News vs. Social Media: Sentiment Impact on Stock Performance of Big Tech Companies

Abstract

With the growing prominence of large technology firms and the shift in news dissemination driven by social media, scholars have increasingly examined how public discourse about these companies shapes financial markets. Focusing on Apple, Amazon, and Microsoft during the transitional period of January 2015–January 2020, this study evaluates attention and sentiment across traditional news media, social media, and web search in relation to stock market outcomes. We use relatively fine-grained weekly data to link media attention and sentiment to stock returns, volatility, and trading volume. To compare media sentiment across sources, we apply FinBERT-based sentiment analysis, drawing on advances in domain-specific language modeling tailored to financial texts. Results show that social media sentiment (Twitter), exerts a consistently positive and significant influence, while the effects of traditional news media (New York Times) and web search activity (Google Trends) are more irregular. The impact also varies across firms: Twitter sentiment is strongly related to trading volume and volatility for Amazon and Microsoft, but appears less influential for Apple, whose large trading base may dilute the effect. These findings offer a historical baseline for media–finance interactions and highlight how text analysis illuminates the pre-COVID era of big technology firms.

1. Introduction

In recent years, the accelerating pace of digital communication has reshaped how financial markets absorb and process information. Among the most dramatic advances is the growth of social media websites, particularly Twitter, which serve as discussion forums for exchanging opinions, gossip, and reactions to company events in real time.

Alongside these newer sources, traditional news media continue to shape investor decision-making through more refined stories and analysis of companies and market conditions. The parallel roles of these two streams of information—fast, perhaps emotive social-media postings and more refined reporting in the press—raise the questions about investor sentiment, expressed across several platforms, which translates into measurable shifts in market outcomes. In addition, studies on investor attention also consider Google Trends search intensity (Da et al., 2011) as a proxy, though its effectiveness relative to media-based indicators remains uncertain.

Research in behavioral economics and finance has long argued that investor mood and perceptions can influence returns, trading activity, and volatility alongside fundamental factors (De Long et al., 1990; Tetlock, 2007). Early studies showed that negative news coverage could move markets (Tetlock, 2007), and subsequent work confirmed that media reports can shape trading activity (Engelberg & Parsons, 2011). Later research demonstrated that online forums and tweets also carried predictive power (Bollen et al., 2011; Sprenger et al., 2014). Yet much of this evidence looks either at broad indices or at single channels of information.

More recently, scholars have begun to focus on the differential impacts of traditional news media and social media (e.g., Jiao et al., 2020). The changing media landscape primarily drives this trend—the dramatic increase in social media use over the past two decades, coupled with the enhanced availability of social media data. While various studies have investigated the topic, providing valuable insights, they primarily focus on coverage (media volume) and aggregate firm-level outcomes, leaving open questions about how media sentiment and coverage effects play out across firms.

Methodologically, most prior studies have relied on relatively basic sentiment proxies that may not fully capture the contextual and domain-specific nature of financial language. To address this limitation, we employ FinBERT, a transformer-based large language model (LLM) trained on financial texts (Huang et al., 2023), to generate sentiment measures from both news and Twitter data. In their recent study, Ruan and Jiang (2025). demonstrate that FinBERT-based multimodal forecasting outperforms technical indicators, while Jiang and Zeng (2023) show that FinBERT consistently surpasses generic lexicons for market prediction tasks. This approach enables us to extract more accurate and context-aware signals of evaluative tone, thereby enhancing the precision of sentiment analysis and facilitating a richer understanding of how media sentiment relates to financial market dynamics.

This paper investigates that issue by focusing on three tech giants: Apple, Amazon, and Microsoft. These companies dominate the technology sector and frequently attract substantial coverage in both mainstream media and online conversations. The study analyzes weekly data over a five-year period, from the first week of 2015 to the first week of 2020. This five-year truncation serves several purposes. First, it excludes the onset of the COVID-19 pandemic, whose unprecedented economic disruption and transformation of media narratives would confound the identification of more structural relationships (Baker et al., 2020). Second, it ensures consistency across data sources while capturing the phase in which social media emerged as a significant channel of investment information, coinciding with major market episodes that tested firm resilience (Chen et al., 2014). Finally, the pre-2020 window is particularly relevant because Apple, Microsoft, and Amazon all underwent substantial increases in market capitalization and global reach during these years, making them salient cases for examining how media sentiment and attention interact with stock performance (Fama & French, 2021). This period is therefore well-suited for examining how differences in sentiment across media outlets translate into stock performance of the three big-tech.

By drawing on media coverage and sentiment from both traditional news and social sources, together with the web search-based proxy for attention this analysis aims to uncover whether Twitter’s immediacy creates market pressure that differs from the stabilizing role often attributed to traditional news. The study also seeks to clarify the extent to which investor reactions to media signals are reflected in stock returns, volatility, and trading volume. In doing so, the paper connects theories of information diffusion with practical questions about market efficiency and investor behavior.

