<p>This research proposes a new approach in nowcasting, which uses high-frequency data to make early predictions of low-frequency economic indicators. Many existing methods rely on univariate models specialized for specific indicators and, as a result, face high development and maintenance costs. Furthermore, these isolated univariate models bring inconsistent results when implementing multivariate predictions. In addition, with the increasing popularity of alternative data, there is a growing need to process high-dimensional data with irregular observation frequencies and release timings. We propose a new latent variable model called the Dynamical Incomplete Variational Autoencoder (DI-VAE) to address these challenges. This model introduces a flexible structure that can deal with missing data and differences in observation frequencies. In experiments conducted with a mixed-frequency, multivariate dataset of the Japanese economy, DI-VAE demonstrated better prediction performance than state-of-the-art benchmark models. Additionally, we observed that the prediction error decreased ideally as the observed data increased. DI-VAE shows potential as a nowcasting method that can adapt to diverse and irregular data structures.</p>