<p>This paper provides empirical insight into the economic value of high-frequency (HF) data applied to portfolio construction from a liquidity perspective. By categorizing the Nikkei 225 components based on trading frequency, we form global minimum variance portfolios stratified by liquidity levels. The estimation of the HF-based covariance matrix utilizes refresh-time sampling for synchronization, the realized kernel estimator and the pre-averaging estimator to eliminate noise in densely sampled data, and the classical 5-minute realized covariance estimator for sparsely sampled data. The findings show that the classical estimator and pre-averaging estimator effectively yield lower portfolio volatility in short-term smoothing, with the realized covariance estimator marginally outperforming. Risk levels decline with decreasing trading frequency, while smoothing can improve portfolio performance comprehensively. Notably, HF-based portfolio overwhelmingly outperforms low-frequency (LF)-based portfolio in terms of risks, especially in volatile periods.</p>