2. Literature Review

There has been growing interest in the intersection of social media coverage, sentiment and market reactions in financial markets over the past decade, especially its impact on stock returns, trading volume, and volatility of large technology firms. The evidence from literature in this area shows a complex connection between these variables, as observed by various articles examining the explanatory potential of sentiments articulated via tools such as Twitter and conventional news outlets.

Early work established that investor communications and news tone embed information that co-moves with markets. Using newspaper column tone, Tetlock (2007) showed that high “pessimism” in daily media predicts lower next-day market returns and subsequent mean reversion, providing evidence that text-based sentiment captures risk-premia-relevant beliefs rather than mere noise. A little earlier than Tetlock (2007) work, Antweiler and Frank (2004) demonstrated that message-board “bullishness” and posting volume are associated with next-day returns and volatility for individual stocks, foreshadowing the role that high-frequency, investor-generated text would later play. These two studies, one news-based and one investor-generated, set the modern blueprint for measuring sentiment and testing short-horizon predictability.

The Twitter era accelerated this agenda. Bollen et al. (2011) used large-scale Twitter mood indices and Granger-causality tools to show that specific emotions, especially “calm”, predict the direction of market moves over short windows, suggesting that broad public mood captured in social platforms can anticipate subsequent price pressure. Subsequent firm-level analyses (e.g., Ranco et al., 2015; Sprenger et al., 2014) connected hashtag-tagged tweets to individual stocks, documenting that tweet sentiment and message volume correlate with abnormal returns, trading volume, and realized volatility, conveying that social signals may proxy for shifts in attention, disagreement, and order flow.

Building upon the early work of Rao and Srivastava (2012), the authors explore the correlation between Twitter sentiments and financial market products and focus on the short-term impacts of social media chatter on stock prices and indices. Through their analysis of more than four million tweets, they identify the strong impact of social media hype on microeconomic market movements and propose that sentiment analysis can play a crucial role in formulating short-term hedging strategies. One of the best aspects of this article is that it places emphasis on the timeliness of social media data to identify market mood. The authors demonstrate that Twitter-based market sentiment affects stock prices and indices before traditional news outlets do because online networks transmit information at high speed. The study demonstrates that real-time market expressions that express optimism or pessimism link directly to market price movements. Twitter sentiment serves as a predictive tool for market movements because it directly influences short-term price fluctuations while also representing investor emotions.

Further elaborating on this issue, Ranco et al. (2015) discusses the power of prediction of social media sentiment and proposes that investors overreact or underreact to news events. The study demonstrates the importance of real-time information to market mood and stock performance, with an indication of positive correlation between media pessimism and trading volumes. Ranco et al. (2015) also shows that sentiment analysis programs can handle large data sets and put social media in an important role for financial predictions. Critically, while the article provides compelling evidence for the predictive power of sentiment on social media, it also indicates that the effect of sentiment could vary according to the intensity and nature of the news, which is in line with investor overreaction and underreaction theories. But automated sentiment analysis raises questions about the accuracy and contextualization of sentiments, which are layered and complex, especially in the fast-paced environment of social media. However, the application of social media data to financial analysis is a rich avenue to feeling out market psychology and predicting market movements, especially for technology giants that are widely discussed on the internet.

Wlodarczak (2017) is more precise in her approach by connecting the frequency of positive and negative tweets about Apple to its share price. This analysis shows that the increase in positive tweets can predict increases in share price, thus the economic significance of social media data. The thesis contributes to the growing emphasis on the predictive capability of social media across various industries, further solidifying the application of Twitter in financial analysis.

Chatterjee and French (2022) develops on these findings by specifically exploring the effect of Twitter and news media sentiments on stock returns, trading volume, and volatility for Apple, Amazon, and Microsoft. This study points to the relationship between tweet volumes and trading volume, indicating that positive investor sentiment can fuel market activity. By categorizing sentiments into positive, neutral, or negative, Chatterjee’s study paints a more detailed picture of how individual stocks react to social media sentiments.

Methodologically, the field has progressed from dictionary counts to supervised learning. Ke et al. (2020) introduced a text-mining approach trained explicitly to predict returns from news, outperforming off-the-shelf sentiment scores. This shift matters practically for volatility and volume: when text features are tailored to the prediction target, they can pick up risk-relevant news intensity, not just average tone. Parallel work in forecasting has shown that mixed measures of sentiment and attention improve daily volatility models, reinforcing the idea that text affects not only mean returns but also second moments of returns through uncertainty and trading activity channels (Audrino & Hu, 2016).

Jiao et al. (2020) provide one of the most comprehensive comparisons of how social media and traditional news shape market dynamics. Using a large panel of U.S. stocks and media data from the Thomson Reuters MarketPsych Indices, they show that the two information channels have systematically opposite effects on subsequent trading activity and volatility. Specifically, higher coverage in social media predicts increases in idiosyncratic volatility and turnover, consistent with the idea that repetition in online platforms fuels disagreement and amplifies investor sentiment. In contrast, greater news media coverage tends to reduce volatility and trading activity, reflecting its role in supplying more structured and credible information that narrows informational asymmetries. Their findings are also supported by a panel vector autoregression, which reveals that news media tends to lead social media in content generation, reinforcing the interpretation of social media as an “echo chamber” of existing news. This study highlights the need to distinguish between the informational quality of media sources when assessing their impact on financial markets.

