequence is designed, enabling parameter updates without requiring the true system state or pre-collected training data. Furthermore, a convergence analysis of the algorithm is conducted, and it is proved that the algorithm possesses a sublinear cumulative regret bound of $O(\sqrt{K})$, which guarantees that the average regret tends to zero. Finally, the effectiveness and robustness of IMM_OAGD in complex maneuvering target tracking are tested by extensive simulation experiments.

For the problem of nonparametric interval estimation for an unknown population variance, this paper constructs a nonparametric unbiased estimator of the population variance based on ranked set sampling method and the idea of U-statistics. According to the asymptotic normality of the new estimator, the confidence interval of variance is constructed, and it is shown that the precision of the confidence interval under ranked set sampling is higher than that under simple random sampling, regardless of whether the ranking is perfect or not. To solve the problem that the confidence interval expression contains unknown quantities, the method of moments is adopted to construct a consistent estimator of the asymptotic variance of the estimator, and then an approximate confidence interval of the variance is further constructed. Simulation calculations are conducted to compare the coverage rates and average lengths of approximate confidence intervals, bootstrap confidence intervals and jackknife confidence intervals, and an empirical data analysis is carried out. The research results show that the precision of the approximate confidence interval under ranked set sampling is higher than that under simple random sampling in both large and small sample cases. In addition, the overall coverage rates of bootstrap and jackknife confidence intervals are higher than those of approximate confidence intervals.

This paper introduces a panel data structure containing firm-specific micro characteristics into the vector autoregressive model with exogenous variables (VARX) and thereby constructs a high-dimensional panel vector autoregressive (HPVAR) volatility forecasting model to characterize the dynamic features of realized volatility. First, to address the difficulty of lag-order selection in high-dimensional volatility series, an improved hierarchical own-other (IHOO) regularization method is proposed to achieve structured regularization and lag-order selection simultaneously in a high-dimensional setting. Next, a proximal gradient descent (PGD) algorithm with adaptive step sizes, together with Bayesian optimization, is employed for parameter estimation and hyperparameter selection. An upper bound for the in-sample mean squared forecasting error (MSFE) is further derived under regularity conditions. Finally, empirical analysis is conducted using high-frequency volatility data from the U.S. and Hong Kong stock markets. The results show that: (1) after firm-specific characteristics and external shocks are incorporated, HPVAR achieves higher forecasting accuracy than VAR and VARX; (2) IHOO produces better forecasting performance than hierarchical lag own-other (HLagO) and Lasso; and (3) heatmaps of autoregressive lag coefficient matrices and estimated maximum-lag matrices indicate that the proposed model maintains both high forecasting performance and good interpretability under high dimensionality.

To address the separation between prediction and optimization in traditional two-stage “predict-then-optimize” approaches, A multilayer perceptron (MLP) prediction-integrated convex quadratic stochastic optimization model is proposed. The resulting model is formulated as a stochastic bilevel programming problem, in which the upper-level problem aims to minimize the final decision loss, while the lower-level problem is a convex quadratic program. A gradient descent algorithm is employed to solve the model, and its convergence, convergence to a stationary point, and sublinear convergence are established theoretically. Numerical experiments are conducted in the context of portfolio decision-making. The results show that, under different parameter combinations, the number of iterations, the average loss, and the final gradient norm of the algorithm remain within reasonable ranges, indicating good practical applicability and numerical stability. Out-of-sample evaluation results show that the proposed model generally outperforms the traditional two-stage method and other benchmark models in terms of cumulative return, risk-adjusted return, and maximum drawdown, thereby verifying the effectiveness of the integrated prediction and optimization framework.

With the rapid development of the platform economy and blockchain technology, enterprises are better positioned to pursue the "dual-carbo" objectives within the carbon quota trading mechanism. This study constructs Stackelberg game models for a platform supply chain operating under reselling and agency selling modes, considering both low-carbon information concealment and disclosure scenarios. The models analyze how consumers' low-carbon awareness and preference, carbon reduction costs, blockchain implementation costs, and market erosion factors jointly affect the strategic decisions and profits of manufacturers and platforms. The results show that: (1) the manufacturer's emission reduction effort and the profits of supply-chain members increase with consumers' low-carbon perception and preference but decrease with higher carbon-reduction costs, blockchain costs, market erosion effects, and service commission rates. (2) Under information concealment, the agency selling mode performs better only within a moderate range of carbon-trading prices and commission rates, whereas after information disclosure, the platform's profit becomes more sensitive to blockchain costs, and the favorable range of the agency selling mode narrows. (3) Blockchain cost acts as a critical enabling factor. When it remains moderate, blockchain adoption can achieve a win-win outcome in both economic and ecological performance.

