Returns from rare coins: a machine learning approach

Abstract Despite the existence of several studies on collectibles, the rare coin market is still underexplored. This paper examines this market with a sample of 5553 Spanish columnarios (1732–1772) auctioned from 1992 to 2021, investigating the key factors influencing auction prices using a dataset with an extensive number of covariates. Traditional hedonic models face challenges with large datasets, including multicollinearity, overfitting, and parameter complexity, which compromise clear and reliable interpretation. To address these limitations, this study employs the cross-fit partialing-out LASSO regression to select key explanatory variables, resulting in unbiased estimates and insights for investors and researchers. An interrupted time series analysis is subsequently conducted to compare indices derived from the traditional and LASSO hedonic methods. Findings confirm that LASSO approach outperforms the traditional hedonic regression method in terms of estimation accuracy.