Since 2017, empirical studies have sharpened three patterns. First, short-horizon predictability: high-frequency or daily Twitter-based sentiment often anticipates intraday or next-day returns, with effects that attenuate quickly. Sun et al. (2016) (a high-frequency precursor) show that half-hour sentiment predicts same-day returns; newer sector- and index-level work finds similar one- to two-day windows. Zeitun et al. (2023) report that Twitter sentiment significantly influences sectoral returns across U.S. industries, suggesting that social signals move beyond single names to broader portfolios.

Divernois and Filipović (2024) document that StockTwits classified sentiment predicts returns, both unconditionally and around events, highlighting that trader-focused platforms may carry especially timely signals. A 2023 Federal Reserve study builds a real-time Twitter-based financial sentiment index that predicts equity returns and reacts to monetary policy surprises, underlining macro-relevance.

Second, volatility and volume amplification: Twitter/news sentiment shocks are often accompanied by higher turnover and realized volatility, consistent with disagreement and attention mechanisms. Ranco et al. (2015) find joint predictability for abnormal returns, trading volume, and volatility using firm-tagged tweets; Audrino and Hu (2016) show that sentiment and attention proxies improve volatility forecasts. During crisis regimes, machine-learning sentiment extracted from news headlines becomes an even stronger predictor of both returns and volatility (Costola et al., 2023), plausibly because negative news intensifies belief dispersion and risk assessments. Cross-country evidence during COVID-19 similarly links positive/negative investor sentiment to both market returns and volatility (Cevik et al., 2022).

Third, news vs. social media: While both sources matter, their roles differ. News is curated, often price-relevant, and its impact can persist when it conveys fundamentals, whereas social media can be timelier but noisier and more sentiment-driven. Work using millions of news stories shows that returns respond within days, with asymmetric and sometimes more persistent effects for negative news. Machine-learning approaches tailored to return prediction (e.g., Ke et al., 2020) extract stronger signals from news than generic sentiment dictionaries. At the same time, trader-centric platforms (Twitter/StockTwits) capture attention and disagreement shocks that translate rapidly into order flow, which helps explain their tight links to volume and intraday volatility. Recent comparative evidence often concludes that news-based sentiment is “more rational” while social sentiment can be more “behavioral,” though both are predictive under the right horizons and during turbulent periods.

Recent work continues to refine tools. FinBERT (Araci, 2019) and later finance-specific Transformer variants have improved sentence-level polarity in headlines and tweets by learning financial context (e.g., “beat/miss,” “guidance,” “downgrade”), which reduces misclassification relative to general-purpose lexicons. Studies applying FinBERT or similar models typically report stronger return/volatility links than dictionary methods, particularly around events (earnings, macro releases) where domain language is specialized.

More recent studies have continued to expand the methodological frontier of market sentiment modeling across both social media and news sources. Verma and Verma (2025) conclude that economic-news sentiment tends to be more “rational,” showing stronger pricing effects on returns, whereas social-media sentiment more prominently amplifies volatility and trading activity. Similarly, Wu and Gu (2023) proposed a lightweight attention-based deep-learning framework to extract market sentiment from Twitter posts, achieving strong predictive accuracy despite limited interpretability. Building on this trend, di Tollo et al. (2023) combined BERT-based natural-language processing with stochastic neural networks to forecast the directional movement of cryptocurrencies and stocks, demonstrating that hybrid models can effectively capture complex sentiment–market linkages.

Collectively, these approaches illustrate the growing sophistication and predictive power of text-based sentiment modeling in finance. At the same time, they highlight the importance of retaining interpretability when connecting sentiment measures to financial variables such as returns, volatility, and trading volume. Balancing predictive performance with economic transparency remains a key direction shaping how sentiment analysis contributes to understanding market dynamics.

In a related work, Nyakurukwa and Seetharam (2024) examined the relative effect of social media and news media on stock market dynamics. Based on their findings, although stock prices have the potential to predict financial news, people’s attitudes on social media, particularly from influencers, may not be as predictive in nature. It further reveals that positive social media updates during large-scale events, such as the COVID-19 pandemic, were linked with larger subsequent stock returns, but the impact of official news was larger. The most important insight of the study is the differential effect of news and social media opinions on stock market behavior. The authors note that past news sentiment has a greater impact on market volatility, which means that investors perceive news outlets to be more credible or influential market-moving information sources. Conversely, social media, especially Twitter posts of prominent individuals such as CEOs and politicians, has a lower correlation with subsequent stock prices but has a significant impact on market return correlation, indicating its role in defining short-term market sentiment and volatility.

In brief, the literature makes a compelling case for sentiment analysis in financial markets, and particularly for big technology firms. Whereas social media sentiment, and Twitter sentiment in specific, demonstrates predictive potential for trading volume and stock volatility, traditional news media continues to have greater influence on stock price movement. This suggests that investors and researchers should consider both these sources of sentiment in their analysis to fully understand market activity.

