Impact of Google searches and social media on digital assets' volatility

Social media has become a common platform for expressing public opinions, which can influence government and business decisions. Drawing on behavioral finance theories (Kahneman and Tversky, Thaler), the study situates investor decisions within both internal and external psychological factors. Prior research shows that user sentiment affects behavior and decision-making, and that external attention can be a key driver of price fluctuations in digital asset markets. A notable co-movement occurred in 2017 when Bitcoin prices and Google search activity surged together, suggesting a relationship between online attention and market volatility. Figures on Bitcoin volatility and Google search trends for “Bitcoin” (2017–2020) illustrate similar patterns, especially around late 2017, echoing evidence from stock markets where search activity correlates with market movements. Given Bitcoin’s dominant role (around 50% of top-100 cryptocurrencies) and potential spillovers to other digital assets, and the documented influence of Twitter sentiment, this paper aims to examine the effects of Google searches—separately for web, news, and YouTube—on digital asset volatility. The study focuses on six large-cap assets in 2019 (Bitcoin, Bitcoin Cash, Ethereum, Ethereum Classic, Litecoin, Ripple; together ~71% of market capitalization). Understanding the role of user attention via Google Trends and sentiment via Twitter can inform policymakers (e.g., USD is dominant in Bitcoin transactions) and help investors devise strategies to mitigate market turbulence. The study collected real-time tweets over five months using RapidMiner to avoid missing tweets, aligning with prior work that used 2–3 months of Twitter data.

Behavioral finance research highlights that investor behavior, influenced by psychological and social factors, can drive market outcomes beyond what is predicted by rational models and the Efficient Market Hypothesis (EMH). Studies suggest that ignoring behavioral components can reduce the accuracy of performance estimates, and that investor sentiment can create price deviations. Although EMH posits prices reflect all available information, real-world evidence shows overreactions and short-term impacts from negative events. Numerous studies link Google Trends with digital asset volatility and returns (e.g., Liu and Tsyvinski; Katsiampa et al.; Chang et al.; Chuffart; Aslanidis et al.; Pinto-Gutiérrez et al.), consistent with the 2017 surge in Bitcoin price and searches. Traditional models like the Fama-French 3-factor omit user attention, potentially mismeasuring volatility dynamics. Social media and conventional media disseminate information rapidly; Twitter has been used to predict Bitcoin, stock markets, election results, and public health risks. Research using Google Trends and Wikipedia indicates informational content and bi-directional relationships with Bitcoin prices, with Google search volume reflecting population behavior and investor attention. Prior studies find that both Twitter and Google searches can affect Bitcoin volatility; however, findings differ on the relative importance of each. Most prior work focused on aggregate Google search metrics without distinguishing among web, news, and YouTube searches, leaving a gap this study addresses by assessing each search channel separately alongside Twitter sentiment.

Data: The study uses daily data for six digital assets (Bitcoin, Bitcoin Cash, Ethereum, Ethereum Classic, Litecoin, Ripple) from CoinMarketCap, and collects tweets via the Twitter Search API (RapidMiner used to capture up to 160 data points per day per asset) using asset-specific keywords. Google Trends data (web, news, YouTube searches) were obtained via RStudio. The sample spans 2019-09-01 to 2020-01-31 with Google Trends scaled 0–100. Volatility estimation: Asset returns were computed as Return_t = (P_t − P_{t−1})/P_{t−1}. Volatility was estimated via a GARCH specification with a mean equation Return_t = β0 + β1 Return_{t−1} + ε_t, ε_t ~ N(0, σ_t^2), and variance equation σ_t^2 = α + α ε_{t−1}^2 + β σ_{t−1}^2, estimated by maximum likelihood. Tweet preprocessing and sentiment: Tweets were cleaned (punctuation except #, $, @, ', ", !, ?, ., and links retained; lowercasing applied). Sentiment was computed using VADER, producing positive, neutral, and negative sentiment classifications based on compound score thresholds (≥ 0.05 positive; > −0.05 and < 0.05 neutral; ≤ −0.05 negative). Daily counts for each sentiment category were compiled. Python was used for cleaning and sentiment analysis. Standardization and stationarity: All series were standardized using Z-transformation Z_i = (X_i − μ_x)/σ_x to harmonize scale and variance. Stationarity was assessed via Augmented Dickey-Fuller (ADF) tests under intercept and trend-and-intercept specifications; series significant at level I(0) proceeded to VAR. VAR, Granger causality, and IRF: A VAR was estimated for each asset of the form Y_it = α + Σ A_ij Y_{j,t−1} + Σ B_ik X_{k,t−1} + ε_it, where Y_it is asset volatility and X_kt includes web/news/YouTube searches and positive/neutral/negative sentiment. Lag length was selected using SC, AIC, and HQ criteria; adjustments were made to address autocorrelation (e.g., lag 3 for Bitcoin, lag 2 for Ripple; lag 1 for Bitcoin Cash, Ethereum, Ethereum Classic, Litecoin). Granger causality tests were conducted on the differenced VAR system. Impulse Response Functions (IRFs) with Cholesky ordering placed Google Trend variables first, sentiment second (more exogenous), and volatility last (most endogenous). Diagnostic tests included VAR residual serial correlation LM tests and stability checks via inverse roots of the AR characteristic polynomial.

