In this master’s thesis, we explore conformal prediction of values, focusing on assessing the reliability of property value estimates in Slovenia, predicted using nonparametric machine learning methods. Unlike traditional methods for quantifying uncertainty, conformal prediction offers robust statistical guarantees in finite samples in a distribution-free manner. Additionally, the basic concepts are broadly applicable, simple to implement and operate independently of specific machine learning models. In this way, instead of using models that provide point estimates, conformal prediction allows their calibration so that they produce intervals dependent on the model’s uncertainty and a predetermined maximum allowable miscoverage rate. In the thesis, we describe and implement various methods of conformal prediction and explore their effectiveness in determining the reliability of real estate value estimates, predicted by random forests. We find that the described methods can produce efficient prediction intervals with a guaranteed coverage rate. We also find that using prediction intervals, some versions of conformal prediction can return a confidence level for each prediction as output, which aligns with the guidelines of relevant regulatory authorities.
|
