The house price index is essential for assessing the housing market conditions and predicting future changes in the market. However, the two main methods for calculating the housing price index, the appraisal-based index and transaction-based index, have significant drawbacks. The appraisal-based index method has issues such as smoothing and lag bias owing to the involvement of human appraisers, and it also requires significant manpower for calculations. The transaction-based index method requires numerous transaction samples to achieve statistical significance, making it challenging to be created for small regions or when transactions are scarce. To address these issues, we propose an alternative approach that employs a machine learning model to estimate time-series transaction prices for individual apartments and builds a Laspeyres index with the estimated prices. We demonstrated the model’s ability to capture serial correlation of house prices, enabling accurate estimations even with unobserved potential prices. Comparing our prediction-based index to existing Korean house price indices, we observed local smoothing but overall alignment with the global trend of the transaction-based index, facilitating smooth index calculation for small areas. This study offers a novel method for house price index calculation that mitigates limitations of traditional approaches by using machine learning for more precise house price estimation, even with limited housing transaction data.