Currently, the U.S. financial and credit sector faces several critical technological challenges. Limitations of traditional credit scoring: the FICO model, which relies on historical data, struggles to assess the true credit risk of groups such as young people and new immigrants. Inaccurate risk assessment: traditional customer risk assessment models typically rely on historical data and fail to adequately consider changes in the economic environment or the real-time financial status of borrowers. These models are not responsive to customer behavior and market fluctuations. Existing models also underestimate the default risks associated with high debt levels and sudden economic shocks (such as pandemics). As a result, U.S. credit card debt has surpassed $1 trillion, with approximately 6% of credit card borrowers in serious arrears. Saturation of the credit market and inefficient conversion: credit products have become highly homogeneous, and the efficiency of customer matching is low, resulting in sluggish industry growth. These problems are rooted in the disconnect between static models and the complex, dynamic market. This not only affects the overall operational efficiency and profitability of the U.S. financial industry but also has a profound impact on consumers and the broader U.S. economy.
Ye Ziyi, a Chief Data Scientist specializing in financial credit technology, has creatively applied technologies such as the Multi-armed Bandit Algorithm (including simple MAB and contextual MAB), Off-Policy Evaluation, Boosting algorithms, Momentum, and Inverse Propensity Scoring. She has also developed and deployed several large-scale financial models, such as the simple multi-armed bandit model with multi-stage design, the multi-stage account-level valuation model, and the contextual MAB recommendation system. She has provided replicable solutions for the entire industry to address existing technological challenges in the U.S. financial credit sector, and achieved breakthroughs in key areas such as credit assessment, risk prediction, and customer conversion in the credit market.
Dynamic Credit Assessment Breaks the Rigid Dilemma of FICO
In credit assessment, the simple multi-armed bandit model with multi-stage design created a feature selection pipeline for multi-armed bandit models, enabling the dynamic screening of high-value features (such as the addition of “mobile payment stability” or “frequency of occupation changes”) and real-time adjustment of feature weights. This significantly improved the loan approval rate for new immigrants. Moreover, the model uses off-policy evaluation for simulation experiments, testing the effects of different feature combinations on historical data, thus avoiding the risks associated with online testing. Overall, this model successfully breaks the data limitations of the FICO credit scoring model, enabling dynamic iteration and personalization of credit scoring.
Accurate Risk Assessment Reduces Default Rates
In terms of risk assessment, the multi-stage account-level valuation model innovatively combines boosting algorithms with a generalized linear models to predict the income of any given account in a particular month. This not only optimizes loan amounts and interest rates based on the borrowers’ repayment behavior and historical data, but also dynamically adjusts loan approval strategies to balance risk and return. Additionally, it helps companies estimate future income from an account and provides key metrics for critical decision-making. During the pandemic, this model successfully addressed the problems of insufficient risk forecasting for high debt ratios and default risks in the U.S. credit sector, by evaluating the credit behavior and potential value of high-risk user groups. It helped companies achieve a balance between risk and return, significantly reduced overall bad debt and overdue rates.
Improving Customer Conversion Rates for Business Growth
In terms of customer conversion, the contextual MAB recommendation system creatively applies new technologies, such as Momentum and Inverse Propensity Scoring. Momentum technology analyzes the frequency and intensity of recent user behaviors to identify shifts in user interests, allowing for timely adjustments to recommendation strategies, thus improving both the timeliness and accuracy of recommendations. Inverse Propensity Scoring is a technique used to correct selection bias, addressing model bias caused by data imbalances in recommendation systems and ensuring that low-frequency but high-value options are thoroughly explored. These two technologies work together to progressively optimize recommendation strategies based on real-time user behavior and historical data, dynamically recommending the most suitable credit products to the right customers and maximizing user conversion rates.
Breaking the Traditional Interest Rate Testing Dilemma, Optimizing Interest Rate Resting with Dynamic Algorithms
Traditional A/B testing methods often result in significant profit loss when optimizing interest rate models. In contrast, the simple multi-armed bandit model with multi-stage design allows for precise control over test group proportions and interest rate adjustments.
These groundbreaking technologies have been standardized through three core software copyrights self-developed by her, including the multi-armed bandit strategy intelligent evaluation selection system, the economical decision support algorithm verification platform, and the economic forecasting model performance evaluation system, they have been licensed to other organizations and generated economic benefits. Ye Ziyi stated: “The future of the credit industry belongs to ‘dynamic models’ – that can adapt to environmental changes like living organisms. Our work proves that through the integration of technological innovation and domain knowledge, we can simultaneously achieve the ‘impossible triangle’ of risk control, user experience, and business value.” Her team is extending these technologies to the small and medium-sized enterprise (SME) loan sector to address the global challenge of “information asymmetry.” Data science has unleashed immense potential in the financial field, and these innovative models provide practical solutions to technical problems in the U.S. credit industry, driving digital transformation and technological change within the sector.
Published by Joseph T.