3. Stock Performance of the Three Big-Tech (Jan 2015–Jan 2020)

This section examines the pre-pandemic expansion period for Apple (AAPL), Microsoft (MSFT), and Amazon (AMZN), from January 2015 through January 2020. We analyze price behavior, trading activity, and realized volatility. Price and volume data are drawn from Nasdaq Historical Data, while volatility is calculated from daily returns. The period is intentionally capped at January 2020 to focus on a relatively stable macro regime preceding the COVID-19 shock (Fama & French, 2021; Nasdaq, 2025).

As illustrated in Figure 1, all three firms display persistent price appreciation across January 2015–January 2020, reflecting strong cash-flow growth and the ongoing digitization of consumption and enterprise workflows. Apple’s ascent is underpinned by the broadening of its services ecosystem—stabilizing earnings through recurring revenue—while hardware cycles add periodic lift. Microsoft’s trajectory is comparatively smooth, consistent with the scalability of cloud and subscription models that convert topline growth into predictable operating margins. Amazon’s path is more convex, combining e-commerce scale effects with AWS margin leadership but retaining sensitivity to consumer and logistics cycles. In aggregate, prices during this window track improving fundamentals in a benign liquidity environment, with 2018 Q4 standing out as a temporary risk-off correction before momentum resumed (International Monetary Fund [IMF], 2023; Nasdaq, 2025).

Similarly, as illustrated in Figure 2, trading volume surges cluster around earnings announcements, guidance revisions, and macro policy signals. Apple exhibits the highest baseline turnover—consistent with broad retail ownership and index inclusion—while Microsoft’s activity appears steadier, suggesting a more institutional flow profile. Amazon’s volume is episodic and event-driven, aligned with debates over growth versus profitability during heavy investment phases. These patterns echo established links between information arrival and trading intensity (Chordia et al., 2011) and are broadly consistent with microstructure evidence that liquidity concentrates around news and uncertainty (Bloomberg Intelligence, 2024).

On the other hand, thirty-day rolling annualized volatility, as captured in Figure 3, reveals a market largely characterized by contained risk punctuated by episodic spikes, notably in late 2018, followed by normalization into early 2020. Amazon tends to post the highest realized volatility, consistent with its discretionary demand exposure and operational leverage, while Microsoft’s volatility is most muted, reflecting subscription-based revenues and balance-sheet strength. Apple sits between the two, with variation linked to product-cycle expectations. The realized-volatility profile aligns with the literature on high-frequency volatility dynamics and information flow (Andersen et al., 2003), and with the macro narrative of accommodative financial conditions through most of the sample (International Monetary Fund [IMF], 2023).

In Short, the pre-pandemic era underscores how the big tech companies combined scalable business models with ample liquidity to generate strong returns, robust participation, and generally subdued risk. Apple, Microsoft, and Amazon functioned as both growth engines and market barometers. Their differentiated volatility and volume signatures—higher for Amazon, lower for Microsoft, and mid-range for Apple—reflect differences in revenue composition, operating leverage, and investor base. For empirical work, this window offers a clean benchmark before the exogenous COVID-19 shock, suitable for identifying structural relationships between information flow, liquidity, and price formation (Henriques & Sadorsky, 2025; Nasdaq, 2025).

4. Methodology

4.1. LLM-Based Sentiment Analysis

Sentiment analysis has become one of the central tools in natural language processing (NLP). It involves evaluating subjective information expressed as text in news articles and social media posts to determine whether content conveys positive, negative, or neutral emotions. Techniques that accurately identify sentiment across these domains offer valuable insights. (Medhat et al., 2014; Nyakurukwa & Seetharam, 2024; Yue et al., 2019).

Social media and news articles represent the two dominant information sources in a broad range of applications. They both generate huge volumes of textual material that often convey opinions and sentiments expressed in natural language. Such content is particularly suitable for data mining techniques designed to identify both contextual and emotional dimensions (Yue et al., 2019; Xiao & Ihnaini, 2023). Several research fields have applied sentiment analysis to social media and news data, including stock price forecasting (Shobayo et al., 2024), brand and movie reviews (Medhat et al., 2014), event detection and summarization (Kolajo et al., 2022; Li et al., 2023), target-dependent opinion analysis (Zhang et al., 2022), and community questioning (Roy et al., 2023).

Sentiment analysis employs various techniques to extract attitudes and emotional states by predicting the polarity conveyed—typically marked as positive, negative, or neutral. Nevertheless, the tools involved face difficulties when processing complex aspects such as human language ambiguity, emoticons, slang, sarcasm, and implicit intentions. Each domain presents distinct challenges in terms of collecting representative datasets and identifying and classifying the polarity. Identifying reliable physical and social evidence remains central in elucidating the text’s connotative assumptions and determining accurate polarities (Bhargava et al., 2025).

