Nowcasting Swiss GDP Growth From Public Lead Texts: Simple Methods Are Sufficient
研究发现，仅从瑞士报纸的导语文本中提取每日主题情绪和衰退指标，就能有效即时预测GDP增长，且简单方法（如关键词评分）与复杂模型（如大语言模型）效果相当，但成本更低。
ABSTRACT Public lead texts from Swiss newspapers contain most of the signal needed to nowcast Swiss GDP growth in real time. I build an indicator from daily topic‐specific sentiment and recession measures extracted from three Swiss newspapers and evaluate it in pseudo‐real time. The indicator is competitive with established Swiss business‐cycle indicators and simple statistical benchmarks once enough within‐quarter news has arrived, and the gains remain after excluding COVID‐19 and the Global Financial Crisis. I then compare three design choices that arise in text‐based nowcasting systems: lead texts versus full articles, keyword‐based scoring versus large‐language‐model classification, and static versus dynamic factor aggregation. None delivers systematic forecast gains over the baseline; the LLM variant is more costly and harder to hold fixed in real time, and the full‐article indicators often perform worse. The main contribution is therefore a design result: in this public‐data setting, lead texts and simple methods already recover most of the useful nowcasting signal.