Network analysis of co-search-based investor attention on stock prices

Abstract

This dissertation explores investor attention as a core behavioral force shaping financial market dynamics, with a focus on return predictability, volatility asymmetry, co-movement, liquidity commonality, and portfolio-level risk. Using two complementary data sources—Google Trends for search-based attention and StockTwits for investment discourse—it investigates how different forms of attention influence market behavior in both developed (S&P 1500) and emerging (BISTALL-Borsa Istanbul) markets. The study combines econometric modeling, network theory, and machine learning to offer a multidimensional framework for understanding the structural and informational complexity of attention-driven market outcomes. Traditional financial theories such as the Efficient Market Hypothesis and the Capital Asset Pricing Model often fall short in capturing behavioral frictions, cognitive biases, and information asymmetries in real-world markets. The concept of limited attention, in particular, has emerged as a key determinant of asset price behavior. This dissertation seeks to answer how investor attention affects return dynamics, volatility persistence, and systemic co-movement, and whether attention can be operationalized through data-driven network structures to enhance financial forecasting. The empirical analysis relies on two primary attention proxies. Search-based attention is measured via the Abnormal Search Volume Index (ASVI), calculated using Google Trends data. Discourse-based attention is captured through co-mention activity on StockTwits, where user-generated messages referencing multiple stocks are used to construct behavioral networks. This dual-source approach allows for a comparative analysis of general public interest versus investment-focused interaction and their respective market implications. In Borsa Istanbul, portfolios sorted by ASVI display significant differences in return and risk characteristics. Long-short strategies, where long positions are taken in high-ASVI stocks and short positions in low-ASVI ones, generate weekly abnormal returns exceeding 190 basis points under the CAPM, Fama-French, and Carhart models. These effects are especially pronounced in equal-weighted, small-cap portfolios, highlighting the sensitivity of less liquid stocks to retail attention. Sharpe ratio analysis confirms the robustness and profitability of these strategies on a risk-adjusted basis. Volatility modeling further reveals that attention plays a predictive role in conditional volatility. GARCH-family models, including asymmetric versions such as GJR-GARCH and their attention-augmented variants (GARCH-X, GJR-GARCH-X), demonstrate that ASVI significantly predicts future volatility. Importantly, negative attention shocks lead to more persistent volatility responses than positive ones, consistent with behavioral theories such as loss aversion. Forecast accuracy tests, including rolling-window evaluations and Diebold-Mariano statistics, show that models incorporating attention systematically outperform traditional approaches. Moreover, Granger causality and panel VAR analyses uncover a feedback mechanism: heightened attention increases volatility, which in turn attracts more attention, creating a self-reinforcing loop. These patterns align with theories of investor overreaction and herding behavior, especially in markets with lower informational efficiency. Periods of uncertainty or media amplification trigger attention surges among retail investors, who often overweight recent or salient information, leading to speculative trading and volatility clustering. These dynamics are more persistent in emerging markets, where arbitrage is weaker and information diffusion is slower. Attention is also found to drive risk transmission across sectors. Using a DCC-GARCH-X framework, the dissertation shows that volatility spillovers intensify when both source and target sectors experience simultaneous attention surges. This asymmetric contagion pattern highlights investor attention as a behavioral source of systemic risk, complementing existing models of volatility transmission. Such findings underscore the need to incorporate behavioral indicators into financial stability assessments and portfolio risk management. Beyond search-based proxies, the study introduces a machine learning–driven methodology to model investor attention using StockTwits data. Co-mention relationships among S&P 1500 stocks are used to construct daily behavioral networks, with each edge representing investor-perceived links. Node2Vec embeddings map these networks into latent vector spaces, and K-means clustering is applied to identify cohesive groups of stocks that share common attention profiles. This framework captures latent behavioral structures that are invisible to traditional industry classifications. Stocks within the same attention-based clusters exhibit elevated return co-movement, even after controlling for market-wide factors, firm fundamentals, and sector affiliation. Placebo clusters constructed using fundamentals fail to reproduce this effect, demonstrating that behavioral attention clustering captures unique investor-driven dynamics. Similarly, these clusters show significant liquidity commonality, as indicated by synchronized movements in Amihud illiquidity across cluster members. These co-movements persist after controlling for firm size, sector, and ownership structure, suggesting that cognitive proximity and shared sentiment influence trading activity more than previously recognized. Moreover, co-mention intensity, proxied by the edge weight in the network, amplifies both return and liquidity co-movement, indicating that the strength of attention matters as much as its direction. From a portfolio construction perspective, attention-based clusters offer tangible benefits. Portfolios formed from these clusters achieve higher Sharpe ratios and more consistent internal alignment compared to sector-based portfolios. Although such portfolios may experience higher short-term volatility, their risk-adjusted performance is superior, particularly in equal-weighted implementations. This suggests that behavioral attention networks reflect investor consensus and expectation alignment, potentially supporting more efficient price discovery. Comparative analysis between Google- and StockTwits-based attention emphasizes how platform structure shapes attention signals. Google Trends reflects broader, often event-driven public interest, not always directly tied to investment. In contrast, StockTwits provides a finance-specific environment where discussions are more technical and investment-relevant. Consequently, attention measures derived from StockTwits exhibit stronger predictive power across return, volatility, co-movement, and liquidity dimensions. The dissertation makes several methodological contributions. It integrates high-frequency attention proxies into standard asset pricing and volatility models, applies graph embedding techniques to construct latent investor networks, and uses advanced econometric methods—including Panel VAR, GMM, and various GARCH-X variants—to assess causal links between attention and market variables. This multi-method framework improves robustness, interpretability, and forecasting accuracy. On the theoretical front, the study contributes to behavioral asset pricing by showing that attention—particularly when mediated through digital platforms—is not a transient anomaly but a persistent and structural feature of market dynamics. By bridging behavioral finance with network science, it reconceptualizes financial markets as dynamic systems driven by sentiment, perception, and social connectivity. The practical implications are wide-ranging. For institutional investors, attention-based peer groupings offer an alternative lens for diversification, especially under stress conditions where conventional correlations collapse. Regulators may use attention metrics such as ASVI and co-mention intensity as early-warning tools for crowding or systemic shocks. For quantitative analysts, attention-augmented models enhance alpha prediction, volatility estimation, and real-time sentiment tracking. Despite these contributions, some limitations remain. ASVI reflects interest, not actual investment activity, and StockTwits largely captures retail investor sentiment. The clustering methods are unsupervised and may benefit from incorporating economic priors. The analysis is confined to post-2016 U.S. equities and the Turkish market, suggesting room for broader, cross-platform and cross-country extensions. In conclusion, this dissertation presents a comprehensive and methodologically rich investigation of investor attention as a behavioral engine of financial outcomes. By operationalizing attention through both search-based and discourse-based proxies, it offers new insight into how investor focus influences asset pricing, volatility, liquidity, and risk transmission. As financial markets evolve with digital information flows, understanding the structure and impact of attention will be crucial for academics, practitioners, and regulators alike.