Forecasting job vacancies in Hong Kong using AI time-series foundation models

Timely labor market data is critical for economic policy, yet official job vacancy statistics in Hong Kong are published with a one-quarter delay. This paper investigates whether high-frequency online job postings (OJPs) can improve short-term vacancy forecasts. We benchmark Amazon’s Chronos-2 foundation model against traditional statistical, machine learning, and deep learning alternatives, demonstrating that a fine-tuned, covariate-enhanced Chronos-2 achieves superior forecasting accuracy. Our model records a Symmetric Mean Absolute Percentage Error (SMAPE) of 4.95% and a Relative Root Mean Squared Error (RRMSE) of 4.99%, significantly outperforming benchmarks like ARIMAX (SMAPE: 134.13%, RRMSE:143.06%) and Linear Regression (SMAPE: 8.02%, RRMSE: 8.67%). Crucially, feature importance analysis reveals that the predictive utility of such digital labor signals is highly sector-dependent. OJP integration is highly predictive in volatile, high-turnover sectors such as Construction and Transportation, where it serves as a “just-in-time” demand signal. Conversely, its value diminishes in legacy sectors like Manufacturing due to noise and, surprisingly, in Information and Communications, where speculative “pipeline hiring” dilutes the correlation between advertisement volume and immediate vacancies. Applying this framework to nowcast the fourth quarter of 2025, we project a 3.57% contraction in aggregate labor demand to 49,616 vacancies. This macro-level cooling masks a bifurcating market: domestic consumer-facing sectors (Retail, Accommodation) show resilient growth, highly digitized sectors plateau, and trade-exposed manual sectors contract sharply. Our findings underscore the necessity of a sector-specific approach in AI-driven economic nowcasting, providing policymakers with a high-resolution map of labor shifts months ahead of official data.