In recommender systems, the lack of historical behavioral data for new users often makes it difficult to characterize their similarity to existing users. To address this issue, we propose a nonparametric regression–based similarity prediction method. Specifically, an incomplete Cholesky decomposition is first employed to capture the similarity among existing individuals. A penalized B-spline is then constructed to fit the underlying predictive model, enabling the prediction of similarities between new users and existing users. Under suitable regularity conditions, we establish the asymptotic distribution of the proposed prediction statistic. Numerical simulations and real-data analysis demonstrate the feasibility and effectiveness of the proposed method.

Multi-layer networks provide a framework for capturing system complexity by being able to model different interaction patterns or states of the same type of entities at different moments. However, how to extract the community structure in multi-layer networks is a fundamental problem. Most of the existing methods use matrix data for computation, ignoring the multi-layer data structure and higher-order information of the network, resulting in limited accuracy of community partitioning. To address the above problems, this paper proposes a community detection method for multi-layer networks based on higher-order information and tensor train decomposition. The method first constructs a weighted adjacency matrix containing higher-order structural information, and then takes advantage of the natural tensor representation capability of multi-layer networks, and adopts the tensor train decomposition technique to realize node embedding representation, so as to fully explore the structural features of different layers. The experimental results show that this method outperforms the current mainstream methods in community detection in both synthetic and real-world multi-layer networks, and has better detection accuracy, especially suitable for community detection in multi-layer networks with a large number of layers.

As important participants in capital markets, institutional investors may influence both market stability and corporate governance through institutional investor cliques. We examine the impact of institutional investor cliques on stock tail market manipulation. Using Chinese A-share listed firms from 2014 to 2023 as the research sample, we construct an institutional investor common ownership network based on institutional shareholding data and identify institutional investor cliques using the Louvain algorithm based on modularity optimization. Meanwhile, using high-frequency intraday trading data of listed firms, we construct a suspicious tail market manipulation indicator based on an end-of-day price dislocation model. The empirical results show that institutional investor cliques significantly suppress stock tail market manipulation. This conclusion remains robust after addressing potential endogeneity and conducting a series of robustness checks, such as instrumental variable estimation, Heckman two-stage models, alternative variable definitions, different network connection thresholds, sample exclusions, and alternative winsorization treatments. Mechanism analysis reveals that institutional investor cliques inhibit stock tail market manipulation by improving the information environment and strengthening external corporate governance. Heterogeneity analysis further indicates that the governance effect of institutional investor cliques is more pronounced when investor sentiment is pessimistic, when firms are in the growth stage, or when firms belong to high-tech industries. These findings provide new evidence for understanding the governance role of institutional investor networks in emerging capital markets and have important implications for regulatory policies aimed at preventing stock tail market manipulation.

Regulating the opportunistic share reduction behavior of executives in listed companies has become a crucial component in the construction of the capital market regulatory system. Using a sample of China’s Shanghai and Shenzhen A-share listed companies from 2007 to 2022, this paper empirically investigates the impact of economic policy uncertainty (EPU) on executives’ opportunistic share reductions and its underlying mechanisms. The findings reveal that EPU significantly inhibits both the volume and frequency of executives’ opportunistic share reductions. This conclusion remains robust after a series of endogeneity and robustness tests, including the Instrumental Variable (IV) approach and Propensity Score Matching (PSM). Mechanism analysis indicates that EPU suppresses opportunistic share reductions through two channels, including the “dilution of information advantages” and the “compression of profitability space”. Further research shows that this inhibitory effect is more pronounced in firms with higher policy dependence, lower audit quality, and lower analyst coverage. This study expands the research on the determinants of executives’ opportunistic share reductions from the perspective of the macroeconomic policy environment and enriches the literature on the economic consequences of EPU. Furthermore, it provides important empirical evidence for regulatory authorities to optimize regulations on share reductions and stabilize market expectations.

