Induced uncertainties during the filament winding process may cause a significant stochastic variation in the mechanical behaviour of composite shells. This paper aims to develop a novel and deep uncertainty quantification (UQ), sensitivity and reliability analyses of filament wound shells considering manufacturing uncertainties. Firstly, a progressive damage analysis is performed to estimate their deterministic burst pressure. Then, a signal-to-noise (SNR) approach is employed using the Taguchi method for sensitivity analysis and screening uncertainties arising from manufacturing. Initial results reveal that the shells are more sensitive to thickness uncertainties for thinner structures. Then, probabilistic and reliability analyses are carried out using the Boosted Decision Trees Regression (BDTR) approach from machine learning algorithms. Despite the complexity and non-linear relationships in the problem, the developed BDTR-based metamodel shows powerful predictive performance. A comparative study shows that ply thickness uncertainty leads to a significant underestimation of failure probability. For expensive and time-consuming models in that only a few runs can be affordable, a modified approximation method for reliability analysis is proposed. Results indicate a high capability at estimating failure probability with high accuracy.