In South Korea, apartments dominate the residential housing market, accounting for 67.5% of total housing transactions in the fourth quarter of 2018. With this figure continuing to rise, apartments are the most significant asset for many families. Consequently, precise and timely valuations are crucial for stakeholders, including homeowners, buyers, and mortgage lenders. Traditionally, these stakeholders have relied on the qualitative judgments of certified real estate agents. Because of market opacity and low liquidity, agents often use a comparative approach, referencing the most recent transaction prices of nearby comparable apartments. However, this method is subjective, potentially biased, time-consuming, and costly. Our study seeks to offer a more objective and quantitative method for determining fair apartment prices in Korea, helping market participants make informed decisions. The prediction target is the monthly representative price of an apartment complex (the within-complex average of transaction prices), from which a complex-level price index is subsequently constructed; we distinguish this target from individual transaction prices throughout. By employing clustering methods to identify similar apartments and imputation techniques for missing values, our model demonstrates promising results, with a mean absolute percentage error as low as 5.38% in the worst-case (consecutive-mask) setting and 4.76% in the typical (random-mask) setting. Because the training (2006–2015) and test (2016–2022) periods are temporally disjoint, these figures reflect out-of-sample performance rather than in-sample fit. We further validate the resulting series against external references: it attains a 5.03% MAPE against actual transactions nationwide—outperforming the appraiser-based Kookmin Bank index (7.29%)—and, once aggregated, closely tracks the official KREB transaction-based index while becoming available earlier; complex-level Granger tests confirm that our series temporally leads the appraiser-based series about 1.5 times as often as the reverse. We also outline the missing-data assumptions under which the approach is valid.
Jo et al. (Mon,) studied this question.