Attention-amplified sentiment and jump risk in Bitcoin and Ethereum markets: A text-augmented Hawkes framework

We document how investor sentiment and collective attention jointly shape jump risk in Bitcoin (BTC) and Ethereum (ETH) markets using minute-level data over 2022–2023 and a Hawkes model with FinBERT Twitter sentiment and Wikipedia attention in its baseline. A moderate positive sentiment shift ( Δ S = 0.5 ), holding attention fixed, is associated with an approximately 16% higher fitted upward-jump baseline intensity at neutral attention ( A h − 1 = 0 ) for BTC, rising to 42% when collective attention is elevated ( A h − 1 = 2 ); for ETH, the corresponding figures are 19% and 49%. Specifications with textual baselines deliver double-digit percentage improvements in out-of-sample likelihood relative to otherwise identical Hawkes specifications without textual covariates. Calibration diagnostics based on the exact score decomposition associated with the consistent, optimally binned, and reproducible (CORP) approach, observed-to-expected ratios, and a dependence-robust Spiegelhalter-type test provide no evidence against calibration of the text-augmented forecasts, whereas the benchmark specifications display significant miscalibration. The findings document economically meaningful state dependence and provide a descriptive basis for high-frequency measurement of text-conditioned jump intensity in BTC and ETH markets.