Recent studies suggest that thin hot-mix asphalt (HMA) overlay is one of the most frequently used preservation techniques for HMA pavements. This preservation treatment is often applied to address functional problems, such as roughness, raveling and weathering, and friction loss. A novel probabilistic model is provided for predicting the international roughness index (IRI) values of asphalt pavements treated with thin HMA overlay. The model consists of two tightly coupled components. The first component is a set of artificial neural networks responsible for predicting the IRI value of the existing pavement if no treatment is applied, and the second one is a set of Bayesian regression models responsible for predicting the reduction in IRI value owing to applying the thin HMA overlay. The model considers key design and material characteristics of both the existing pavement and the thin overlay, and site factors. The developed Bayesian models have less than 5% outliers (i.e., data points falling within either 2.5% tail area of model predictions), indicating high goodness of fit for these models. It is hoped that this IRI prediction model will enable pavement engineers to estimate the performance, service life, and life-cycle costs of thin HMA overlays.