A1139
Title: Forecasting Japanese recessions using machine learning and text data
Authors: Yusuke Oh - Bank of Japan (Japan)
Mototsugu Shintani - University of Tokyo (Japan) [presenting]
Abstract: The forecasting of Japanese recessions is investigated, and a machine learning-based framework is proposed to address the relative scarcity of such research for Japan. Building on well-established predictors such as government bond term spreads and financial cycle measures, the focus is particularly on evaluating the additional predictive power of text-based metrics constructed from newspaper content, a relatively unexplored dimension in Japanese recession forecasting. A pseudo-real-time, out-of-sample forecasting exercise is conducted, examining comprehensive combinations of models and predictors to identify the optimal configuration. Model selection using the model confidence set (MCS) procedure demonstrates that machine learning models substantially outperform traditional logistic regression benchmarks and that text-based indicators significantly enhance short-term forecasting performance.
