arXiv subjects

Yuyu Chen

Publications and source records attributed to Yuyu Chen.

At least 19 recordsLinked to original sources

arXiv2026

Directional AI Advice: Experimental Evidence from Healthcare

Generative AI is fast becoming the first place people turn for expert advice. The advice it provides can be directional rather than neutral, shaped in part by the choices of its designers and regulators. When clients consult AI before meeting an expert, they carry this directional advice into a relationship that once rested on the expert's judgment alone. We study its consequences in healthcare through a large-scale preregistered field experiment at a Chinese hospital, where we randomize patients' access to an AI chatbot before their outpatient visit. Examination of the conversation logs shows that the chatbot routinely cautions against the use of medications, especially Traditional Chinese Medicine and antibiotics, while issuing clean recommendations for diagnostic testing, consistent with the liability-driven guardrails encoded in AI training. This directionality propagates into clinical practice. Prescription rates decline among treated patients while diagnostic testing increases, and these effects are more pronounced among physicians who are receptive to patient input and those with more intensive prescribing styles. Beyond shifting healthcare utilization, survey results show that AI access reduces patient compliance and satisfaction, shifting the balance of authority between patients and physicians.

econ.GN↗

arXiv2026

Universal Value-at-Risk superadditivity

Value-at-Risk (VaR) is a standard regulatory risk measure, and its failure of subadditivity is well known. Much less appreciated is that for sufficiently heavy-tailed losses, VaR can be superadditive uniformly across all probability levels, a phenomenon strictly stronger than the asymptotic superadditivity studied in extreme value theory. We call this property universal VaR superadditivity (UVS). We study UVS and its stronger weighted version (WUVS) as properties of random vectors rather than of marginal distributions. This perspective unifies and extends a recent line of work on iid infinite-mean models. UVS, except for trivial cases, imposes an infinite-mean structure. We establish preservation properties of UVS and WUVS under increasing and convex transformations, weak convergence, and certain distributional mixtures, and use these tools to prove UVS and WUVS for non-identically distributed risks in several large families including completely subscalable, super-Cauchy, and inverted subadditive risks, extending results previously available only in the iid case. In many results, we also establish strict versions of UVS and WUVS, which lead to stronger decision-theoretic implications. As a consequence, for any portfolio satisfying WUVS, every distortion risk measure is superadditive, so an optimal allocation concentrates on a single asset, and diversification is never beneficial.

q-fin.RM↗

arXiv2026

TabSODA: Tabular Diffusion based Imputation with Skip Pattern Detection and Ordinal Awareness

Missing data imputation in large-scale surveys faces two challenges that are not well handled by current tabular diffusion methods. First, \emph{structural skips}, cells made inapplicable by questionnaire design, should not be imputed but are often conflated with item nonresponse. Second, \emph{ordinal} responses encode ordered categories, yet most pipelines treat them as nominal levels through one-hot or analog-bit encodings. We introduce \textbf{TabSODA} (\textbf{Tab}ular diffusion with \textbf{S}kip pattern detection and \textbf{O}r\textbf{d}inal \textbf{A}wareness), an Expectation-Maximization (EM)-based diffusion imputer built on the Elucidated Diffusion Model (EDM) framework. TabSODA propagates structural skips through the denoising loss and reverse-time sampler, and represents ordinal variables with cumulative-probit scalar latents while retaining analog-bit encodings for nominal variables. When a codebook skip mask is available, TabSODA uses it directly; otherwise, the TabSODA+SKIP variant estimates the mask from raw responses and questionnaire order using a CART-based skip-pattern miner. On Population Assessment of Tobacco and Health (PATH) study and the National Survey on Drug Use and Health (NSDUH), two nationally representative U.S.\ surveys, TabSODA reduces ordinal MACE by up to $23.7\%$ and improves categorical accuracy by up to $9\%$ over the strongest baseline across MCAR, MAR, and MNAR masking. The skip miner achieves near-perfect precision on both datasets, allowing TabSODA+SKIP to closely track the codebook-mask variant.

stat.ML↗

arXiv2026

Capital allocation and tail central moments for the multivariate normal mean-variance mixture distribution