We develop two models for the cases where products are launched first through direct-sale channel and first through resale channel. The optimal product priority launch strategy and its impact on the pricing of the dual-channel supply chain and social welfare are investigated. We uncover that both the manufacturer and the retailer may have the motivation to prioritize the launch of products in their own channels, as it can help their channels have demand and gain more profits. Strong bargaining power can increase the motivation of manufacturer or retailer to prioritize the launch of products in its channels. Meanwhile, in different cases of product priority launch, both manufacturer and retailer need to adjust their pricing strategies based on the degree of channel differentiation and bargaining power. If there is demand occurring in dual-channel, then product priority launch through direct sale-channel can increase the demand for products in each channel and the total demand, bringing about the "social welfare effect". However, in either case, strong bargaining power can always motivate the manufacturer and the retailer to implement a high-price strategy, exacerbating the inequality in profit distribution within the supply chain. Finally, the dual-channel simultaneous product launch is expanded, indicating manufacturer may choose to launch product simultaneously through the direct-sale channel and the resale channel.

The scientific selection of performance targets in Valuation Adjustment Mechanisms (VAM) is a crucial foundation for balancing risks between investors and investees, ensuring the fairness and enforceability of agreements, and securing the stable growth of corporate long-term value. Based on the Newsvendor model, this paper takes the retailer in a two-echelon supply chain as the primary research subject and constructs a VAM model between the retailer and Private Equity (PE) institutions under demand uncertainty. It compares the differences in incentive intensity and effects on the retailer between two types of targets: sales-based VAM and net profit-based VAM. The research findings indicate that sales-based VAM can effectively incentivize retailers to increase market investment and achieve rapid sales growth; however, it is also prone to inducing over-expansion, leading to a growth pattern reliant on continuous capital injections. Such short-term growth, achieved at the expense of long-term robustness, is fundamentally unsustainable. In contrast, net profit-based VAM effectively curbs the retailer's tendency toward blind expansion by strengthening cost constraints. It improves capital utilization efficiency and contributes to achieving a larger scale of both net and total assets.

Generalized Case-Cohort (GCC) sampling is a classic design for controlling covariate measurement costs in large-scale cohort studies, and is widely applied in epidemiological studies with moderate to high incidence rates. Existing studies have largely focused on the large-sample asymptotic properties of variable selection methods under the GCC framework, while paying insufficient attention to estimation robustness in finite-sample settings with heavy censoring. Conventional inverse probability weighted estimators often suffer from degraded finite-sample performance and unstable variable selection results due to excessively large weight variance. This paper proposes a variance-shrinkage based robust weighted SCAD-penalized variable selection method. By constructing shrinkage-type robust sampling weights, the method substantially reduces weight variance with controllable bias, and mitigates estimation fluctuations induced by biased sampling in finite samples. Under a high-dimensional framework where the covariate dimension diverges with the sample size, we derive finite-sample uniform convergence bounds and explicit estimation error bounds of the estimator, establish the element-wise uniform boundedness and Frobenius norm bound of the Hessian matrix, and prove the oracle property of the proposed method. We also present the Berry–Esseen normal approximation bound and the explicit solution for the optimal shrinkage parameter, and quantify the impacts of sampling fraction, censoring rate and covariate dimension on estimation performance. Numerical simulation results demonstrate that the proposed method significantly outperforms the conventional weighted SCAD method in both estimation accuracy and variable selection stability under extreme scenarios with small sample sizes and heavy censoring. Finally, we apply the proposed method to data from the Busselton Health Study, providing practical guidance for GCC sampling design and variable selection across different research settings.

This paper proposes a comprehensive infant undercount estimator designed to capture the full spectrum of infant undercounts in the population census, thereby replacing the currently widely used single-source undercount estimator, which is known to significantly underestimate the true scale of infant undercount. Employing an integrated methodology that combines mathematical modeling, multiple sampling techniques, and field surveys, the study addresses the construction and related issues of the comprehensive infant undercount estimator. Both theoretical and empirical findings demonstrate that the single-source undercount estimator fails to provide complete coverage of infant undercounts in the total population. Furthermore, the comprehensive infant undercount estimator must be constructed on the basis of homogeneous population strata; otherwise, it is susceptible to heterogeneity bias. Compared to single source, double source, and triple source undercount estimators, the comprehensive infant undercount estimator exhibits superior estimation accuracy, making it suitable for estimating infant undercounts. As a biased estimator, its precision should be assessed using its mean squared error estimator. This research makes an original contribution by introducing the comprehensive infant undercount estimator and establishing a systematic framework for estimating infant undercounts in population censuses. The comprehensive infant undercount estimator is expected to provide theoretical support and practical guidance for the National Bureau of Statistics in designing future infant undercount estimation programs, thereby enhancing their scientific rigor and operational feasibility.

