Current churn prediction research mainly emphasizes static accuracy on curated datasets, and they often overlook label noise arising from heuristic CRM rules, uncontrollable user attrition, and suboptimal retention timing. This study proposes a robust two-stage retention framework targeting e-commerce users with tenure of more than four years. Firstly, structural churn cases are filtered from the dataset. Then, 5% label noise is added to simulate real-world annotation errors, and benchmark evaluations are conducted across nine machine learning models. Ensemble methods, specifically GBM, XGBoost, and LightGBM, achieve F1 scores ranging from 0.43 to 0.44 with cross-validation standard deviations below 0.02. About the method, the analysis progresses from binary classification to Cox proportional hazards modeling, identifying a proactive intervention window between the fourth and fifth year of user tenure. Business-oriented threshold tuning enables GBM campaigns to attain a 345.7% return on investment. This significantly surpasses simple methods. Further analysis via SHAP values shows service call fatigue is the primary churn driver. The assumed impact of fee sensitivity lacks empirical support within this mature cohort. The proposed project effectively bridges statistical prediction and operational retention, providing an effective methodology for protecting high-value customer assets.

The Black–Scholes option pricing model assumes that the volatility of the underlying asset is constant, yet observed market option prices imply a volatility that changes with the strike price. This study investigates whether adjusting the volatility input to match market prices improves option pricing and delta hedging, and whether the benefit differs between a high-volatility individual stock and a broad market index. Daily option data for Tesla and the SPDR S&P 500 exchange-traded fund from 2019 to 2022 are used, covering calm markets, the 2020 market crash, and the subsequent monetary tightening. Two settings are compared: one using historical volatility estimated from past returns, and one using an implied volatility obtained by fitting the Black–Scholes price to market option prices through least squares. Pricing accuracy is assessed by the pricing residuals across strike prices, and hedging effectiveness by the variability of the hedging profit and loss over forty-eight monthly positions. The results show that a single fitted volatility cannot match option prices at all strike prices, confirming that the constant-volatility assumption does not hold. Using the fitted volatility reduces the variability of the hedging profit and loss by about twelve percent for the individual stock and twenty-three percent for the index. The larger improvement for the index indicates that the value of calibration depends on the volatility structure of the underlying asset.

How Artificial Intelligence reshapes wealth distribution has become a key issue. Existing research is mainly divided into two independent branches: within-country and between-country studies. This paper aims to integrate these two dimensions by constructing a unified analytical framework. Firstly, the three factors affecting the distribution effect of Artificial Intelligence have been identified as the industrial structure, institutional environment and position in the Global Value Chain. Based on the empirical evidence from the United States, China and Latin America, the review shows that the inequality patterns within different countries and regions vary due to institutional and structural differences. At the international level, this paper links forecast data from the International Monetary Fund with potential causal paths to show how Artificial Intelligence is concentrating high-value-added activities in developed countries and driving developing countries into low-skilled and easily replaceable segments, thus widening the North-South divide. By connecting these two perspectives, this review provides a more comprehensive understanding of the uneven distribution of wealth caused by Artificial Intelligence.

Credit card fraud has become a central issue for the financial security of banks and cardholders. However, fraudulent transactions account for only a very small proportion of all transaction cases. If it is missed, it may directly cause huge property damage. As a result, this study focuses on how to identify fraudulent transactions as much as possible, while controlling false positives and human audit costs. Specifically, logistic regression, gradient boosting, XGBoost, and random forest are compared, while class weighting, SMOTE, and Random Undersampling are evaluated for handling class imbalance. It is more critical to translate the model results into a risk warning system and a loss simulation system. The results show that random forest with class weighting has the best overall performance, with 90.6% F1-Score, 94.1% precision, and 87.3% recall. Meanwhile, under the cost assumptions, this model reduces simulation costs by 85.4% relative to the no-model baseline. These findings suggest that in data environments with severe class imbalances, a machine learning model can be designed as an effective triage tool. Its actual value is not only to predict fraud, but also to establish a low-risk release, medium-risk verification, and high-risk manual review of the decision-making process for institutions and truly reduce losses.

The complexity and dynamics of the stock market make traditional absolute price prediction models based on supervised learning face great risks and limitations in practical business applications. This report proposes an innovative Hybrid Deep Reinforcement Learning trading framework to shift the forecast target from stock prices to optimal trading decisions in specific market conditions. The dataset for this experiment comes from a Malaysian listed company, GAMUDA BHD, from 2000 to 2024 on the Kaggle platform, and builds a decision-making structure with three stages. Use the Gaussian Mixture Model, the long-term and short-term memory network, and finally the Deep Q-Network to train the intelligent body to learn the optimal discrete trading strategy (buy, sell, hold) in a continuous complex environment. The research results show that, compared with the traditional buy-and-hold strategy and the single intensive learning model, the hybrid architecture that integrates market state perception shows significant commercial application value in terms of maximum pullback control and Sharp rate improvement. During a systemic market downturn, the GMM clusterer detected a surge in volatility, identifying the crisis. Based on this, DQN forced a "sell" order, successfully limiting the maximum drawdown. This strongly demonstrates the significance of hybrid models in mitigating black swan events in the market.

