Dingyuan Liu

Large language models (LLMs) are increasingly considered as potential interfaces for communicating model‐based explanations in credit risk decision‐support settings. This study evaluates whether LLMs can serve as post hoc explainability tool for credit risk models, focusing on their ability to preserve supplied feature‐importance rankings and autonomously generate feature‐importance rankings. Using a LendingClub dataset, we compare LLM outputs with SHAP and coefficient‐based attributions. Results indicate that LLMs reliably reproduce reference rankings under controlled prompts but show limited alignment when generating explanations autonomously. These findings suggest that LLMs are best deployed as narrative interfaces rather than substitutes for formal attribution methods in credit risk governance. The paper is organized as follows. Section 1 introduces the research problem. Section 2 reviews the related literature on explainable AI and the use of LLMs in financial applications. Section 3 describes the experimental design, including data processing, baseline model training, prompt engineering strategies and evaluation metrics. Section 4 presents the empirical results for both research questions. Section 5 concludes with a discussion of implications and directions for future research.

This paper develops a Business Resilience Index (BRI) that measures county-level resilience to natural disasters at a county-quarter frequency for the United States over 2014–2024. The index integrates high-frequency labor market outcomes from the Quarterly Census of Employment and Wages with flood insurance policy information from FEMA, disaster damages from the NOAA Storm Events Database, and social and health determinants from County Health Rankings. Starting from a broad candidate set, we apply an interpretable feature-screening pipeline to retain 79 variables and then use principal component analysis to extract four orthogonal structural dimensions of resilience: market scale, socioeconomic resilience, urban density risk, and industrial economy profile. We construct a domain-weighted strategic index and benchmark it against data-driven and equal-weight alternatives, showing that county rankings are highly stable across weighting schemes. To evaluate whether the BRI aligns with recovery behavior under acute shocks, we implement a matched difference-in-differences event study around two major flood episodes—Texas in 2015Q2 and North Carolina in 2018Q3. Conditional on exposure intensity and matched comparability, higher pre-event BRI counties exhibit earlier stabilization and a stronger post-event employment path relative to lower BRI counties, with differences in magnitude and timing across cases. Overall, the BRI provides an interpretable, high-frequency baseline for identifying capacity constraints that may slow recovery and for supporting preparedness targeting and post-disaster monitoring.

Standard employment forecasting often absorbs national drift, which can obscure the reasons states persistently differ in employment intensity. This study develops and evaluates a forecasting framework for state employment relativity, defined as the log ratio of a state employment per capita measure to its contemporaneous national counterpart. We construct a state year panel for 2014 to 2024 by integrating administrative labor market aggregates, population weighted health and socioeconomic indicators, and small establishment structure measures that proxy local service capacity. To prevent temporal leakage, we use expanding window rolling origin evaluation and treat 2022 to 2024 as a forward held out period for model comparison and interpretation. Averaged over 2022 to 2024, Light Gradient-Boosting Machine (LightGBM) delivers the strongest forward performance with R 2 = 0.87, RMSE = 0.046, and MAE = 0.036 in log points, while tree bagging methods remain competitive but are less stable in the more volatile 2023 environment. SHapley Additive exPlanations (SHAP) based interpretation, evaluated on both the held out period and rolling origin stability diagnostics, identifies a compact and temporally persistent predictor set centered on education composition, health burden proxies, food environment quality, small business services density, and hires intensity. The results support a forecasting interpretation in which state deviations from national employment intensity reflect nonlinear interactions among human capital composition, health access and morbidity, and local enterprise structure. They also show why interpretability claims are stronger when assessed for temporal robustness rather than only pooled averages.

Standard employment forecasting often absorbs common national drift, which can obscure the reasons states persistently differ in employment intensity. This study develops and evaluates a forecasting framework for state employment relativity, defined as the log ratio of a state employment per capita measure to its contemporaneous national counterpart. We construct a state year panel for 2014 to 2024 by integrating administrative labor market aggregates, population weighted health and socioeconomic indicators, and small establishment structure measures that proxy local service capacity. To prevent temporal leakage, all tuning and validation procedures preserve chronological ordering, and the main model comparison is based on a fixed forward holdout that trains on 2014 to 2021 and evaluates on 2022 to 2024. Among the prespecified learners, LightGBM delivers the strongest forward performance, with holdout R^2=0.831, RMSE =0.0536, and MAE =0.0430 in log points. The main interpretability analysis therefore focuses on SHAP values from the selected LightGBM model, while a stacked ensemble is used only as a supplementary robustness diagnostic for temporal stability of predictor families. Across these analyses, the most important signals center on education composition, health burden and access, local service ecosystem capacity, and hiring intensity. A state level case study for Texas and North Carolina shows how similar relative employment positions can arise from different feature combinations and different model error profiles. The results support a forecasting interpretation in which state deviations from national employment intensity reflect nonlinear interactions among human capital composition, population health conditions, and local business service structure, and they show why clear separation between model selection, model interpretation, and robustness analysis is essential for credible empirical claims.

This paper develops a Business Resilience Index (BRI) that measures county-level resilience to natural disasters at a county quarter frequency for the United States over 2014-2024. The index integrates high-frequency labor-market outcomes from the Quarterly Census of Employment and Wages with flood insurance policy information from FEMA, disaster damages from the NOAA Storm Events Database, and social and health determinants from County Health Rankings. Starting from a broad candidate set, we apply an interpretable feature-screening pipeline to retain 79 variables and then use principal component analysis to extract four orthogonal structural dimensions of resilience: Market Scale, Resilience Capacity, Safety and Risk Profile, and Workforce Dynamics. We construct a domain-weighted Strategic Index and benchmark it against data-driven and equal-weight alternatives, showing that county rankings are highly stable across weighting schemes. To evaluate whether the BRI aligns with recovery behavior under acute shocks, we implement a matched difference-in-differences event study around two major flood episodes, Texas 2015Q2 and North Carolina 2018Q3. Conditional on exposure intensity and matched comparability, higher pre-event BRI counties exhibit earlier stabilization and a stronger post-event employment path relative to lower BRI counties, with differences in magnitude and timing across cases. Overall, the BRI provides an interpretable, high-frequency baseline for identifying capacity constraints that may slow recovery and for supporting preparedness targeting and post-disaster monitoring.

Using U.S. county health measures and labor-market fundamentals, we predict cross-state employment performance out of sample. We construct an employment relativity index, state employment per capita relative to the contemporaneous U.S. rate, to net out national shocks. We use health measures and non-employment labor-market fundamentals (hiring intensity and earnings), tune tree-based models within the training window, and achieve R2=0.846 on a 2022-2024 holdout. A leakage-safe ridge-stacked ensemble yields R2=0.863. SHAP and rolling-origin tests highlight education, labor-market dynamism, and health burden as stable predictors.

Traditional labor market forecasting relies heavily on economic time-series data, often overlooking the "health-wealth" gradient that links population health to workforce participation. This study develops a machine learning framework integrating non-traditional health and social metrics to predict state-level employment density. Methods: We constructed a multi-source longitudinal dataset (2014–2024) combining Quarterly Census of Employment and Wages (QCEW) data with County Health Rankings. Using a time-aware split to evaluate performance across the COVID-19 structural break, we compared LASSO, Random Forest, and regularized XGBoost models, employing SHAP values for interpretability.Results: The tuned, regularized XGBoost model achieved strong out-of-sample performance (Test R ² = 0.800)}. A leakage-safe stacked Ridge ensemble yielded comparable performance (Test R ² = 0.827), while preserving the interpretability of the underlying tree model used for SHAP analysis.