Using machine learning to aggregate apartment prices: comparing the performance of different Luxembourg indices

This paper presents three different methods to estimate an apartment price index for Luxembourg and evaluates their performance by comparing volatility, proneness to revisions, coherence and out-of-sample fit. In addition to the standard hedonic  and repeat sales methods, we apply a machine learning algorithm (the interpretable random forest approach) to produce a new index for Luxembourg. The three methods indicate similar trends in residential property prices. The new random forest index closely tracks the two more traditional indices, providing evidence supporting the viability of this new approach. In comparing the three methods, the random forest index is more stable and therefore provides information that is easier to interpret. However, all three methods are subject to revisions when new observations are released and these tend to be larger for the random forest than for traditional indices.
JEL Codes: C40, C53, R31
Keywords: Residential property price index, hedonic model, repeat sales model, machine learning, random forest algorithm