Selected work

Representative Papers

Beyond Differences: Doubly Robust Meta-Learners for Ratio-Based Treatment Effects

May 25, 2026

This study addresses the challenge of robustly estimating the conditional average treatment effect expressed as a ratio (ratio-CATE), which arises in domains such as medicine, pricing, and marketing. Existing methods often rely on restrictive log-linear assumptions, limiting their applicability. To overcome this, the authors propose Q-Learner, a novel nonparametric framework that decomposes ratio-CATE into a product of two odds ratios and reformulates estimation as two propensity score-based classification tasks. Building on this formulation, they develop S/T- and Q-type doubly robust meta-learners with distinct robustness properties. Empirical evaluation demonstrates that Q-Learner achieves state-of-the-art performance across seven randomized controlled trials with low conversion rates, while the proposed doubly robust variants significantly outperform existing baselines on four observational datasets with confounding.

0 citationsRead paper

Gini-based Model Monitoring: A General Framework with an Application to Non-life Insurance Pricing

Oct 06, 2025

Non-life insurance pricing models suffer performance degradation in dynamic environments due to concept drift, yet systematic approaches for its detection and monitoring remain underexplored. Method: This paper proposes a statistically grounded monitoring framework based on the Gini coefficient. Theoretically, we derive the asymptotic distribution of the Gini coefficient and establish a rigorous hypothesis testing procedure to distinguish spurious drift from genuine concept drift, clarifying their respective impact boundaries. Practically, we integrate the Gini index with deviation loss to formulate an operational monitoring workflow that informs model retraining decisions. Results: Evaluated on real-world datasets with controlled concept drift injection, the framework accurately pinpoints performance degradation onset points, significantly enhancing both the stability and predictive accuracy of non-life insurance pricing models.

0 citationsRead paper

Recent publications

Latest Papers

Beyond Differences: Doubly Robust Meta-Learners for Ratio-Based Treatment Effects

May 25, 2026

This study addresses the challenge of robustly estimating the conditional average treatment effect expressed as a ratio (ratio-CATE), which arises in domains such as medicine, pricing, and marketing. Existing methods often rely on restrictive log-linear assumptions, limiting their applicability. To overcome this, the authors propose Q-Learner, a novel nonparametric framework that decomposes ratio-CATE into a product of two odds ratios and reformulates estimation as two propensity score-based classification tasks. Building on this formulation, they develop S/T- and Q-type doubly robust meta-learners with distinct robustness properties. Empirical evaluation demonstrates that Q-Learner achieves state-of-the-art performance across seven randomized controlled trials with low conversion rates, while the proposed doubly robust variants significantly outperform existing baselines on four observational datasets with confounding.

0 citationsRead paper

Gini-based Model Monitoring: A General Framework with an Application to Non-life Insurance Pricing

Oct 06, 2025

Non-life insurance pricing models suffer performance degradation in dynamic environments due to concept drift, yet systematic approaches for its detection and monitoring remain underexplored. Method: This paper proposes a statistically grounded monitoring framework based on the Gini coefficient. Theoretically, we derive the asymptotic distribution of the Gini coefficient and establish a rigorous hypothesis testing procedure to distinguish spurious drift from genuine concept drift, clarifying their respective impact boundaries. Practically, we integrate the Gini index with deviation loss to formulate an operational monitoring workflow that informs model retraining decisions. Results: Evaluated on real-world datasets with controlled concept drift injection, the framework accurately pinpoints performance degradation onset points, significantly enhancing both the stability and predictive accuracy of non-life insurance pricing models.

0 citationsRead paper