This paper studies coordinated routing of a truck and truck-mounted drones for emergency patrol of transmission lines under stringent timeliness requirements. Multiple drone flight modes and the capability of consecutively visiting multiple task nodes are considered. The problem is formulated as an Orienteering Problem with multiple Drones and Multi-Stop Sorties (OP-mD-MS), with the objective of maximizing patrol rewards. A Benders decomposition algorithm is developed by separating the original problem into a master problem for truck routing and drone launch-and-recovery decisions and a subproblem for drone routing and truck-drone time coordination. Feasibility and optimality cuts are used to exchange information between the two problems and guide the iterative search for an optimal solution. Computational experiments show that the proposed algorithm performs well in terms of solution quality and computational time and is effective and practical in a real-world case study.

White noise testing is an important fundamental procedure in statistics and has been widely used in practice. However, when the dimension of the time series is relatively large, the construction of white noise tests becomes challenging. Although the existing literature has extensively explored this issue, most studies rely on the assumption that the innovations are independent and identically distributed (i.i.d.). In particular, Li et al. proposed a method based on random matrix theory, which has been shown to possess desirable statistical properties. This paper finds, however, that when the dimension is relatively large and the innovations form a martingale difference sequence with zero conditional mean, for example, when the innovations follow a process with GARCH-type conditional heteroskedasticity, the test of Li et al. suffers from severe over-rejection, resulting in a pronounced oversized phenomenon. The primary reason is that the asymptotic distribution of the test statistic under martingale difference innovations differs from that derived under the i.i.d. assumption. Accordingly, this paper proposes a wild bootstrap correction and establishes its asymptotic validity. Simulation studies demonstrate that the refined method exhibits robust finite-sample performance across a variety of data-generating mechanisms, significantly alleviates the oversized problem of the original test, and remains robust when the dimensionality is relatively large. Finally, we apply the corrected procedure to the daily stock returns of the 50 companies in the S&P 500 to illustrate its practical effectiveness.

Accurate forecasting of commodity prices helps policymakers respond promptly to external economic shocks and enhances macroeconomic resilience. This paper investigates the interdependencies within commodity markets and the transmission of risks across different assets, with a focus on price dynamics. Based on Bayesian inference principles, a prior heterogeneous graph is constructed incorporating three types of semantic relationships (causal, similarity, and correlation) to capture multidimensional influence pathways among commodities. Serving as the foundation for spatial graph convolution layers, this prior graph enables the model to effectively learn complex spatial dependencies among commodities. Leveraging an attention mechanism, self-attention value matrix is derived as posterior dynamic dependency matrices, representing the evolving interdependencies among commodities under price fluctuations, thereby helping identify potential market linkages and risk transmission paths, such as spillover effects from upstream industrial goods to downstream products. Ultimately, we propose a heterogeneous graph convolutional neural network model based on multi-head self-attention mechanisms for price fluctuation prediction. Using 50 actively traded commodity futures contracts from China's major trading markets as case studies, we conduct comparative experiments and ablation analyses based on preprocessed real-world trading data. Experimental results show that our proposed algorithm outperforms existing methods across multiple evaluation metrics, reflecting its effectiveness and practical value.

Data in fields such as finance, industry, medicine, and meteorology commonly exhibit characteristics including heteroscedasticity, skewness, and heavy-tailed distributions. The identification and testing of change-points in such data holds significant practical relevance and theoretical value. This study focuses on the joint location and scale models. Utilizing a modified information criterion (MIC), we propose a method for identifying change-point locations under the assumption of a skew-t-normal(StN) distribution. To evaluate the performance of this method, we conduct Monte Carlo simulations involving simultaneous shifts in both the location parameter and scale parameter, comparing the proposed method against the classical likelihood ratio test (LRT). Simulation studies demonstrate the superior performance of the proposed method over the LRT. To further validate the method, we applied it to air quality data recorded in Lanzhou, China, from September 1, 2023, to August 31, 2024. The two change-points identified through this analysis coincide with documented environmentally significant change-points, demonstrating the high accuracy of the proposed change-point detection and identification methodology.

