Interpretable hybrid credit scoring for thin-file and underbanked populations
研究通过结合逻辑回归和梯度提升方法，为薄文件和欠银行服务人群提供可解释的信用评分，并在东非数据上进行实证分析与公平性审计。
0 citationsRead paper
Selected work
Representative Papers
Policy-Invariant Reward Shaping from LLM Feedback: A Framework for Hybrid RL Agents
研究解决了结合大语言模型与强化学习时奖励信号理论地位不明确的问题，通过将LLM反馈作为有界势函数，确保即使在LLM评分不准确的情况下也能保持最优策略集。
0 citationsRead paper
Intelligent resource allocation in wireless networks via deep reinforcement learning
This work addresses the challenge of optimal power allocation in stochastic wireless networks without requiring an accurate system model. By formulating resource scheduling as a Markov decision process, the authors employ a deep Q-network (DQN) to learn an adaptive power control policy directly from observed channel states. As the first model-free approach leveraging deep reinforcement learning to achieve performance close to the theoretical optimum, the proposed method attains a system throughput of 3.88 Mbps—nearly matching the water-filling algorithm’s upper bound—and improves upon random and fixed allocation strategies by 73% and 27%, respectively. Moreover, the solution maintains high fairness and energy efficiency, achieving a Jain’s fairness index of 0.91.
0 citationsRead paper
Applications of synthetic financial data in portfolio and risk modeling
Empirical financial research is hindered by limited access to real market data due to privacy constraints and data scarcity. Method: This paper systematically evaluates TimeGAN and VAE for generating synthetic financial return series, proposing a multidimensional validation framework grounded in statistical similarity metrics, temporal structure tests, and mean–variance optimization. Contribution/Results: We demonstrate—first in a financial context—that TimeGAN faithfully replicates key stylized facts, including volatility clustering, autocorrelation, and extreme-event dynamics; synthetic returns yield portfolio weights, Sharpe ratios, and risk metrics deviating by less than 8% from empirical counterparts. In contrast, VAEs—though more stable in training—substantially underrepresent tail risk. Our work establishes an empirical benchmark for deploying synthetic data in downstream financial modeling (e.g., portfolio construction, trading strategy analysis, and risk modeling), enabling reproducible, low-cost, privacy-preserving quantitative research.
0 citationsRead paper
Send to which account? Evaluation of an LLM-based Scambaiting System
Traditional detection and response mechanisms are ineffective against large-scale, generative-AI-driven phishing scams, as they fail to disrupt the underlying financial infrastructure—such as mule accounts and cryptocurrency wallets—that enables fraud. Method: We propose the first large-scale, real-world deployed LLM-powered conversational honeypot system, which engages scammers in human-like dialogues to proactively elicit sensitive financial information. Our approach introduces a rule-augmented dialogue management framework and formally defines two core evaluation metrics: *information leakage rate* and *human acceptance rate*. Contribution/Results: Over a five-month field deployment, the system conducted 2,600+ interactions and collected 18,000 messages, achieving a 32% information leakage rate and a 70% human acceptance rate. It successfully identified critical accounts involved in fund transfers, empirically validating the feasibility and effectiveness of proactive anti-fraud paradigms.
0 citationsRead paper
Recent publications
Latest Papers
Interpretable hybrid credit scoring for thin-file and underbanked populations
研究通过结合逻辑回归和梯度提升方法，为薄文件和欠银行服务人群提供可解释的信用评分，并在东非数据上进行实证分析与公平性审计。
0 citationsRead paper
Policy-Invariant Reward Shaping from LLM Feedback: A Framework for Hybrid RL Agents
研究解决了结合大语言模型与强化学习时奖励信号理论地位不明确的问题，通过将LLM反馈作为有界势函数，确保即使在LLM评分不准确的情况下也能保持最优策略集。
0 citationsRead paper
Intelligent resource allocation in wireless networks via deep reinforcement learning
This work addresses the challenge of optimal power allocation in stochastic wireless networks without requiring an accurate system model. By formulating resource scheduling as a Markov decision process, the authors employ a deep Q-network (DQN) to learn an adaptive power control policy directly from observed channel states. As the first model-free approach leveraging deep reinforcement learning to achieve performance close to the theoretical optimum, the proposed method attains a system throughput of 3.88 Mbps—nearly matching the water-filling algorithm’s upper bound—and improves upon random and fixed allocation strategies by 73% and 27%, respectively. Moreover, the solution maintains high fairness and energy efficiency, achieving a Jain’s fairness index of 0.91.
0 citationsRead paper
Applications of synthetic financial data in portfolio and risk modeling
Empirical financial research is hindered by limited access to real market data due to privacy constraints and data scarcity. Method: This paper systematically evaluates TimeGAN and VAE for generating synthetic financial return series, proposing a multidimensional validation framework grounded in statistical similarity metrics, temporal structure tests, and mean–variance optimization. Contribution/Results: We demonstrate—first in a financial context—that TimeGAN faithfully replicates key stylized facts, including volatility clustering, autocorrelation, and extreme-event dynamics; synthetic returns yield portfolio weights, Sharpe ratios, and risk metrics deviating by less than 8% from empirical counterparts. In contrast, VAEs—though more stable in training—substantially underrepresent tail risk. Our work establishes an empirical benchmark for deploying synthetic data in downstream financial modeling (e.g., portfolio construction, trading strategy analysis, and risk modeling), enabling reproducible, low-cost, privacy-preserving quantitative research.
0 citationsRead paper
Send to which account? Evaluation of an LLM-based Scambaiting System
Traditional detection and response mechanisms are ineffective against large-scale, generative-AI-driven phishing scams, as they fail to disrupt the underlying financial infrastructure—such as mule accounts and cryptocurrency wallets—that enables fraud. Method: We propose the first large-scale, real-world deployed LLM-powered conversational honeypot system, which engages scammers in human-like dialogues to proactively elicit sensitive financial information. Our approach introduces a rule-augmented dialogue management framework and formally defines two core evaluation metrics: *information leakage rate* and *human acceptance rate*. Contribution/Results: Over a five-month field deployment, the system conducted 2,600+ interactions and collected 18,000 messages, achieving a 32% information leakage rate and a 70% human acceptance rate. It successfully identified critical accounts involved in fund transfers, empirically validating the feasibility and effectiveness of proactive anti-fraud paradigms.
0 citationsRead paper