From guesswork to guidance: meta-learning for algorithm selection in real estate market forecasting

What the paper says

Purpose The increasing complexities of real estate market forecasting, in combination with the accelerated evolution of Machine Learning (ML) algorithms, necessitates the optimisation of algorithm selection to reduce computational demands and enhance model accuracy. While numerous studies have examined the performance of individual algorithms, a significant research gap remains concerning the impact of dataset characteristics on algorithmic performance within this specific domain. Design/methodology/approach The present study aims to address this research gap by undertaking a model-based meta-learning approach, in which a Random Forest classifier is trained on prior dataset characteristics and associated ML performances. In a final step, these results are illustrated using empirical data. Findings The findings suggest that, in this proof-of-concept setting, mapping dataset characteristics to an algorithm recommendation is feasible and yields encouraging predictive performance. The evaluation achieved an average AUC of 0.85 and an accuracy of 0.88, exceeding the No Information Rate of 0.38. However, results should be interpreted as exploratory given the limited meta-sample size. Originality/value This study is the first to apply meta-learning within a domain where datasets are heterogeneous and not publicly shared. It was shown that there is a systematic relationship between data structure and model performances which confirms the “no free lunch” theorem. This study may be considered as the initial attempt that can be developed further through subsequent studies.