To address the challenges of strong model uncertainty, insufficient generalization under complex operating conditions, and high operational cost in the denitrification dosing control of wastewater treatment plants, this paper proposes a mechanism-guided residual learning and flexible constraint optimization method. In this approach, a mechanistic model is employed to characterize the underlying dynamics of the dosing process, while LightGBM is utilized to learn and compensate for the residuals of the mechanistic model. Lag terms and temporal features are further incorporated to enhance the model's representational capacity. Building upon this hybrid framework, a prediction-based flexible dosage-saving strategy is introduced to achieve economic optimization of the dosing process. Experimental results demonstrate that the proposed hybrid model substantially outperforms the mechanistic model in predictive accuracy, reducing the Mean Absolute Percentage Error (MAPE) from 27.26% to 5.79%. Compared with purely data-driven models, it also exhibits superior out-of-distribution stability and generalization capability. Under Gap Temporal Split scenarios, the mechanism-guided model markedly mitigates performance degradation, while achieving a 3.54% reduction in chemical consumption under the constraint that effluent quality standards are satisfied. SHapley Additive exPlanations (SHAP) analysis further reveals that the incorporation of mechanistic structure enhances both the interpretability and operating-condition adaptability of the model. Overall, the proposed method achieves a favorable balance among predictive accuracy, robustness, and operational economy, offering an effective approach for mechanism-data fusion modeling and optimal control of complex wastewater treatment processes.

Air cargo serves as a crucial link in supporting the efficient operation of global supply chains, with its demand fluctuations driven by multiple complex factors. Based on the TEI@I methodology, this study applies complementary ensemble empirical mode decomposition to the monthly cargo data of ten major Chinese airports from 2015 to 2024. According to the stationarity and complexity of each subsequence, modeling and forecasting are conducted using seasonal autoregressive integrated moving average models, long short-term memory networks, and support vector regression models, with an event-driven adjustment module incorporated. Furthermore, this paper analyzes the correlation characteristics of cargo volumes within the northern, eastern, and southern airport clusters and compares differences in factor sensitivity and prediction error structures across different clusters. Empirical results demonstrate that the TEI@I prediction model constructed in this study performs well for most airports and evaluation indicators, providing reliable methodological support for multi-airport cargo volume forecasting. It also offers important reference value for strategic planning and resource allocation by local governments, airport operators, and air logistics enterprises.

China is one of the countries most severely affected by natural disasters, and emergency evacuation is a common and effective measure for disaster prevention and mitigation. In practice, emergency evacuations often require the rapid transfer of some vulnerable populations (e.g., the elderly, pregnant women, and the injured). However, post-disaster road networks may be damaged and congested, severely constraining evacuation efficiency. Electric Vertical Takeoff and Landing (eVTOL) aircraft feature vertical takeoff and landing, obstacle-crossing, and high-speed capabilities, enabling rapid transfer of vulnerable populations and complementing ground transportation to overcome road evacuation bottlenecks. Considering capacity differences and transfer service delays during ground-to-air mode transitions, this paper proposes a ground-air coordinated dynamic evacuation optimization model. In the model, the objective function is to minimize total system evacuation time by jointly deciding the ground-air splitting ratio and path traffic flow. First, the Cell Transmission Model (CTM) is used to characterize the dynamic evolution of ground-air evacuation traffic flows, and a transfer cell mechanism incorporating passenger capacity conversion and transfer service delays is designed. Second, a decomposition algorithm based on the Alternating Direction Method of Multipliers (ADMM) is proposed to solve our proposed model. Finally, the effectiveness of the model and algorithm is validated on the Sioux Falls network. The results indicate that: (1) ground-air collaborative evacuation significantly outperforms only using ground evacuation. The coordinated benefits show a concave growth trend with increasing evacuation demand and tend to stabilize. There exists a globally optimal splitting ratio (approximately 0.38 in the numerical example of this paper) to achieve dynamic resource matching; (2) ground-air transfer nodes are the core bottleneck restricting the evacuation efficiency of air corridors, where transfer waiting time is significantly longer than flight time and increases at a faster rate; (3) improving the passenger capacity of eVTOL and the service capacity of take-off and landing points can reduce the total evacuation time, but both exhibit diminishing marginal returns, and the optimal diversion ratio is more sensitive to changes in passenger capacity; (4) the takeoff and landing service capacity and transfer speed of ground-air transfer nodes determine the evacuation efficiency of the air corridor. The improvement of air transport capacity should be matched with transfer service capacity to improve the system's evacuation efficiency. This paper can provide a decision-making support for evacuation plans, route planning, and the allocation of eVTOLs and vehicles during an emergency evacuation.

