In this study, we explore the use of generative models for time series prediction and construction of prediction intervals, addressing the challenge of quantifying uncertainty in deep learning models. Specifically, we employ a Variational Autoencoder (VAE), a form of a Bayesian neural network also a part of generative AI models, to model and generate latent factors for the exchange rates of ten currencies. These latent factors enable the approximate reconstruction of the exchange rate series through the decoder part of VAE. By generating a thousand sets of latent factors and reconstructing exchange rates, we create prediction intervals through a Multi-Layer Perceptron (MLP) applied to the reconstructed series. This approach provides valuable insights into evaluation of the uncertainty associated with time series predictions using neural networks.