An important tool for quantifying news tone is the Janis–Fadner (JF) coefficient of imbalance, widely used to characterize the evaluative tenor of firm media coverage (Janis & Fadner, 1965; Deephouse, 2000; Pfarrer et al., 2010). Pfarrer et al. (2010, p. 1139) apply the JF coefficient operationalize the overall affectivity of coverage.

The coefficient is computed as:

$$JF=\left\{\begin{array}{c}({\mathrm{P}}^2-\mathrm{P}·\mathrm{N})/{\mathrm{V}}^2 \mathrm{P}>\mathrm{N}\\ 0 \mathrm{P}=\mathrm{N}\\ \left(\mathrm{P}·\mathrm{N}-{\mathrm{N}}^2\right)/{\mathrm{V}}^2 \mathrm{P}<\mathrm{N}\end{array}\right.$$

(1)

where P is the number of positive articles, N the number of negative articles, and V the total number of articles (P + N). JF ranges from −1 (entirely negative coverage) to +1 (entirely positive). In Pfarrer et al. (2010), the coefficient is rescaled to –100… +100 for interpretability.

In this regard, Araci (2019) and Huang et al. (2023), among others, have recommended the use of FinBERT—a transformer-based language model derived from the Bidirectional Encoder Representations from Transformers (BERT) architecture (Devlin et al., 2019) for conducting sentiment analysis, particularly in financial and economic contexts. Unlike general-purpose sentiment models, FinBERT is pre-trained on large-scale financial corpora—such as Reuters TRC2 news articles and corporate filings—and subsequently fine-tuned on the Financial PhraseBank, a manually annotated dataset of financial news sentences labeled as positive, negative, or neutral (Malo et al., 2014). This domain-specific training allows FinBERT to interpret nuanced expressions in business discourse. For example, phrases such as “shares rallied following the earnings call” or “profits declined despite higher revenue” are accurately classified in terms of their financial sentiment, whereas generic sentiment models often misinterpret them (ProsusAI, 2020).

We employed the publicly available FinBERT-tone variant via the Hugging Face Transformers library (Wolf et al., 2020). For each text sample, the model returned a probability distribution over the three sentiment categories (positive, negative, neutral), calculated via the final softmax layer. The highest-probability label was assigned as the sentiment classification, and the full probability vector was retained to quantify sentiment intensity.

When applying FinBERT, we experimented with three levels of textual granularity. The first approach analyzed each article as a whole, treating it as a single unit. The second approach aggregated paragraphs where the company of interest was mentioned, combining them to form a more focused context. The third examined each sentence mentioning the company individually. After testing all three methods, we chose to retain the paragraph-level aggregation, as it consistently produced the most accurate sentiment classification while preserving sufficient contextual information. This granularity allowed FinBERT to capture nuanced financial expressions without losing interpretability, balancing detail and computational efficiency.

4.2. Model Specification

This section adopts a behavioral-finance perspective in which investor attention and affect (tone) shape belief updating and trading intensity. Conventional media outlets and social platforms act as attention catalysts; negative tone is expected to carry disproportionate weight (negativity bias), and volatility is expected to react asymmetrically (leverage effect). Based on this view, we develop the following four baseline specifications to test the impact of media outlets on stock performance: return, trading volume, and volatility of the three major technology companies: Amazon, Apple, and Microsoft.

Return_i,t = α + β₁ NewsMedia_i,t + β₂ SocialMedia_i,t + β₃ Trend_i,t + β₄ Financial_i,t + ε_i,t

(2)

T_Volume_i,t = ϕ + γ₁ NewsMedia_i,t + γ₂ SocialMedia_i,t + γ₃ Trend_i,t + γ₄ Financial_i,t + u_i,t

(3)

Volatility_i,t = θ + δ₁ * NewsMedia_i,t + δ₂ * SocialMedia_i,t + δ₃ * Trend_i,t + δ₄ * Financial_i,t + v_i,t

(4)

where the dependent variables that explain the stock performance of the giant techs are:

• $Retur{n}_{i,t}$: weekly stock return for firm $i$, calculated as the percentage change in adjusted closing price.
• $T\_Volum{e}_{i,t}$: weekly trading volume, measured as the total number of shares traded (scaled by 1000).
• $Volatilit{y}_{i,t}$: the absolute value of weekly stock returns, approximating realized variability

And a set of explanatory and control variables that capture media coverage, sentiments, public search interest, and firms’ financial fundamentals.

