Economic growth nowcasting through deep learning: A hybrid model of variational autoencoders and transformers

Accurate GDP quarter-on-quarter (QoQ) nowcasting is crucial for timely economic decisions and policy formulation, requiring models that effectively capture complex economic dynamics. Traditional methods, like dynamic factor models, have been widely used but face two key limitations: (i) limited representation of latent factors, which inadequately capture economic dynamics, and (ii) modest nowcasting performance due to reliance on simple regression-based estimations. This paper introduces a hybrid approach that utilizes variational autoencoders to extract latent factors more effectively, enhancing factor representation. Simultaneously, a transformer encoder improves nowcasting accuracy by capturing intricate relationships among these factors. Our model is further augmented with uncertainty projection, auxiliary input, and cross-attention modules, enhancing both accuracy and interpretability. Experimental results show that our approach significantly outperforms traditional models across key metrics. This paper highlights the advantages of integrating advanced deep learning techniques into GDP QoQ economic forecasting, with the potential to influence future research and set a new standard for accuracy in GDP nowcasting.