<p>Publication lags in macroeconomic indicators impede timely policy responses, particularly during economic crises. This study develops a nowcasting framework to estimate South Korea’s monthly unemployment rate during the 1990s, including the 1997–1998 financial crisis, using high-frequency news sentiment data. The proposed Mixed Data Sampling with Machine Learning and Polynomial Features (MIDAS-ML-P) framework extends standard MIDAS by engineering polynomial features (squares and interactions) from daily sentiment lags. This enables machine learning algorithms to systematically capture complex nonlinear dynamics often overlooked in traditional models. Daily sentiment indices are constructed from a decade of Korean news articles (1991–2000) using multiple lexicons and benchmarked against raw article volume (<i>N articles</i>). Predictive accuracy is rigorously assessed through rolling-origin validation, bootstrap MSE distributions, and formal statistical tests. Results reveal context-dependent predictor superiority; while lexicon-based sentiment is informative, raw article volume demonstrates greater robustness in the adaptive evaluation. Further analysis shows that sentiment shocks lead unemployment by 3–6 months, providing an economic rationale for the model’s extended lag structure. This study contributes a robust methodology for real-time macroeconomic surveillance, demonstrating the value of modeling nonlinearities in high-frequency textual data to nowcast key indicators during periods of structural instability.</p>