The increase in vacant houses has become a serious social issue in many developed countries, including Japan. 　Therefore, to support mid- to long-term policy planning, there is a growing need to understand future vacancy distributions. In this study, we develop a machine learning model to predict future municipal-level vacancy rates by incorporating not only demographic and building information, but also spatial indicators related to road development conditions. First, we constructed a road mesh dataset at the 500-meter grid level by aggregating physical road indicators, block rectangularity, and the proportion of buildings with front roads. We then combined these variables with data from the Population Census and the Housing and Land Survey and developed a vacancy prediction model using LightGBM. The results show that incorporating road-related indicators improves prediction accuracy. In particular, we found that municipalities with a higher density of narrow roads, more irregularly shaped blocks, and a larger proportion of buildings lacking direct road access tend to have higher vacancy rates. This study demonstrates the value of road development information, which has received limited attention in previous research, in improving vacancy prediction, suggesting that road environments can influence the spatial distribution of vacant houses. Moreover, the findings of this study contribute to the early identification of areas at risk of high vacancy and the planning of preventive measures, thereby supporting urban management through the use of smart data.