APJITM UKM

Article Info

Explaining Bitcoin Volatility with GARCH Models and SHAP: A Human-Centered XAI Approach to Financial Analytics

Suleiman Dahir Mohamed, Mohd Tahir Ismail, Lubna Hamzalouh
dx.doi.org/10.17576/apjitm-2026-1501-13

Page 224-237

Abstract

This paper investigates Bitcoin (BTC) volatility dynamics using a GARCH (1,1) model supplemented with major financial regressors to better understand the causes that drive market variations. Initially, a baseline GARCH (1,1) model with a normal distribution was used to estimate volatility in weekly BTC log returns from January 2016 to May 2025. The model was then improved by adding three regressors, and the fitted sigma values were used as a dependent variable. We use external factors like Google Trends data on the term "Bitcoin crash" as a proxy for public fear and sentiment, the CBOE VIX (market-wide risk sentiment), and Bitcoin trade volume (market activity intensity) to capture behavioral, macro-financial, and transactional influences on extracted volatility. The enhanced GARCH model performed better, with a log-likelihood of 499.94 and significantly improved information criteria values (AIC = -2.02, BIC = -1.96). The SHAP (SHapley Additive exPlanations) analysis was used to interpret the regressors' contributions to BTC volatility. The results showed that Bitcoin crashes (public fear) had the greatest impact on volatility, followed by BTC trading volume and the VIX. BTC crash incidents increased and decreased volatility based on market conditions, although trade volume constantly contributed positively. VIX had the least effect but offered a stabilizing influence. Diagnostic tests confirmed the model's adequacy and robustness, revealing no autocorrelation or heteroscedasticity in the residuals. This study shows how statistical modeling and explainable AI can be combined to generate interpretable insights about Bitcoin market dynamics, with an emphasis on human-centered, data-driven analysis.