Real-time nowcasting models not only provide valuable information for assessing current economic conditions but also help enhance the forecasting accuracy of medium-term economic models by compensating for the publication lags of major macroeconomic i...
Real-time nowcasting models not only provide valuable information for assessing current economic conditions but also help enhance the forecasting accuracy of medium-term economic models by compensating for the publication lags of major macroeconomic indicators. Therefore, central banks in major countries have developed and operated various real-time nowcasting models, with the model from the Federal Reserve Bank of New York (FRBNY) being a representative example. Nowcasting models require a large number of datasets with short publication frequencies and include many missing values (ragged edge), thus they are based on dynamic factor models (DFM). However, during the period of significant fluctuations in key macroeconomic indicators, caused by the COVID-19 pandemic in 2020, the performance of the GDP nowcasting model was poor. Moreover, even after the economy stabilized, the instability of the model continued, leading to the suspension of nowcasting announcements from September 2021. Subsequently, a new and improved model (FRBNY-2) was released, and nowcasting announcements were resumed in September 2023. This paper shows that the decline in the forecasting accuracy of the original model was due to excessive constraints in identifying common factors, which hindered the appropriate extraction of these factors. It is suggested that if these constraints are relaxed, the original model could maintain a forecasting accuracy comparable to that of FRBNY-2. However, in terms of model expansion, such as incorporating weekly data, the traditional dynamic factor model appears to have limitations. Therefore, this paper attempts to nowcast GDP and other macroeconomic indicators using recently developed deep learning models. Unlike traditional dynamic factor models that assume linear relationships and normal distributions, deep learning models consist of nonlinear equations and do not assume specific distributions, making them potentially more suitable for explaining rapid economic fluctuations. Among the deep learning models, the recently developed Mamba model is known to be highly efficient and well-suited for large language models (LLMs). It is also based on a state-space model structure, enabling dimensionality reduction of data and making it more comparable to dynamic factor models. Using the same economic indicators as the FRBNY model from 1985 to 2019, the model was estimated and the forecasting accuracy of quarterly GDP nowcasting was compared for the period after 2020, including the COVID-19 period. The Mamba model exhibited a smaller mean absolute error (MAE) of 2.2% compared to the 3.9% of the traditional dynamic factor model. In particular, during the second and third quarters of 2020, when GDP volatility was significant, the Mamba model had smaller forecast errors, and it also showed smaller GDP nowcasting errors for the remaining periods except for these two quarters. Notably, when the model was expanded to include weekly data such as financial market prices, which were not included in the original model, the Mamba model's mean forecast error further decreased to 1.9%, whereas the dynamic factor model's forecasting accuracy deteriorated. Additionally, the Mamba model demonstrated superior forecasting accuracy in nowcasting Korean GDP. This paper confirms that deep learning models, which have recently shown excellent performance in analyzing language, images, and sound, can also yield superior results in economic time series analysis. This suggests the possibility of developing economic forecasting models that include not only time series data but also unstructured data such as text and images. Furthermore, compared to traditional models, deep learning models can efficiently estimate large-scale datasets that include mixed-frequency macroeconomic indicators and can reflect nonlinear relationships between variables, contributing to elucidating complex relationships between economic variables in the future. Keywords : Nowcasting, Dynamic factor model, Machine learning, Deep learning, Mamba Student Number : 2010-30909