Descriptive and stationarity: Standardization yields unit standard deviations across variables. ADF tests show stationarity for all variables. For Ethereum and Ripple, all variables are stationary at the 1% level; for Bitcoin, Bitcoin Cash, Ethereum Classic, and Litecoin, volatility is stationary at 5–10% depending on specification, while other variables are stationary at 1%. VAR and Granger causality:

• Bitcoin: Web search significantly affects volatility (positive at lag 1, negative at longer lag), and news search is positive at lag 2 (5%). YouTube search shows a positive effect in VAR but not supported by Granger causality. Granger tests indicate web search and news search Granger-cause volatility.
• Bitcoin Cash: Volatility is primarily explained by its own lagged volatility; other variables are not significant.
• Ethereum: Web search positively and significantly affects volatility at lag 1 (1%); supported by Granger causality.
• Ethereum Classic: Web search positively affects volatility at lag 1 (10%). Positive sentiment negatively affects volatility at lag 1 (5%), while negative sentiment positively affects volatility at lag 1 (1%); both sentiments Granger-cause volatility.
• Litecoin: Web search and negative sentiment both positively affect volatility at lag 1 (1%); supported by Granger causality.
• Ripple: Web search positively affects volatility at lag 1 (1%). YouTube search positively affects at lag 1 (5%) and negatively at lag 2 (1%); both web and YouTube searches Granger-cause volatility. News search appears significant in VAR but not confirmed by Granger causality. Model fit and diagnostics: Reported R^2 values for volatility equations include approximately 0.84 (Bitcoin, Bitcoin Cash), 0.76 (Ethereum Classic, Litecoin), and 0.52 (Ethereum, Ripple). Durbin–Watson statistics are near 2, suggesting limited autocorrelation. VAR residual LM tests indicate no serial correlation across assets, and inverse roots lie within the unit circle, indicating VAR stability. Impulse Response Functions (IRFs):
• Bitcoin: One standard deviation shocks to web search (day 1–7) and news search (day 3–9) raise volatility; other attention/sentiment shocks are insignificant. Own-shock effects persist positively through day 7.
• Bitcoin Cash: Volatility responds slightly and positively to news search shocks (day 1–6) and strongly to own-shock (day 1–11); other shocks are insignificant.
• Ethereum: Web search shocks raise volatility (day 1–5); own-shock significant (day 1–7); other shocks insignificant.
• Ethereum Classic: Negative sentiment shocks raise volatility (day 1–10); own-shock significant; other shocks largely insignificant.
• Litecoin: Web search (day 1–9) and negative sentiment (day 1–8) shocks raise volatility; own-shock significant (day 1–8).
• Ripple: Web search shocks raise volatility (day 1–4); YouTube search shocks positive around day 3 (with later negative at lagged horizon per VAR), own-shock significant (day 1–4); negative sentiment shows a short-lived positive effect around day 3; other shocks insignificant. Overall: Google search variables (web, news, YouTube) significantly influence the volatility of Bitcoin, Ethereum, Litecoin, and Ripple, often more robustly than Twitter sentiment. Sentiment plays a notable role for Ethereum Classic (and for Litecoin via negative sentiment). Bitcoin Cash appears less sensitive to attention/sentiment variables.

The study’s central question—whether and how distinct Google search channels (web, news, YouTube) and social media sentiment affect digital asset volatility—is addressed by consistent evidence that Google search variables exert significant and dynamic effects on major cryptocurrencies. For Bitcoin and Ethereum, web search activity reliably transmits shocks to volatility, aligning with the notion that heightened investor attention can lead to increased trading and price fluctuations. Bitcoin also responds to news search, indicating the information content embedded in news-focused attention. Ripple’s sensitivity to both web and YouTube searches emphasizes that multimedia search behavior also contains predictive information about market volatility. In contrast, Ethereum Classic and Litecoin demonstrate clearer responsiveness to Twitter-derived sentiment (especially negative sentiment), supporting behavioral finance views that emotions and sentiment shocks can drive price variability. Bitcoin Cash’s volatility is dominated by its own dynamics, suggesting it may be relatively insulated from attention and sentiment in the studied period. These results underscore that attention proxies derived from Google Trends capture broad, population-level interest with material implications for volatility, often outperforming sentiment proxies from Twitter in this context. This has relevance for market participants and policymakers: monitoring search-based attention can help anticipate volatility spikes, informing risk management and policy responses to market instability.

Google search activity—disaggregated into web, news, and YouTube channels—provides significant explanatory and predictive power for the volatility of leading digital assets (notably Bitcoin, Ethereum, Litecoin, and Ripple), generally surpassing Twitter sentiment during the study period. Sentiment still matters for select assets (e.g., Ethereum Classic and Litecoin via negative sentiment). These findings aid policymakers in understanding drivers of digital asset volatility and support investors in developing strategies that incorporate attention metrics, particularly during turbulent periods. The approach demonstrates the value of separately analyzing different search channels rather than relying solely on aggregate Google Trends. Future research could extend coverage to a broader set of digital assets, longer horizons, alternative sentiment sources, higher-frequency data, and models capturing nonlinearities or regime changes to deepen understanding of attention–volatility dynamics.

The analysis is limited to six major digital assets over a relatively short window (September 2019 to January 2020). Twitter data, while comprehensive within the collection period, may not capture all sentiment dimensions beyond the platform. Findings may not generalize to smaller-cap assets, different market regimes, or longer time spans.