• $NewsMedi{a}_{i,t}$ is a vector that captures traditional media coverage derived from The New York Times via the LexisNexis database. Articles were identified by searching for each company name in the “Company” field, ensuring they were a central topic. To reduce the noise common in article-level sentiment analysis, where multiple entities are often discussed together, we tested several levels of granularity: full article, title, paragraph, and company-specific sentences. Our final model used paragraph-level sentiment, which gave the strongest results. Sentence-level analysis produced similar outcomes, while article- and title-level analysis were more ambiguous.
  • Paragraph Counts: the weekly number of paragraphs mentioning firm $i$, which represents the weekly number of paragraphs from news articles that mention the firm. This serves as a proxy for the volume of traditional media coverage, capturing news-related investor attention.
  • News Sentiment: a composite sentiment score reflecting the overall tone of coverage, or polarity (positive/negative) of news media content. JF coefficient $\times$ 100 (to scale)
• $SocialMedi{a}_{i,t}$ is a vector that represents real-time discourse on Twitter (X). Data was sourced from the existing dataset published by Doğan et al. (2020), which comprises posts from X (formerly Twitter) that mention the three companies. This corpus provides insight into real-time public discourse and reaction on a major social media platform during the study’s timeframe.
  • Tweet Counts: the total number of tweets mentioning firm $i$ in week $t$, regardless of the content. It measures online visibility, public attention, and social buzz, and not necessarily have any financial relevance.
  • Twitter Sentiment: a sentiment score capturing the polarity of tweets, scaled similarly to news sentiment. JF coefficient $\times$ 100
• ${Trend}_{i,t}$ is a Google Trends Index that measures the weekly normalized search intensity for each firm. This data provides a normalized index (0–100) of search volume, as it tracks Google search interest in the firm over time. It serves as a measure of real-time public interest and retail investor attention.

Together, these sentiment variables aim to capture investor attention, tone, and media visibility—factors that have become increasingly influential in shaping market activity, particularly in the age of algorithmic and high-frequency trading.

On the other hand, the financial variables include control variables like Return on Assets and EPS surprise, and ε_i,t is an idiosyncratic error term. Since financial variables are reported quarterly, while sentiments are reported weekly, the financial variables are interpolated linearly to come up with a weekly series for all financial variables, as follows.

• $Financia{l}_{i,t}$ is a vector that contains firm-specific financial fundamentals interpolated from quarterly to weekly frequency. All financial variables were obtained from CRSP or Compustat via WRDS, or calculated from data provided by these sources.
  • Return on Assets (ROA): profitability relative to total assets. It indicates how efficiently a firm is using its assets to generate earnings.
  • Revenue Growth: quarter-over-quarter sales expansion. This reflects the firm’s ability to expand operations and market share.
  • Free Cash Flow: cash generated after capital expenditure. This reflects the firm’s ability to expand operations and market share.
  • Price-to-Book Ratio: market value relative to book value. It is often used to assess valuation, particularly in asset-heavy firms.
  • Earnings Surprise: difference between reported EPS and analyst expectations. Positive surprises often drive short-term stock price increases, while negative ones can lead to declines.

The disturbance terms ${\epsilon }_{i,t}$, $u_{i,t}$, and $v_{i,t}$ represent idiosyncratic shocks for returns, trading volume, and volatility, respectively.

This study leverages a multi-source dataset to analyze public and media attention. All data was collected weekly for a consistent five-year period, from January 2015 to January 2020, ensuring temporal alignment for comparative analysis. The sample period was truncated at the beginning of 2020 for various reasons: first, to avoid the confounding influence of the COVID-19 pandemic, which fundamentally altered both media discourse and financial market dynamics. Secondly, to ensure consistent coverage across multiple data sources and coincides with the rapid rise of social media as an investment information channel, as well as major market events that tested firm resilience. Thirdly and importantly, all three companies, during this period, experienced substantial growth in market capitalization and global influence, making them especially salient cases for studying the interaction between media sentiment and stock performance.

The model is estimated individually for Apple, Amazon, and Microsoft using heteroskedasticity-robust standard errors. All variables are measured contemporaneously, ensuring alignment with the timing of stock return calculations to better capture immediate market reactions. These three companies were chosen not only because reliable, multi-source sentiment data are available for them, but also because they share key characteristics: all three are globally visible technology leaders, extensively covered in both traditional and social media, and widely held by investors. At the same time, they differ in important respects, which makes comparison especially valuable. These similarities and differences together make the three firms suitable for analyzing how media sentiment produces heterogeneous effects on market behavior.

5. Results and Discussion

The empirical investigation begins with both firm-level regressions and a panel fixed-effects model to explore the determinants of daily stock returns. The firm-level regressions allow for capturing company-specific responsiveness to explanatory variables, while the panel fixed-effects model exploits within-firm variation, controlling for time-invariant heterogeneity across Apple, Amazon, and Microsoft.

The following figures illustrate descriptive trends for key variables, including weekly stock returns (Figure 4) and Twitter activity and sentiment (Figure 5) for the three companies. These visuals provide a general sense of how the variables move over time and offer preliminary context for the analysis. However, to assess whether these patterns reflect meaningful relationships, we proceed to formal statistical testing.

Table 1. Descriptive Statistics and Pairwise Correlation matrix (N = 759).