Aiming at the problem of insufficient engineering accuracy caused by additive noise interference in weak signal detection of Duffing system, this paper models the deterministic Duffing system excited by signal and noise as a stochastic differential equation driven by additive Gauss noise. This paper analyzes the characteristics that the drift term of the system contains a cubic nonlinear term and satisfies the superlinear growth, and points out that it does not satisfy the linear growth condition required by the classical solution existence theory. To this end, this paper applies Mao's classical framework and generalized local Lipschitz condition system to the system, constructs a Lyapunov function that satisfies the Khasminskii condition, and strictly proves that the system has a global solution under any initial conditions. On this basis, the Lyapunov function with undetermined parameters is further constructed, and the explicit upper bound of the second-order moment of the solution is derived by combining It\^{o} formula and Bihari inequality. Then, the quantitative conclusion that the upper limit of the second-order moment Lyapunov exponent is not positive is obtained directly from the upper bound of the second-order growth, which shows that the system will not appear exponential runaway growth.

This paper proposes a cooperative control algorithm for multi omnidirectional mobile robot systems over directed topologies based on the Udwadia-Kalaba (U-K) approach, aiming to achieve formation tracking and obstacle avoidance control. First, within a leader-follower framework, the formation tracking task is formulated as equality constraints and further reformulated into a second-order differential form to match the acceleration-level constraint paradigm of the U-K method. Then, under the condition that the initial states lie within the safe set, obstacle avoidance inequality constraints are transformed into second-order constraints that can be embedded into the U-K framework by constructing control barrier functions. On this basis, the formation tracking and obstacle avoidance constraints are integrated to establish a weighted composite constraint set. By assigning different weights, safety-critical obstacle avoidance is approximately prioritized under finite weights, while a lexicographic solution with obstacle avoidance priority is obtained as the weight tends to infinity. Finally, an explicit control law is derived in the sense of weighted least squares based on the U-K method. Simulation results demonstrate that, in environments containing both static and moving obstacles, the proposed algorithm can restore formation tracking while satisfying safety distance constraints. Compared with a Lyapunov-based model predictive control method, the proposed algorithm exhibits superior real-time performance and control accuracy.

As the core driver of high-quality development in digital finance, the digital transformation of financial institutions requires in-depth research into its evolutionary mechanisms and implementation pathways. Based on the framework of evolutionary game theory, this paper constructs a three-party stochastic evolutionary game model involving local governments, financial institutions, and fintech companies. It systematically reveals the dynamic evolutionary mechanisms and strategic choice processes underlying the digital transformation of financial institutions, with a particular focus on analyzing the differential impacts of external environmental uncertainty on the transformation process. The study finds that: (1) the three parties exhibit nonlinear interactions during the evolutionary process, and their strategy choices are significantly interdependent; (2) local government policy interventions have a dual effect on the transformation process: moderate fiscal subsidies can significantly enhance the motivation for financial institutions to transform, but there is a critical threshold for policy effectiveness, and as market mechanisms improve, a policy substitution effect may occur; (3) Under deterministic evolutionary games, the optimal path for financial institutions’ digital transformation is characterized by “government guidance—market coordination—steady-state symbiosis”; however, this path is prone to deviation and distortion under stochastic evolutionary games, particularly during the market coordination phase. Under conditions of stochastic disturbances, the regulatory effectiveness of subsidy policies is significantly weakened, and fintech firms exhibit a stronger risk-averse tendency; The synergistic effects between financial institutions and fintech firms exhibit phased fluctuations, slowing the system’s convergence rate. Based on this, it is recommended that government departments establish a dynamically optimized policy toolkit, while financial institutions should refine their strategic plans for digital transformation. By deepening the dual-wheel synergy mechanism of “data empowerment + technology-driven innovation,” they can build a sustainable digital financial ecosystem.