Capital allocation is a procedure used to assess the risk contributions of individual risk components to the total risk of a portfolio. While the conditional tail expectation (CTE)-based capital allocation is arguably the most popular capital allocation method, its inability to reflect important tail behaviour of losses necessitates a more accurate approach. In this paper, we introduce a new capital allocation method based on the tail central moments (TCM), generalising the tail covariance allocation informed by the tail variance. We develop analytical expressions of the TCM as well as the TCM-based capital allocation for the class of normal mean-variance mixture distributions, which is widely used to model asymmetric and heavy-tailed data in finance and insurance. As demonstrated by a numerical analysis, the TCM-based capital allocation captures several significant patterns in the tail region of equity losses that remain undetected by the CTE, enhancing the understanding of the tail risk contributions of risk components.

q-fin.PM↗

arXiv2025

Fine-Tuned Language Models for Domain-Specific Summarization and Tagging

This paper presents a pipeline integrating fine-tuned large language models (LLMs) with named entity recognition (NER) for efficient domain-specific text summarization and tagging. The authors address the challenge posed by rapidly evolving sub-cultural languages and slang, which complicate automated information extraction and law enforcement monitoring. By leveraging the LLaMA Factory framework, the study fine-tunes LLMs on both generalpurpose and custom domain-specific datasets, particularly in the political and security domains. The models are evaluated using BLEU and ROUGE metrics, demonstrating that instruction fine-tuning significantly enhances summarization and tagging accuracy, especially for specialized corpora. Notably, the LLaMA3-8B-Instruct model, despite its initial limitations in Chinese comprehension, outperforms its Chinese-trained counterpart after domainspecific fine-tuning, suggesting that underlying reasoning capabilities can transfer across languages. The pipeline enables concise summaries and structured entity tagging, facilitating rapid document categorization and distribution. This approach proves scalable and adaptable for real-time applications, supporting efficient information management and the ongoing need to capture emerging language trends. The integration of LLMs and NER offers a robust solution for transforming unstructured text into actionable insights, crucial for modern knowledge management and security operations.

cs.CL↗

arXiv2025

Optimal insurance design with Lambda-Value-at-Risk

This paper explores optimal insurance solutions based on the Lambda-Value-at-Risk ($Λ\VaR$). If the expected value premium principle is used, our findings confirm that, similar to the VaR model, a truncated stop-loss indemnity is optimal in the $Λ\VaR$ model. We further provide a closed-form expression of the deductible parameter under certain conditions. Moreover, we study the use of a $Λ'\VaR$ as premium principle as well, and show that full or no insurance is optimal. Dual stop-loss is shown to be optimal if we use a $Λ'\VaR$ only to determine the risk-loading in the premium principle. Moreover, we study the impact of model uncertainty, considering situations where the loss distribution is unknown but falls within a defined uncertainty set. Our findings indicate that a truncated stop-loss indemnity is optimal when the uncertainty set is based on a likelihood ratio. However, when uncertainty arises from the first two moments of the loss variable, we provide the closed-form optimal deductible in a stop-loss indemnity.

q-fin.RM↗

arXiv2025

Understanding Inconsistent State Update Vulnerabilities in Smart Contracts

Smart contracts enable contract terms to be automatically executed and verified on the blockchain, and recent years have witnessed numerous applications of them in areas such as financial institutions and supply chains. The execution logic of a smart contract is closely related to the contract state, and thus the correct and safe execution of the contract depends heavily on the precise control and update of the contract state. However, the contract state update process can have issues. In particular, inconsistent state update issues can arise for reasons such as unsynchronized modifications. Inconsistent state update bugs have been exploited by attackers many times, but existing detection tools still have difficulty in effectively identifying them. This paper conducts the first large-scale empirical study about inconsistent state update vulnerabilities (that is, inconsistent state update bugs that are exploitable) in smart contracts, aiming to shed light for developers, researchers, tool builders, and language or library designers in order to avoid inconsistent state update vulnerabilities. We systematically investigate 116 inconsistent state update vulnerabilities in 352 real-world smart contract projects, summarizing their root causes, fix strategies, and exploitation methods. Our study provides 11 original and important findings, and we also give the implications of our findings. To illustrate the potential benefits of our research, we also develop a proof-of-concept checker based on one of our findings. The checker effectively detects issues in 64 popular GitHub projects, and 19 project owners have confirmed the detected issues at the time of writing. The result demonstrates the usefulness and importance of our findings for avoiding inconsistent state update vulnerabilities in smart contracts.