	Variable	Mean	S.D.	1	2	3	4	5	6	7	8	9	10	11
1	Trading volume (Thousands)	25,088.480	21,192.420
2	Stock Volatility (%)	2.572	2.411	0.22
3	Stock Return (%)	0.591	3.477	−0.10	0.18
4	News Paragraph Counts	24.947	32.792	0.37	0.10	0.07
5	News Sentiment	−0.512	19.307	0.04	0.04	0.09	0.02
6	Twitter Counts (Thousands)	3.219	2.329	0.30	0.20	−0.01	0.38	0.00
7	Twitter Sentiment	1.403	1.053	0.11	0.11	0.28	0.05	0.06	−0.24
8	Google Trends index	32.999	25.190	0.19	0.06	−0.04	0.10	−0.03	0.04	0.02
9	Return on Assets	2.506	1.620	0.52	−0.00	−0.04	0.38	0.01	0.39	−0.12	0.02
10	Revenue Growth	5.828	16.927	−0.03	0.05	−0.02	−0.01	0.00	0.02	−0.05	0.08	0.17
11	Free Cash Flow	21,088.190	18,121.530	0.55	−0.02	−0.05	0.41	0.03	0.49	−0.09	0.13	0.68	−0.03
12	Price-To-Book Ratio	11.130	6.314	−0.68	0.07	0.04	−0.39	−0.05	−0.21	−0.05	−0.28	−0.57	0.11	−0.69

reports the descriptive statistics and pairwise correlations among the variables of the model. The results indicate considerable variation across trading activity, media attention, and firm fundamentals, which are consistent with the heterogeneous nature of both information flows and financial performance. Several noteworthy patterns emerge: media and social variables display moderate correlations with trading volume and returns, while financial fundamentals such as return on assets and free cash flow exhibit stronger associations with valuation measures like the price-to-book ratio. Initially, these patterns suggest that both sentiment-related and financial drivers play a complementary role in shaping stock performance.

Table 2. Estimates of the Three Specifications.

	Trading Volume			Stock Volatility			Stock Return
Variables	Apple	Amazon	Microsoft	Apple	Amazon	Microsoft	Apple	Amazon	Microsoft
News Paragraph Counts	6.3	23.7	2.2	0.000	0.003	0.011 *	0.012 **	−0.010	0.013
	(24.3)	(15.1)	(56.7)	(0.004)	(0.015)	(0.006)	(0.005)	(0.020)	(0.009)
News Sentiment	56.8	0.7	6.9	−0.001	0.002	0.007	0.022	0.007	0.029 **
	(101.6)	(5.5)	(92.1)	(0.015)	(0.006)	(0.010)	(0.020)	(0.007)	(0.015)
Twitter Counts (Thousands)	1823.0 ***	1735.2 ***	7534.6 ***	0.326 ***	0.496 **	0.396 *	−0.103	0.033	0.052
	(479.1)	(199.3)	(1936.2)	(0.070)	(0.202)	(0.215)	(0.098)	(0.269)	(0.307)
Twitter Sentiment	−1224.8	951.9 ***	2395.7 ***	0.408 *	0.929 ***	0.299 ***	2.789 ***	2.603 ***	0.469 ***
	(1605.0)	(254.8)	(922.1)	(0.233)	(0.268)	(0.104)	(0.324)	(0.357)	(0.148)
Google Trends index	58.4	−9.4	162.1 ***	0.018	−0.004	0.011	−0.009	−0.003	−0.010
	(111.8)	(10.3)	(60.9)	(0.016)	(0.010)	(0.007)	(0.023)	(0.014)	(0.010)
Return on Assets	8705.9 ***	34.8	696.5	0.065	−0.463	0.143	−0.393	0.016	0.090
	(1421.0)	(410.1)	(1029.3)	(0.208)	(0.426)	(0.116)	(0.289)	(0.567)	(0.165)
Revenue Growth	−99.7 *	32.9 *	1.9	−0.003	0.023	−0.003	0.003	0.023	−0.020
	(56.5)	(19.7)	(204.3)	(0.008)	(0.021)	(0.023)	(0.011)	(0.028)	(0.033)
Free Cash Flow	0.4 ***	−0.0	−0.3	−0.000	−0.000	−0.000	−0.000 *	−0.000	0.000
	(0.1)	(0.1)	(0.2)	(0.000)	(0.000)	(0.000)	(0.000)	(0.000)	(0.000)
Price-To-Book Ratio	−1283.7 **	−152.9	314.9	0.100	−0.047	0.027	−0.096	−0.192	−0.082
	(562.5)	(100.3)	(1138.4)	(0.082)	(0.105)	(0.128)	(0.114)	(0.140)	(0.183)
Constant	−12,671.7	9065.3 ***	41,001.6 ***	1.456	4.527 **	1.948 *	5.496 ***	5.022 *	0.987
	(8515.8)	(1856.4)	(9423.2)	(1.237)	(1.938)	(1.045)	(1.718)	(2.580)	(1.489)
Observations	253	234	245	252	233	244	252	233	244

Standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

reports the regression estimates for the three specifications linking both sentiment-related measures and financial controls to trading volume, stock volatility, and stock returns across Apple, Amazon, and Microsoft. Each column corresponds to a distinct outcome variable, with standard errors reported in parentheses and significance levels denoted by conventional thresholds.

The discussion proceeds in two stages. First, the analysis is conducted vertically, by interpreting results within each outcome variable—trading volume, volatility, and returns—separately. Second, the analysis is extended horizontally, by comparing the influence of news media versus social media variables across the three firms, highlighting similarities and differences in their responsiveness to different information channels.

