Priset är (förmodligen) rätt : En applicering och evaluering av Laplacian-regulariserad stratifierad modellering för prisestimering av bostäder

Real estate price prediction is a persistent challenge due to the complexity of housing markets, spatial heterogeneity and rapidly changing economic conditions. With the recent surge of machine learning research, an opportunity opens up to address this problem by enabling data-driven models to learn patterns from previously sold housing data. This thesis investigates the predictive accuracy of a Laplacian Regularized Stratified Model (LRSM) for predicting residential property prices in Sweden, based on historical data sourced from real estate agency Fastighetsbyrån. The LRSM provides geographical structure through stratification by municipality and longitude-latitude bins. It further utilizes Laplacian regularization across bins to lend predictive strength to data-sparse areas. A dynamic price index, reflecting the average price per m2 in each bin per year, is introduced to provide context to the fluctuating economy and real estate market, improving its accuracy over time. The LRSM is evaluated against several machine learning models and a traditional method for property valuation; Ridge and Lasso regression, a Random Forest regressor, and valuation data provided by Fastighetsbyrån. All models are assessed using root mean square error (RMSE) and R2 score. Statistical significance is tested using Levene’s and Wilcoxon signed-rank tests. The LRSM outperforms the linear models and the traditional valuations in both accuracy and robustness over time and geography. While the Random Forest regressor slightly surpasses the LRSM in raw accuracy, the difference is small and the LRSM offers advantages in computational efficiency. The results demonstrate that LRSM is a competitive and robust model for large-scale real estate prediction. Future work may improve its accuracy by incorporating behavioral features that were unavailable in the data set.