Spare parts are often associated with intermittent demand patterns that render their forecasting a challenging task. Forecasting of spare parts demand has been researched through both parametric and non-parametric approaches. However, little has been contributed in this area from a Bayesian perspective, and most of such research is built around the Poisson demand distributional assumption. However, the Poisson distribution is known to have certain limitations and, further, empirical evidence on the inventory performance of Bayesian methods is lacking. In this paper, we propose a new Bayesian method based on compound Poisson distributions. The proposed method is compared to the Poisson-based Bayesian method with a Gamma prior distribution as well as to a parametric frequentist method and to a non-parametric one. A numerical investigation (on 7400 theoretically generated series) is complemented by an empirical assessment on demand data from about 3000 stock keeping units in the automotive sector to analyse the performance of the four forecasting methods. We find that both Bayesian methods outperform the other methods with a higher inventory efficiency reported for the Poisson Bayesian method with a Gamma prior. This outperformance increases for higher demand variability. From a practical perspective, the outperformance of the proposed method is associated with some added complexity. We also find that the performance of the non-parametric method improves for longer lead-times and higher demand variability when compared to the parametric one.