Boosted regression trees

Purpose – The purpose of this paper is to apply boosted regression trees (BRT) to a heterogeneous data set of residential property drawn from a jurisdiction in Malaysia, with the objective to evaluate its application within the mass appraisal environment in Malaysia. Machine learning (ML) techniques have been applied to real estate mass appraisal with varying degrees of success. Design/methodology/approach – To evaluate the performance of the BRT model two multiple regression analysis (MRA) models have been specified (linear and non-linear). One of the weaknesses of traditional regression is the need to a priori specify the functional form of the model and to ensure that all non-linearities have been accounted for. For a BRT model the algorithm does not require any predetermined model or variable transformations, making the process much simpler. Findings – The results show that the BRT model outperformed the MRA-specified models in terms of the coefficient of dispersion and mean absolute percentage error. While the results are encouraging, BRT models still lack transparency and suffer from the inability to translate variable importance into quantifiable variable effects. Practical implications – This paper presents a useful alternative modelling technique, BRT, for use within the mass appraisal environment in Malaysia. Its advantages include less intensive data cleansing, no requirement to specify the predictive underlying model, ability to utilise categorical variables without the need to transform them and not as data hungry, as for example, MRA. Originality/value – This paper adds to the knowledge in this area by applying a relatively new ML model, BRT to residential property data from a jurisdiction in Malaysia. BRT has shown promise as a strong predictive model when applied in other disciplines; therefore this research empirically tests this finding within real estate valuation.