How Much Do Sentiment and On-Chain Data Add? A HAR-Anchored Multimodal Mamba for Bitcoin Volatility Forecasting
Academic Journal of Finance and Accounting · 03.11.2026 · 1 (5) · 44-59
DOI: 10.54691/fr3m0946 ↗Kurz gefasst
Fachartikel in der Zeitschrift Academic Journal of Finance and Accounting, online erschienen am 03.11.2026. Verfasst von Lily Jenkins und 1 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Sentiment indices and blockchain activity are widely proposed as predictors of cryptocurrency risk, yet their value for volatility forecasting beyond strong econometric benchmarks is rarely tested under strict out-of-sample conditions. We present SOC-Mamba, a multimodal selective state space framework that learns a residual on top of a heterogeneous autoregressive (HAR) forecast and issues heteroscedastic 1-, 7- and 30-day forecasts from market, Fear & Greed sentiment and Coin Metrics on-chain streams. Two fusion strategies are compared: early fusion (EF) and a cross-modal selective scan (XS), which extends the volatility-aware selection of CryptoMamba-SSM so that a reliability-gated auxiliary context also conditions the step size and projections of the Mamba encoder. On a public-data benchmark of daily Bitcoin realized range-based volatility from 4-hour bars (2020–2026) with explicit publication lags, 13 forecasters are compared over five expanding-window folds, with capacity and training length chosen on validation data only. Without the anchor, deep sequence models have 9.3–53.5% higher MSE than HAR, which identifies the anchor as the key ingredient. Both anchored variants join HAR in the 90% model confidence set at every horizon; SOC-Mamba-EF attains the lowest 1-day MSE and reduces MSE by 10.7–23.7% during the FTX collapse, and SOC-Mamba-XS improves significantly on HAR in 1-day QLIKE. Econometric anchoring makes deep multimodal forecasters competitive with strong econometric benchmarks, and the auxiliary sentiment and on-chain data deliver their largest gains during stress episodes.
Zitieren
APA
Jenkins, L., & Rao, Y. (2026). How Much Do Sentiment and On-Chain Data Add? A HAR-Anchored Multimodal Mamba for Bitcoin Volatility Forecasting. Academic Journal of Finance and Accounting, 1(5), 44-59. https://doi.org/10.54691/fr3m0946
Vancouver
Jenkins L, Rao Y. How Much Do Sentiment and On-Chain Data Add? A HAR-Anchored Multimodal Mamba for Bitcoin Volatility Forecasting. Academic Journal of Finance and Accounting. 2026;1(5):44-59. doi:10.54691/fr3m0946
BibTeX
@article{Jenkins2026,
author = {Jenkins, Lily and Rao, Yuna},
title = {How Much Do Sentiment and On-Chain Data Add? A HAR-Anchored Multimodal Mamba for Bitcoin Volatility Forecasting},
journal = {Academic Journal of Finance and Accounting},
year = {2026},
volume = {1},
number = {5},
pages = {44--59},
doi = {10.54691/fr3m0946}
}
EUVIMED