Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Forecasting bitcoin implied volatility with market, behavioural and event driven information

Han, Beining 2026. Forecasting bitcoin implied volatility with market, behavioural and event driven information. PhD Thesis, Cardiff University.
Item availability restricted.

[thumbnail of Thesis_corrections (1).pdf]
Preview
PDF - Accepted Post-Print Version
Download (11MB) | Preview
[thumbnail of Cardiff University Electronic Publication Form] PDF (Cardiff University Electronic Publication Form) - Supplemental Material
Restricted to Repository staff only

Download (148kB)

Abstract

This thesis investigates Bitcoin implied-volatility forecasting by separating volat ility persistence, behavioural information, and event-driven investor attention. The target is the Bitcoin implied volatility index, which reflects forward-looking uncertainty from the options market. The central argument is that forecast im provement depends not simply on adding predictors but on how each information source is represented, processed, and evaluated across forecast horizons. The thesis first establishes a market-based benchmark by using lagged implied volatility and GARCH-family forecasts as inputs to an LSTM model. The GARCH family models capture volatility persistence and conditional heteroskedasticity, but their accuracy weakens when implied volatility changes sharply. Relative to the GARCH(1,1) benchmark, the GARCH–LSTM reduces test RMSPE by 55.2% to 5.70% at one day and by 42.7% to 8.22% at five days. The thesis then examines whether sentiment and Google Trends add predict ive value beyond volatility persistence and GARCH-based forecasts. The results show that behavioural variables are not automatically useful when simply ad ded to the input set. Their contribution depends on how neural architectures extract temporal information and assign importance across market, economet ric, and behavioural signals. Among the LSTM, CNN–LSTM, Transformer, and Attention–LSTM models considered, Attention–LSTM provides the most effect ive integration. The proposed dual behavioural attention mechanism improves five-day forecasting, reducing testing RMSPE by 43.32% and RMSE by 40.06% relative to the GARCH–LSTM baseline. Finally, the thesis examines whether investor attention should be represented as i Abstract raw search activity or as an event-driven intensity process. Hawkes intensity captures the clustering and decay of attention shocks, but its value is horizon dependent. It weakens one-day performance, where forecasts are mainly driven by short-term volatility persistence, but improves five-day forecasts by reducing RMSE by 27.11%, RMSPE by 30.81%, and MAE by 32.54% relative to raw Google Trends. Overall, the thesis shows that Bitcoin implied-volatility forecasting benefits from evaluating volatility persistence, behavioural information, and event-driven at tention as distinct but connected sources of predictive information

Item Type: Thesis (PhD)
Date Type: Completion
Status: Unpublished
Schools: Schools > Mathematics
Subjects: Q Science > QA Mathematics
Date of First Compliant Deposit: 26 October 2026
Last Modified: 27 Oct 2026 12:28
URI: https://orca.cardiff.ac.uk/id/eprint/190406

Actions (repository staff only)

Edit Item Edit Item

Downloads

Downloads per month over past year

View more statistics