大規模言語モデルを用いたベージュブックのセンチメント評価と暗号資産市場の分析

抄録

In this study, we examine whether sentiment in the Federal Reserve's Beige Book is reflected in the Bitcoin market, focusing on the magnitude of price movements and mining difficulty changes around release dates. Prior work using financial sentiment dictionaries and FinBERT suggests that energy-related sentiment is linked to mining difficulty, but such approaches may miss context-dependent meaning. We therefore compare dictionary-based scores, FinBERT, a dovish–hawkish measure, and LLM-based sentiment scores computed at both the sentence level and the document level using gemini-2.5-flash/pro, Claude Opus 4.5/Haiku 4.5 with a generic and a cryptocurrency-oriented prompt. Our results indicate that dovishness in the economic-activity summary is weakly associated with larger Bitcoin price moves, while energy-topic dictionary sentiment is associated with larger mining difficulty changes; only the document-level, cryptocurrency-oriented LLM score for the full report is significant for price-move magnitude, and no LLM specification is significant for mining difficulty.