Under the backdrop of intensifying pressure from global environmental governance and the accelerated advancement of the “dual carbon” goals, heavily polluting enterprises face higher requirements for environmental compliance, stricter emission supervision, and greater pressure to pursue green transformation. This makes them more inclined to engage in greenwashing in concealed and sophisticated ways, thereby gaining social recognition, brand premium, and policy inclinations through a false pro-environmental image. However, existing studies on greenwashing behavior identification mostly rely on single-modal data such as numerical or textual data, which not only lack the integrated application of multi-modal data but also generally overlook the impact of enterprises' top management affiliation relationships on the accuracy of greenwashing behavior identification. Therefore, this paper integrates numerical, textual, and affiliation network data to construct a greenwashing behavior identification model based on the Multi-modal Hierarchical Relational Graph Attention Network (MHRGAT). Empirical results based on 1,691 A-share listed heavily polluting enterprises in Shanghai and Shenzhen from 2014 to 2023 show that: (1) Compared with machine learning methods such as LR and RF, and deep learning models such as CNN and GAT, the MHRGAT model achieves the optimal identification performance. (2) The transmission efficiency of director affiliation information with different historical durations in the network varies, and selecting the director affiliation relationships of enterprises over the past 3 years is relatively appropriate for greenwashing behavior identification. (3) The MHRGAT model based on multi-modal fusion significantly outperforms models using single-modal or dual-modal data in identifying greenwashing behaviors of heavily polluting enterprises across multiple evaluation metrics, demonstrating that multi-modal data fusion enhances the overall performance of greenwashing detection. These research findings can provide regulatory authorities with a greenwashing behavior identification method based on multi-modal fusion and cross-modal attention, thereby guiding heavily polluting enterprises to actively fulfill their environmental responsibilities.

The energy transformation has promoted the development of the lithium-ion battery industry chain. There are complicated international trade relations among the participating economies in the industrial chain. Unexpected events such as geopolitical games make trade or production risks exist in all links of the industrial chain, and the spatio-temporal transmission of risks will pose a threat to the international trade resilience of the entire industrial chain. In the face of risks, positive or negative feedback from other participating economies will affect the international trade resilience of the industrial chain. However, in the current research on international trade in the lithium-ion battery industry chain, there is a lack of studies focusing on the impact of economies' feedback on trade resilience, and most of them do not take into account the production relations among products. Therefore, this paper constructs a “production-trade” multi-layer network model. Through scenario setting and simulation, the international trade resilience of the lithium-ion battery industry chain combined with the positive and negative feedback of the economy is measured, and the following conclusions are obtained: (1) China holds an important position in multiple links of the international trade industrial chain of lithium-ion batteries. Chile and Australia play important positions in the upstream trade. (2) Risk transmission varies in different links of the international trade industrial chain of lithium-ion batteries. Especially lithium carbonate, its risk impact can carry out a long spatio-temporal conduction on a global scale. (3) Positive feedback helps to improve the international trade resilience of the industrial chain, and vice versa. Positive and negative feedback has a more significant impact on the resilience of middle and upstream products. The research results can provide a reference for improving the toughness of international trade of lithium-ion battery industry chain. The research framework proposed in this article can be extended to other industrial chain studies.

This study focuses on two fundamental challenges in the real scenario of financial fraud detection: the limited generalization ability of traditional supervised models under highly imbalanced data, and the high costs of fraud identification. Using annual report data of Chinese A-share listed firms from 2012 to 2022, we construct a financial indicator system consisting of 31 indicators covering thirteen types of fraud schemes, guided by the extended Fraud Triangle Theory, regulatory concerns, and audit practice. Methodologically, this study proposes a generative financial fraud detection framework based on a Convolutional Variational Autoencoder (CNN-VAE). By learning the latent distribution of normal firms’ financial characteristics, the framework identifies anomalous samples and adapts well to scarce fraud samples and imbalanced class distributions. A retrieval-augmented anomaly scoring method is further introduced to characterize potential fraud risk, together with a threshold-optimization mechanism to maximize economic gains from detection. Empirical results show that the proposed method outperforms mainstream machine-learning and deep-learning models across most evaluation metrics, even when benchmark models are enhanced with resampling strategies. Cost analysis further indicates that threshold optimization substantially improves the economic value of fraud detection in the real scenario, enabling most models to achieve positive net benefits.

