About Me

Hui-Ching Chuang is an Associate Professor in the Department of Statistics at National Taipei University. Her research interests include econometrics and investment, with a focus on applying machine learning and natural language processing techniques.[cv]

WORKING PAPERS

1. Estimating Persistent Explanatory Variables: A Machine Learning Approach (with Po-Hsuan Hsu, Chung-Ming Kuan, and Jui-Chung Yang)[ssrn][slide][code] The SFS Cavalcade Asia-Pacific 2024; 2024 UC Davis-FMA Napa Finance Conference; Max Planck I&E Seminar*; 16th NYCU Finance Conference (Keynote)*; 2024 FMA Asia Pacific Conference. (*Presented by Po-Hsuan Hsu)
   Abstract

   The common practice of using firm fixed effects models may weaken or even eliminate the explanatory power of important variables that are persistent. Because these variables correlate with time-invariant unobservables, their coefficients in firm fixed effects models rely only on within-firm variation and can thus be estimated imprecisely. In this paper, we propose advanced machine learning methods that select informative firm dummies for inclusion in the regression. These methods produce estimates with lower estimation errors and achieve higher power in an extensive set of simulations accommodating various persistence patterns. To facilitate use of our methods, we provide general guidance for empirical research and detailed decision steps under specific data-generating processes. When we apply these methods to two important economic relations subject to persistence, they point to a positive R&D-patent relation and a concave managerial ownership-firm value relation, consistent with economic theory.
2. What Share of Patents Is Commercialized? (with Po-Hsuan Hsu, You-Na Lee, and John P. Walsh) [data] TPRI Brownbag Seminar*; NBER Productivity Seminar*; Max Planck I&E Seminar*; TES 2023; Academia Sinica; NTU; NTPU; YZU. (*Presented by John P. Walsh)
   Abstract

   This paper develops a scalable NLP framework for measuring U.S. patent commercialization from patent text. It uses inventor-labeled commercialization outcomes from three surveys to train supervised models on transformer-based patent embeddings, benchmarking BERT for Patents, PaECTER, Pat-SPECTER, and Gemini across title-and-abstract, claims, brief-summary, and combined-text inputs. The preferred specification achieves 86.43% ROC-AUC and 91.57% PR-AUC on the holdout sample and is applied to 7.98 million USPTO patents. To improve model transparency, the paper introduces a SHAP-to-phrase interpretation workflow that maps influential embedding dimensions back to readable patent language, linking high-dimensional transformer predictions to recognizable technological concepts.
3. Design Patent Similarity with ArcFace Metric Learning (with Yueh-Heng Wu)[paper][slide]
   Abstract

   We introduce a large-scale image retrieval system for USPTO design patent drawings spanning 1976--2025. On the DeepPatent benchmark, our best-performing ConvNeXtV2 backbone achieves mAP scores of 0.819 and 0.649 on the Standard and Extended Galleries, respectively. Analyses of the embedding space further indicate that the learned representations capture patent-specific visual forms rather than merely reproducing USPC functional classifications. As a byproduct, we develop an entropy-based procedure for distinguishing photographic-style images from conventional patent drawings. The resulting framework enables end-to-end prior-art search and supports large-scale empirical analysis of design patents.
4. Estimating (High-Dimensional) Systems of Equations for Expectiles (with O-Chia Chuang, Zaichao Du) [package] International Symposia on Nonparametric Statistics 2026*, 28th Conference on the Theories and Practices of Securities and Financial Markets; NTU; TFA; TES; 6th Annual Meeting of Young Econometricians in Asia-Pacific*. (*Presented by Zaichao Du, O-Chia Chuang)
   Abstract

   This paper develops a framework for systems of conditional expectile equations, encompassing seemingly unrelated regressions, panel expectile models, and their high-dimensional counterparts. Our two-stage estimator augments each equation with estimated residuals from the other equations to account for contemporaneous dependence. For fixed-dimensional systems, we study consistency and asymptotic normality and derive a covariance formula accounting for first-stage estimation error. For sparse high-dimensional systems with dependent observations, we use LASSO in both stages and derive conditions for second-stage consistency that account for first-stage residual estimation error. An application to global systemically important banks finds asymmetric responses to past losses and substantial reductions in residual correlation after the second-stage adjustment. We also examine cross-bank networks and model-implied impulse responses before, during, and after COVID-19.
5. Classifying Hedge Fund Strategies with Large Language Models: Systematic vs. Discretionary Performance (with Chung-Ming Kuan) 2025 FMA Asia Pacific Conference, European Financial Management Association 2025 Annual Meeting, Greece; Quantitative Finance Workshop 3 (Asset Pricing & Risk Management), IMS-NUS, Singapore; 26th Conference on the Theories and Practices of Securities and Financial Markets.
   Abstract

