Online customer reviews and stock price volatility: An empirical analysis of Trustpilot reviews for UK LSE consumer companies

The proliferation of user-generated content (UGC) presents a novel, high-frequency data source for financial markets. This study investigates whether the sentiment embedded in this data contains predictive information about firm-level stock market risk. The primary objective is to determine if aggregated online customer sentiment has a statistically and economically significant predictive effect on the weekly stock volatility of publicly traded UK retail companies, after controlling for market trends. A large-scale dataset was constructed, combining over 900,000 Trustpilot reviews for five major UK retailers (ASOS, Boohoo, Currys, JD Sports, Next) with financial market data from 2020 to 2024. Sentiment was quantified from online reviews using the VADER lexicon-based model. The study employed a multi-stage econometric approach. Initial attempts to fit advanced time-series models (ARIMAX, GARCH-X) consistently failed to converge due to the noisy characteristics of the differenced weekly data. Consequently, the primary analytical tool was a robust Ordinary Least Squares (OLS) regression using Newey-West (1986) Heteroskedasticity and Autocorrelation-Consistent (HAC) standard errors, which provides valid inference in the presence of serial correlation. The analysis was conducted across 4, 8, and 12-week lag structures. The OLS models revealed a nuanced relationship. For certain firms, notably ASOS and Boohoo, a statistically significant negative relationship was found between lagged customer sentiment and next-period volatility at longer (e.g., 12-week) horizons. This suggests that positive sentiment may predict lower future risk. However, this effect was not uniformly present across all firms or lag structures, indicating the relationship is context-dependent. Supplementary Granger causality tests did not provide strong evidence of a predictive relationship. This thesis contributes to the intersection of behavioral finance and marketing literature by testing the semi-strong form of the Efficient Market Hypothesis with a novel, high-frequency dataset. The findings suggest that markets are not perfectly efficient at incorporating this "soft" information, creating a potential, albeit inconsistent, signal for risk analysis. The study highlights both the potential and the challenges of using noisy UGC for quantitative financial modeling.