Against the backdrop of the global consensus on carbon neutrality, an irreversible industrial revolution driven by energy transition has arrived. The global automotive industry landscape is being reshaped, transitioning from the internal combustion engine era to an era of electrification and intelligence.Therefore, technology-based auto manufacturers, such as NIO Inc., Li Auto Inc., XPeng Inc., and Tesla, Inc., have seized the historical opportunity to make a strategic leap and successfully secure leading positions. In addition, companies that develop new energy vehicles can mitigate the risk of dependence on foreign energy sources. This paper analyzes the annual financial reports of four new energy vehicle companies (NIO Inc., Li Auto Inc., XPeng Inc., and Tesla, Inc.) from 2021 to 2025. By employing quantitative numerical analysis, three aspects-namely solvency, operational capability, and profitability-are evaluated to identify optimal investment targets. Consequently, risk-tolerant investors are advised to choose NIO Inc. and XPeng Inc., whereas risk-averse investors may prefer Li Auto Inc. and Tesla, Inc.

In today's era of increasingly advanced information technology across various fields, deep learning models such as large-scale AI models serve as powerful algorithmic assistants, effectively driving the positive development of numerous enterprises and capital markets. However, more complex and covert information disclosure frauds have also emerged accordingly. This paper selects Xiamen Red Phase Inc. as the research subject, which features newly implemented penalties, a complete and continuous fraud cycle, and diverse methods. The study first reviews and summarizes existing domestic and international research on information disclosure. It then traces the timeline of Red Phase Inc.'s illegal information disclosure and fraud case, using the fraud triangle theory as the research framework. From the perspectives of pressure, opportunity, and rationalization, relevant data is incorporated to analyze corporate information disclosure fraud (particularly financial fraud). Through case studies, the motivations and potential consequences of corporate information disclosure fraud are further examined, ultimately proposing relevant governance recommendations.

Limit order book (LOB) mid-price prediction plays an important role in high-frequency trading. This paper presents a systematic evaluation of deep learning models and hyperparameter settings for LOB prediction using the FI-2010 benchmark. Under the same experimental settings, eleven representative neural network architectures, including recurrent, CNN-based, bilinear-based, and transformer-based models are compared. The results show that BiNTABL achieves the best overall performance in terms of macro-F1, precision, recall, and accuracy. Extensive ablation experiments are then conducted to investigate the effects of network layer number, learning rate, hidden dimensions, batch size, historical input length, and prediction horizon. The results show that prediction performance is highly sensitive to these settings. And the existence of bilinear layer, learning rate together with forecast horizon are the top three most important factors in model performance. The findings provide practical guidance for neural network design and hyperparameter selection in LOB prediction and offer a useful reference for future research on financial time-series forecasting.

Global supply chains are constantly hit by unpredictable disruptions, including extreme weather and transport congestion, rendering conventional experience-based risk management slow and passive. Lean TPS effectively reduces operational variations yet struggles with fragmented data and limited end-to-end visibility across logistics networks. Relevant empirical research on combining big data analytics and lean risk control within large domestic logistics providers still remains insufficient. This study examines how Big Data Analytics Capability (BDAC) can strengthen Lean/TPS-based risk governance in JD Logistics by improving end-to-end visibility, speeding detection, and enabling proactive, data-driven responses. Using JD Logistics as a focal case, this paper explores how technologies like JD Super Brain and digital twins translate Lean concepts—jidoka, standard work, and Kaizen—into scalable, real-time capabilities across regions and partners. The findings reveal that while BDAC enhances sensing and cadence, persistent gaps remain in emergency escalation, cross-regional handoffs, and learning diffusion. Durable resilience emerges when predictive analytics are tightly integrated with scenario-based standard work, unified data governance, and a network-wide Kaizen pipeline.

Option pricing has moved away from the previous analytical and stochastic models to machine-learning-based methods. Therefore, a good model of the market and an efficient one for computation is required. This paper reviews the development of the Black-Scholes and Heston models to DNN- and DDN-based surrogate models, as well as their applications in quantitative finance. Give some attention to computational efficiency in the repeated pricing and calibration. Another issue in the paper is the trade-off between the accuracy of prices and the cost of computation; high-speed calculation may not be entirely precise in finance. Differentiability in the context of risk management and hedging, as well as Delta and Vega sensitivity measures, will be introduced. The problems with the current ones are: dependence on training data, lack of interpretability, and issues of robustness and generalization. In the future, one will aim to build a more stable and flexible surrogate model, enhance its interpretability, address financial constraints, and integrate traditional financial models with machine learning in a hybrid manner.