Exchange rates are characterized by nonlinearity, non-stationarity, and complex fluctuations. Existing methods often suffer from limited prediction accuracy due to single modeling perspectives and inadequate utilization of unstructured data. This paper proposes a combined exchange rate forecasting approach that integrates LSTM and GNN for spatiotemporal feature fusion, enhanced with sentiment analysis based on large language models. First, a pre-trained large model is employed to perform sentiment analysis on unstructured data, quantifying market sentiment scores. These scores are then integrated with historical exchange rate data and fed into a Long Short-Term Memory (LSTM) network to extract temporal dependency features. Second, based on economic theory, relevant macroeconomic indicators are selected to construct a multidimensional graph structure, and a Graph Neural Network (GNN) is used to model the spatial dependencies among variables. Finally, an optimal weighted combination method is applied to integrate the individual predictions, yielding the final exchange rate forecast. To validate the effectiveness of the proposed combined forecasting model, an empirical prediction analysis is conducted on the daily USD/RMB exchange rate from January 2018 to February 2025. The results demonstrate that the proposed method is suitable for forecasting nonlinearly fluctuating exchange rates and achieves higher accuracy compared to existing approaches.

Fractional factorial designs are widely used in industry, agriculture, biopharmaceuticals, high technology and other fields due to their high efficiency. Assessing the optimality of fractional factorial designs has always been an important problem in the area of experimental design. Based on absolute distance and incorporating level permutations of factors, this paper defines the average absolute-moment as a criterion for measuring and screening good designs. Analytical relationships between average absolute moment and generalized word length pattern, moments, and orthogonal vectors for three-level designs are established. Using the wordlength enumerator, a fast method for computing average absolute moment of three-level designs is provided, and its lower bound is derived. This lower bound can serve as a benchmark for evaluating the goodness of designs in various dimensions. Finally, numerical examples are presented to illustrate the theoretical results.

As intelligent transformation reshapes the global manufacturing competitive landscape, how traditional manufacturing firms upgrade to develop New Quality Productivity (NQP) has become a growing research focus. Grounded in a two-factor productivity decomposition perspective, this study develops a partial equilibrium model to examine the impact of intelligent manufacturing on firms' NQP and its underlying mechanisms. We then apply a difference-in-differences design to Chinese A-share listed manufacturing firms. The results show that intelligent manufacturing significantly promotes firms' NQP. Mechanism evidence indicates that intelligent manufacturing reshapes the production function and drives the transition toward NQP through three channels, namely robotics innovation, green innovation, and labor structure optimization, with additional gains arising from their complementarities. Heterogeneity analyses further show that the productivity upgrading effect is more pronounced for technology-intensive firms, capital-intensive firms, and firms in the growth stage. These findings provide theoretical and empirical support for the micro-level pathways through which intelligent manufacturing fosters NQP and inform policies for high-quality development in China's manufacturing sector.

In the era of the digital economy, digital transformation, as an emerging development model for microeconomic entities, has become a new driving force for the transformation, upgrading and high-quality development of traditional brick-and-mortar enterprises. From the perspective of accounting conservatism, this paper takes Shanghai and Shenzhen A-share listed companies from 2007 to 2021 as research samples and constructs a two-way fixed effects model for empirical testing. The results show that corporate digital transformation can significantly improve accounting conservatism, and the digital transformation of peer enterprises has a spillover effect on the improvement of accounting conservatism. In particular, the application of underlying digital technologies exerts a more prominent promoting effect on accounting conservatism. Mechanism tests indicate that corporate digital transformation improves accounting conservatism through multiple channels: increasing corporate information transparency, enhancing internal control quality, reducing real earnings management behaviors, and lowering operational uncertainty. Heterogeneity analysis shows that the positive impact of digital transformation on accounting conservatism is more pronounced in enterprises with higher environmental uncertainty, enterprises in non-digital industries, and large-scale enterprises; meanwhile, the higher the industry information transparency, the more significant the industry spillover effect of digital transformation. This study enriches the literature on the economic consequences of corporate digital transformation and the influencing factors of accounting conservatism, and provides important reference and enlightenment for realizing coordinated industrial development against the background of digital transformation and helping traditional physical enterprises better capture digital dividends.