Selected work

Representative Papers

Artificial intelligences and human scientists exhibit complementary strengths in theory building

Sep 26, 2026

This study investigates the comparative efficacy of large language models (LLMs) and human scientists in social science theory construction, prediction, and revision. By evaluating 25 LLMs against 73 scholars on gender and racial inequality topics, the research employs large-scale blind reviews, empirical predictive accuracy metrics, and Bayesian belief updating models for quantitative assessment. Results indicate that individual LLMs outperform most humans in theory generation, excelling at handling theoretical complexity yet exhibiting ornamental redundancy. Conversely, aggregated human judgments demonstrate greater creativity, superior predictive efficiency, and more precise error-driven belief updating. These findings reveal complementary mechanisms between artificial intelligence and human cognition in scientific discovery, providing empirical foundations for AI-assisted social science research.

0 citationsRead paper

Leveraging Minute-by-Minute Soccer Match Event Data to Adjust Team's Offensive Production for Game Context

Aug 05, 2025

Football offensive statistics are confounded by match context—such as goal difference, red cards, home/away status, and pre-match win probability—leading to biased performance assessments. To address this, we develop a generalized additive model (GAM) with count-valued responses, trained on minute-level event data from 15 seasons across Europe’s top five leagues. The model systematically incorporates nonlinear effects and interactions among contextual covariates, enabling context-aware calibration of offensive metrics (e.g., shots, corners). We propose a novel “context-standardized offensive performance adjustment” framework that maps raw statistics onto a common baseline, thereby enhancing fairness and cross-match/cross-team comparability. This approach effectively disentangles outcome-driven bias and delivers an interpretable, reproducible statistical framework for objective offensive performance evaluation.

0 citationsRead paper

Partially Regularized Ordinal Regression to Adjust Teams' Scoring for Strength of Schedule and Complementary Unit Performance in American Football

Jun 03, 2025

In American football, offensive and defensive units exhibit strong complementary interactions (e.g., defensive turnovers generate offensive opportunities), yet traditional performance evaluation treats them in isolation and fails to adequately adjust for schedule strength bias. Method: We propose an elastic-penalized, bias-regularized ordinal regression model that jointly models unit complementarity and schedule strength: (i) complementarity features quantify synergistic effects (e.g., defense-to-offense transitions); (ii) penalty-free covariate embedding enables full schedule-strength calibration; and (iii) surrogate residual diagnostics are extended—first time—to non-proportional odds ordinal models, partially relaxing the proportional odds assumption. Contribution/Results: Experiments demonstrate significantly improved cross-season out-of-sample prediction accuracy; complementarity explains 12–18% additional variance in scoring outcomes; and calibrated team rankings better reflect true competitive standing.

0 citationsRead paper

Recent publications

Latest Papers

Artificial intelligences and human scientists exhibit complementary strengths in theory building

Sep 26, 2026

This study investigates the comparative efficacy of large language models (LLMs) and human scientists in social science theory construction, prediction, and revision. By evaluating 25 LLMs against 73 scholars on gender and racial inequality topics, the research employs large-scale blind reviews, empirical predictive accuracy metrics, and Bayesian belief updating models for quantitative assessment. Results indicate that individual LLMs outperform most humans in theory generation, excelling at handling theoretical complexity yet exhibiting ornamental redundancy. Conversely, aggregated human judgments demonstrate greater creativity, superior predictive efficiency, and more precise error-driven belief updating. These findings reveal complementary mechanisms between artificial intelligence and human cognition in scientific discovery, providing empirical foundations for AI-assisted social science research.

0 citationsRead paper

Leveraging Minute-by-Minute Soccer Match Event Data to Adjust Team's Offensive Production for Game Context

Aug 05, 2025

Football offensive statistics are confounded by match context—such as goal difference, red cards, home/away status, and pre-match win probability—leading to biased performance assessments. To address this, we develop a generalized additive model (GAM) with count-valued responses, trained on minute-level event data from 15 seasons across Europe’s top five leagues. The model systematically incorporates nonlinear effects and interactions among contextual covariates, enabling context-aware calibration of offensive metrics (e.g., shots, corners). We propose a novel “context-standardized offensive performance adjustment” framework that maps raw statistics onto a common baseline, thereby enhancing fairness and cross-match/cross-team comparability. This approach effectively disentangles outcome-driven bias and delivers an interpretable, reproducible statistical framework for objective offensive performance evaluation.

0 citationsRead paper

Partially Regularized Ordinal Regression to Adjust Teams' Scoring for Strength of Schedule and Complementary Unit Performance in American Football

Jun 03, 2025

In American football, offensive and defensive units exhibit strong complementary interactions (e.g., defensive turnovers generate offensive opportunities), yet traditional performance evaluation treats them in isolation and fails to adequately adjust for schedule strength bias. Method: We propose an elastic-penalized, bias-regularized ordinal regression model that jointly models unit complementarity and schedule strength: (i) complementarity features quantify synergistic effects (e.g., defense-to-offense transitions); (ii) penalty-free covariate embedding enables full schedule-strength calibration; and (iii) surrogate residual diagnostics are extended—first time—to non-proportional odds ordinal models, partially relaxing the proportional odds assumption. Contribution/Results: Experiments demonstrate significantly improved cross-season out-of-sample prediction accuracy; complementarity explains 12–18% additional variance in scoring outcomes; and calibrated team rankings better reflect true competitive standing.

0 citationsRead paper