   We fine-tune FinBERT, a finance-specific large language model, to classify hedge funds as systematic or discretionary. Removing manual subjectivity yields cleaner style labels and reveals that systematic funds, on average, generate higher factor-adjusted returns than discretionary funds. After a false-discovery-rate adjustment, 10-20 % of funds still show statistically significant positive alphas in models that include both observable and latent risk factors.
6. Hawkish by Voice: Financial Market Reactions to FOMC Communication (with Yueh-Heng Wu and Ju-Fang Yen)[ssrn][slide]
   Abstract

   This paper examines whether vocal delivery during Federal Open Market Committee (FOMC) press-conference Q&A sessions contains market-relevant information beyond the words spoken. Using WavLM, a large-scale speech processing model, we construct a vocal-based measure of monetary-policy stance, defined as hawkish versus dovish rather than generic emotion, applied to 89 press conferences from 2011–2025. For each sentence, we compare the original human audio with a text-to-speech (TTS) rendering of the same text. The Human–TTS difference provides a fixed-text measure of Q&A delivery and speaker composition. Larger gaps are associated with higher short-window cumulative abnormal returns on Treasury ETFs across maturities, conditional on textual stance, macroeconomic conditions, monetary-policy surprises, and chair regimes. When economic policy uncertainty is high, larger gaps are also associated with higher S&P 500 ETF returns and declines in VIX and VVIX. The evidence indicates that the vocal delivery and speaker composition of the Q&A contain state-dependent information not captured by text alone.
7. Superior or Spurious? Mutual Fund Alpha in Japan with Latent Factors and False Discovery Control (with Yu-Hao Chang and Chung-Ming Kuan)[paper]
   Abstract

   This paper evaluates the performance of Japanese equity mutual funds over a period characterized by substantial changes in Japan's financial and macroeconomic environment. Using 1,483 funds from 2002 to 2023, we account for latent common factors and multiple-testing bias within a two-pass asset-pricing framework. Over the full sample, evidence of superior performance is limited: only 0.20% of funds are classified as outperformers when the false discovery rate (FDR) is controlled at levels from 5% to 20%. Over shorter subperiods, however, outperformance is more common, and fund performance rankings exhibit weak persistence. Portfolios formed from Benjamini–Hochberg-selected funds subsequently earn compound annual returns of 15.08%–15.61%, compared with 10.31% for TOPIX, and generate significant factor-adjusted monthly alphas of approximately 31 basis points. These portfolios also exhibit higher Sharpe and Sortino ratios and smaller maximum drawdowns than TOPIX.

PUBLISHED PAPERS

1. Jui-Chung Yang, Hui-Ching Chuang, and Chung-Ming Kuan (2020).“Double Machine Learning with Gradient Boosting and Its Application to the Big N Audit Quality Effect.” Journal of Econometrics, 216 (1), 268-283.

2. O-Chia Chuang, Hui-Ching Chuang, Zixuan Wang, and Jin Xu (2024).“Profitability of Technical Trading Rules in the Chinese Stock Market.” Pacific-Basin Finance Journal, 84, 102278.

3. Yin-Siang Huang, Hui-Ching Chuang, Iftekhar Hasan, and Chih-Yung Lin (2024).“Search Symbols, Trading Performance, and Investor Participation.”International Review of Economics and Finance, 92, 380-393.

4. Yin-Siang Huang, Hui-Ching Chuang, Iftekhar Hasan, and Chih-Yung Lin (2021).“The Effect of Language on Investing: Evidence from Searches in Chinese versus English.”Pacific-Basin Finance Journal, 67, 101553.

5. Hui-Ching Chuang and Chung-Ming Kuan (2020).“Identifying and Assessing Superior Mutual Funds: An Application of the New Step-wise Data-Snooping-Bias-Free Test.”Review of Securities & Futures Markets, 32 (1), 1-32.

6. Hui-Ching Chuang and Chung-Ming Kuan (2010).“Testing the Performance of Taiwan Mutual Funds Based on the Tests without Data-Snooping Bias.”Review of Securities & Futures Markets, 22 (3), 181-206.

7. Hui-Ching Chuang and Jauer Chen (2023).“Exploring Industry-Distress Effects on Loan Recovery: A Double Machine Learning Approach for Quantiles.”Econometrics, 11 (6).

8. Hui-Ching Chuang and Jui-Chung Yang (2022).“Dynamic Panel Data Estimators in Leverage Adjustments Model.”Advances in Financial Planning and Forecasting, 10, 67-111.