cs.SE↗

arXiv2025

Gradient boosted multi-population mortality modelling with high-frequency data

High-frequency mortality data have attracted growing attention, but their use has largely been confined to specific applications rather than general modelling and forecasting. Such data pose new challenges to traditional mortality models due to pronounced seasonal patterns and short-term fluctuations. To address these challenges and produce more accurate forecasts with the high-frequency mortality data, this paper introduces a novel integration of gradient boosting techniques into traditional stochastic mortality models under a multi-population setting. Our key innovation lies in using the Li and Lee model as the weak learner within the gradient boosting framework, replacing conventional decision trees. Empirical studies are conducted using weekly mortality data from 30 countries (Human Mortality Database, 2015-2019). Empirical evidence highlights that the proposed methodology not only enhances model fit by accurately capturing underlying mortality trends and seasonal patterns, but also achieves superior forecast accuracy, compared to the benchmark models. We also investigate a key challenge in multi-population mortality modelling: how to select appropriate sub-populations with sufficiently similar mortality experiences. A comprehensive clustering exercise is conducted based on mortality improvement rates and seasonal strength. The empirical results demonstrate that our proposed model maintains strong forecast accuracy across different clustering configurations, thereby reducing the need for extensive data preprocessing.

stat.AP↗

arXiv2025

De-Simplifying Pseudo Labels to Enhancing Domain Adaptive Object Detection

Despite its significant success, object detection in traffic and transportation scenarios requires time-consuming and laborious efforts in acquiring high-quality labeled data. Therefore, Unsupervised Domain Adaptation (UDA) for object detection has recently gained increasing research attention. UDA for object detection has been dominated by domain alignment methods, which achieve top performance. Recently, self-labeling methods have gained popularity due to their simplicity and efficiency. In this paper, we investigate the limitations that prevent self-labeling detectors from achieving commensurate performance with domain alignment methods. Specifically, we identify the high proportion of simple samples during training, i.e., the simple-label bias, as the central cause. We propose a novel approach called De-Simplifying Pseudo Labels (DeSimPL) to mitigate the issue. DeSimPL utilizes an instance-level memory bank to implement an innovative pseudo label updating strategy. Then, adversarial samples are introduced during training to enhance the proportion. Furthermore, we propose an adaptive weighted loss to avoid the model suffering from an abundance of false positive pseudo labels in the late training period. Experimental results demonstrate that DeSimPL effectively reduces the proportion of simple samples during training, leading to a significant performance improvement for self-labeling detectors. Extensive experiments conducted on four benchmarks validate our analysis and conclusions.

cs.CV↗

arXiv2025

Risk exchange under infinite-mean Pareto models

We study the optimal decisions and equilibria of agents who aim to minimize their risks by allocating their positions over extremely heavy-tailed (i.e., infinite-mean) and possibly dependent losses. The loss distributions of our focus are super-Pareto distributions, which include the class of extremely heavy-tailed Pareto distributions. Using a recent result on stochastic dominance, we show that for a portfolio of super-Pareto losses, non-diversification is preferred by decision makers equipped with well-defined and monotone risk measures. The phenomenon that diversification is not beneficial in the presence of super-Pareto losses is further illustrated by an equilibrium analysis in a risk exchange market. First, agents with super-Pareto losses will not share risks in a market equilibrium. Second, transferring losses from agents bearing super-Pareto losses to external parties without any losses may arrive at an equilibrium which benefits every party involved.

q-fin.RM↗

arXiv2025

Risk aggregation and stochastic dominance for a class of heavy-tailed distributions

We introduce a new class of heavy-tailed distributions for which any weighted average of independent and identically distributed random variables is larger than one such random variable in (usual) stochastic order. We show that many commonly used extremely heavy-tailed (i.e., infinite-mean) distributions, such as the Pareto, Fréchet, and Burr distributions, belong to this class. The established stochastic dominance relation can be further generalized to allow negatively dependent or non-identically distributed random variables. In particular, the weighted average of non-identically distributed random variables dominates their distribution mixtures in stochastic order.

math.PR↗

arXiv2025

Stochastic dominance for linear combinations of infinite-mean risks

In this paper, we establish a sufficient condition to compare linear combinations of independent and identically distributed (iid) infinite-mean random variables under usual stochastic order. We introduce a new class of distributions that includes many commonly used heavy-tailed distributions and show that within this class, a linear combination of random variables is stochastically larger when its weight vector is smaller in the sense of majorization order. We proceed to study the case where each random variable is a compound Poisson sum and demonstrate that if the stochastic dominance relation holds, the summand of the compound Poisson sum belongs to our new class of distributions. Additional discussions are presented for stable distributions.

