Research

I study how digital technology and public policy change tourism and hospitality markets, using causal inference, time-series models and large language models as measurement tools for text.

Publications

Design sketch: annual-report text, scored with a dictionary expanded by word embeddings, feeds a firm-by-year green index; pilots that start in three waves, with some firms never covered, are compared with a staggered difference-in-differences design. No results shown.
Design sketch.
PublishedJournal of Cleaner Production, 581, 149668 (2026)

The green push: How does China's carbon-emissions trading drive corporate green transformation?

Yang, J., Lai, L.†, Hsu, S.-C., & Wei, H.-H.

China's emissions-trading pilots, introduced in waves across regions, studied with a panel of listed firms and a new index of green transformation built partly from the text of annual reports.

Finding: Coverage by the carbon market raised firms' green transformation index by about 7% of its sample mean, partly through upgraded human capital and more R&D; the effect was larger for non-state-owned firms and for firms in central and western China.
Staggered DiD (Callaway–Sant'Anna)Goodman-Bacon decompositionDictionary expanded with word embeddingsCRITIC and entropy weightingBootstrapped mediation
Article (DOI)

Under review

Design sketch: a hexagon map of regions shaded by the wave in which the policy reached them, and an illustrative event study that stays flat at first and rises only after a lag. Not real estimates.
Design sketch, not estimates.
Major revisionAnnals of Tourism Research

Beyond tourism boom: Demand-side tourism promotion policies and housing prices

Wang, J., Hao, F., Yang, J.†, & Song, H.

A visa liberalisation policy rolled out region by region in an East Asian country, studied with transaction-level housing data.

Finding: Housing prices rose where the policy applied; the effect appeared only after a few years and worked partly through exposure to tourist flows.
Staggered difference-in-differencesEvent studySun–Abraham estimatorSpatial (Conley) standard errorsPlacebo and randomisation tests
Methods code, simulated data
Design sketch: social-media posts labelled by an LLM become daily series; in three illustrative series, sentiment dips first, avoidance talk rises next and arrivals dip last. Not real data.
Design sketch, not estimates.
Major revision, round 2Journal of Travel Research

Destination sentiment climate and tourist demand: LLM-based aspect sentiment, avoidance signals, and time-series evidence

Hao, F., Wang, K., Yang, J., & Qiu, X.

Daily indicators built from social-media posts about a destination in an Asian city, linked to official visitor arrivals.

Finding: Shifts in online sentiment came before changes in how often people said they would avoid the destination, and those came before changes in arrivals.
LLM aspect-based sentimentReliability checksUnit-root testsVAR and Granger causalityImpulse responses
Text-measurement code

Work in progress

Working paper

Disclosed digital strategy in tourism: The automation–augmentation paradox in experiential service firms

Yang, J., Hao, F.†, & Wang, K.

Overall Best Paper Award and Hospitality Management Best Paper Award, CHME 2026 (conference version)

How tourism and hospitality firms present automation and augmentation strategies in their disclosures, measured with LLM labels that must quote their evidence.

LLM classification with quoted evidenceConfidence-based human reviewPanel models for bounded outcomes
PhD thesis · in preparation

Digital transformation and performance in the hotel industry

Two studies on how self-service check-in technology changes hotel operations and the guest experience, using large-scale operational data and online reviews.

Multilevel variance decompositionQuantile regression with fixed effectsStaggered DiD for count outcomesLLM measurement of review text

Methods and tools

Each area links to open code that runs the methods on simulated or public data, so the implementation can be checked without the confidential research data.

Causal inference for policy

  • Staggered DiD: TWFE against Callaway–Sant'Anna and Sun–Abraham
  • Goodman-Bacon decomposition, event studies
  • Randomisation inference, Conley standard errors
  • Sensitivity to parallel trends (HonestDiD), power simulation
Econometrics toolkit

Text as data and LLMs

  • LLM labelling with verbatim evidence checks
  • Reliability: Krippendorff's α, κ, repeat and cross-model runs
  • Negative controls for invented labels
  • Word-embedding dictionaries, distilled classifiers
LLM text measurement

Time series

  • Unit-root tests, VAR, Granger causality
  • Impulse responses with bootstrap bands
  • Change-point detection (PELT), Bass diffusion
  • Rolling z-score early warnings
Econometrics toolkit

Multilevel and distributional models

  • Intraclass correlation, hierarchical regression
  • Shapley–Owen R² decomposition
  • Quantile regression with fixed effects (MMQR)
  • Wild cluster bootstrap, survival models
Econometrics toolkit

Machine-learning forecasting

  • LASSO, random forest, XGBoost
  • Rolling-window evaluation, Diebold–Mariano tests
  • Model confidence set, SHAP explanations
Inflation forecasting

Software

  • Python (pandas, statsmodels, scikit-learn), R, Stata, SQL
  • LLM APIs, retrieval, tool-using agents
  • Git, automated tests, GitHub Actions
GitHub

† Corresponding author. Papers under review are shown without estimates or original figures, to respect the review process; the sketches show each research design, not results.