2 outcomes. The authors first collected Hong Kong’s annual Policy Addresses from 1997 to 2023, excluding 2002 when no Policy Address was released, and used them as the main source for measuring land policy intensity. Statistical data on land transactions, land-related fiscal revenue, economic development, population, land auction area, and fiscal capacity were obtained from official sources, including the Hong Kong Land Registry, government financial reports, the Hong Kong Monetary Authority, and the Hong Kong Annual Digest of Statistics. To construct the land policy intensity index, the study used a large language model, ERNIE, to identify policy sentences related to land resource allocation, land management, urban planning, administrative authority, and local governance. The model was trained and tested using manually labelled policy sentences, and its classification accuracy outperformed several traditional machine-learning models. The resulting index was then incorporated into regression analysis, error correction models, and Granger causality tests to examine both the short-term and long-term effects of land policy intensity on land revenue and fiscal structure. Findings and Analysis Rising but Fluctuating Land Policy Intensity Hong Kong’s land policy intensity shows a pattern of phased fluctuation and gradual long-term growth from 1997 to 2023. The policy intensity score was relatively low in 1999 and 2006, at 1.8 and 1.3 points respectively, but became more prominent in 2008, 2014, and 2019, reaching 4.6, 5.0, and 5.7 points. This suggests that Hong Kong’s land governance has become more active and policy-driven over time, especially in areas such as land resource allocation, urban planning approval, market regulation, and development control. Reliable Measurement through Large Language Models Before examining the fiscal effects of land policy intensity, the study tests the reliability of its text-based measurement. The ERNIE model achieved a classification accuracy of 85.11%, outperforming traditional machine-learning models, including the Voting Classifier at 76.60%, Neural Network at 75.00%, SVM at 73.40%, Gaussian Naive Bayes at 69.68%, and KNN at 68.09%. This indicates that the large language model can more effectively identify land-policy-related meanings embedded in official Policy Addresses. Land Policy Intensity Reduces Absolute Land Revenue The regression results show that stronger land policy intensity is significantly associated with a decline in the absolute amount of land revenue. In the baseline error correction model, a 1% increase in policy intensity is associated with a 0.66% decrease in land fiscal revenue in the same period. In practical terms, this means that a 10% increase in policy intensity would lead to an approximate 6.6% decline in land fiscal revenue. When lagged dynamics are included, the negative effect remains, though it becomes smaller: a 10% increase in policy intensity corresponds to an approximate 4.5% decrease in land fiscal revenue. This finding suggests that intensified land policies do not necessarily aim at maximizing land income. Instead, they may reflect the government’s growing concern with housing affordability, market stabilization, public housing supply, and sustainable urban development. Stronger intervention may reduce land transactions or limit revenue-oriented land disposal, thereby lowering land revenue in absolute terms. Figure 1: Annual trend of land policy intensity level Hong Kong land development and policy intensity: strategic embedding in a period of transition Public Policy BULLETIN
RkJQdWJsaXNoZXIy NDk5Njg=