math.PR↗

arXiv2025

Sub-uniformity of harmonic mean p-values

We obtain several inequalities on the generalized means of dependent p-values. In particular, the weighted harmonic mean of p-values is strictly sub-uniform under several dependence assumptions of p-values, including independence, negative upper orthant dependence, the class of extremal mixture copulas, and some Clayton copulas. Sub-uniformity of the harmonic mean of p-values has an important implication in multiple hypothesis testing: It is statistically invalid (anti-conservative) to merge p-values using the harmonic mean unless a proper threshold or multiplier adjustment is used, and this applies across all significance levels. The required multiplier adjustment on the harmonic mean p-value grows sub-linearly to infinity as the number of p-values increases, and hence there does not exist a constant multiplier that works for any number of p-values, even under independence.

math.ST↗

arXiv2025

Diversification for infinite-mean Pareto models without risk aversion

We study stochastic dominance between portfolios of independent and identically distributed (iid) extremely heavy-tailed (i.e., infinite-mean) Pareto random variables. With the notion of majorization order, we show that a more diversified portfolio of iid extremely heavy-tailed Pareto random variables is larger in the sense of first-order stochastic dominance. This result is further generalized for Pareto random variables caused by triggering events, random variables with tails being Pareto, bounded Pareto random variables, and positively dependent Pareto random variables. These results provide an important implication in investment: Diversification of extremely heavy-tailed Pareto profits uniformly increases investors' profitability, leading to a diversification benefit. Remarkably, different from the finite-mean setting, such a diversification benefit does not depend on the decision maker's risk aversion.

q-fin.PM↗

arXiv2024

Infinite-mean models in risk management: Discussions and recent advances

In statistical analysis, many classic results require the assumption that models have finite mean or variance, including the most standard versions of the laws of large numbers and the central limit theorems. Such an assumption may not be completely innocent, and it may not be appropriate for datasets with heavy tails (e.g., catastrophic losses), relevant to financial risk management. In this paper, we discuss the importance of infinite-mean models in economics, finance, and related fields, with recent results and examples. We emphasize that many results or intuitions that hold for finite-mean models turn out to fail or even flip for infinite-mean models. Due to the breakdown of standard thinking for infinite-mean models, we argue that if the possibility of using infinite-mean models cannot be excluded, great caution should be taken when applying classic methods that are usually designed for finite-mean cases in finance and insurance.

q-fin.RM↗

arXiv2024

An unexpected stochastic dominance: Pareto distributions, dependence, and diversification

We find the perhaps surprising inequality that the weighted average of independent and identically distributed Pareto random variables with infinite mean is larger than one such random variable in the sense of first-order stochastic dominance. This result holds for more general models including super-Pareto distributions, negative dependence, and triggering events, and yields superadditivity of the risk measure Value-at-Risk for these models.

q-fin.RM↗

arXiv2021

Risk Aggregation under Dependence Uncertainty and an Order Constraint

We study the aggregation of two risks when the marginal distributions are known and the dependence structure is unknown, under the additional constraint that one risk is smaller than or equal to the other. Risk aggregation problems with the order constraint are closely related to the recently introduced notion of the directional lower (DL) coupling. The largest aggregate risk in concave order (thus, the smallest aggregate risk in convex order) is attained by the DL coupling. These results are further generalized to calculate the best-case and worst-case values of tail risk measures. In particular, we obtain analytical formulas for bounds on Value-at-Risk. Our numerical results suggest that the new bounds on risk measures with the extra order constraint can greatly improve those with full dependence uncertainty.

q-fin.RM↗

arXiv2021

Trade-off between validity and efficiency of merging p-values under arbitrary dependence

Various methods of combining individual p-values into one p-value are widely used in many areas of statistical applications. We say that a combining method is valid for arbitrary dependence (VAD) if it does not require any assumption on the dependence structure of the p-values, whereas it is valid for some dependence (VSD) if it requires some specific, perhaps realistic but unjustifiable, dependence structures. The trade-off between validity and efficiency of these methods is studied via analyzing the choices of critical values under different dependence assumptions. We introduce the notions of independence-comonotonicity balance (IC-balance) and the price for validity. In particular, IC-balanced methods always produce an identical critical value for independent and perfectly positively dependent p-values, a specific type of insensitivity to a family of dependence assumptions. We show that, among two very general classes of merging methods commonly used in practice, the Cauchy combination method and the Simes method are the only IC-balanced ones. Simulation studies and a real data analysis are conducted to analyze the sizes and powers of various combining methods in the presence of weak and strong dependence.

math.ST↗