5.1. Impact on Trading Volume

Variables from traditional news media did not show a significant impact on the weekly Trading Volume of the three firms: Apple, Microsoft, and Amazon. However, all companies experience significantly increased trading volume with a rise in Twitter Counts. “Attention Hypothesis” states that spikes in media coverage or online chatter can increase investor attention and subsequently trading activity, even in the absence of financial fundamental news (Barber & Odean, 2008). In these firms, while prices may remain anchored by companies’ financial fundamentals, trading volume is a more responsive indicator of sentiment.

Conversely to sentiments from traditional news outlets, Twitter Sentiments has a significant impact on the trading volume for Microsoft and Amazon, but not for Apple. Simply Wall Street (2025), a financial research and analysis platform, reports that Microsoft’s recent involvement in AI developments, especially its strategic alliance with OpenAI, has made it a frequent topic in financial social media. While this fuels attention and trading volume, institutional holders continue to act as a stabilizing force, preventing sentiment from fully translating into price volatility. On the other hand, Amazon’s nature of investor base and the types of conversations the firm generates online may also have a significant impact on these results. While Amazon is a highly visible company, much of its social media engagement centers on consumer products, logistics delays, or labor relations—topics that are not always directly relevant to investor sentiment. Additionally, Amazon’s earnings surprises and long-term investment strategies are often better understood through traditional financial disclosures and macroeconomic indicators, rather than short-form sentiment captured on social media platforms.

As for Apple, trading volume did not show a significant response to Twitter Sentiment, possibly because the stock is already one of the most heavily traded in the world. With such high baseline liquidity, marginal changes in tweet volume may be absorbed without observable changes in volume metrics. Instead, the reaction manifests more in prices than trades (Bollen et al., 2011).

These findings carry important implications for investors and analysts. It suggests that Twitter sentiment metrics should be interpreted differently depending on firm-specific characteristics—not all firms respond similarly to social buzz. Amazon and Microsoft offer potential alpha-generating signals through Twitter Sentiment, while Apple appears more resilient to this type of sentiment noise.

5.2. Impact on Stock Volatility

Subsequently, the results indicate that traditional media news and sentiment variables indicate a negligible role in explaining stock volatility. News paragraph counts are largely insignificant, except for Microsoft, where a weak positive association is observed. This clearly suggests that neither the intensity of the coverage nor the tone in conventional news media outlets can systematically increase volatility across the three firms.

In contrast, social media variables show strong and consistent effects. Both Twitter mention counts and sentiment exhibit a significant positive impact on stock volatility across the three companies. Higher volumes of Twitter activity, coupled with more extreme sentiment, amplify fluctuations in stock prices. This finding supports the interpretation that social media operates as a channel through which market uncertainty and risk are amplified, while the influence of traditional media remains marginal.

Furthermore, the parallel patterns observed in trading volume and volatility reinforce this conclusion. Social media activity exerts a powerful effect on both indicators, whereas the impact of traditional news media is weak in both cases. Given that trading volume and volatility reflect different dimensions of investor behavior, participation intensity, and uncertainty, this suggests that social media simultaneously fuels both the breadth of market engagement and the magnitude of price swings.

5.3. Impact on Stock Returns

Traditional News media influence on stock returns is selective–News Paragraph Counts has a significant positive impact only on Apple, and News Sentiment has a significant positive impact only for Microsoft. Other cases are statistically insignificant. These differences suggest that firm-specific characteristics such as investor composition, visibility, and news sensitivity may mediate the strength of social sentiment effects on asset prices.

On the other hand, insights from social media revealed a clear link between Twitter sentiment and stock performance. Specifically, positive sentiment on Twitter consistently predicted positive stock returns for all three companies. This suggests that social media platforms act as an early reflection of investor mood and market expectations, where collective opinions and emotions expressed online can translate into measurable financial outcomes. In other words, Twitter sentiment not only mirrors public perception but also serves as a leading indicator of short-term stock movements.

For Apple, a one-unit increase in Twitter Sentiment is associated with a 2.789 percentage point rise in weekly stock returns. Amazon shows a nearly identical effect size of 2.603, suggesting that sentiment shocks during this period had a similarly strong influence on both firms. Microsoft also responds positively, though the magnitude of its coefficient is more modest at 0.469. The comparable impact for Apple and Amazon is particularly noteworthy given the 2015–2019 sample period, when both firms were at the center of retail investor enthusiasm and intense media attention. Apple’s continuous product cycles and brand visibility, combined with Amazon’s rapid expansion in e-commerce and cloud services, made both companies frequent subjects of social media chatter and investor speculation.

Although Apple and Amazon both have substantial institutional ownership, their large and highly active retail investor bases created conditions under which Twitter sentiment could quickly spill over into asset prices. This contrasts with Microsoft, where institutional investors account for over 70% of ownership (WallStreetZen, 2025), providing a stabilizing influence that dampens—but does not eliminate—the effect of sentiment on weekly returns (Yoon & Oh, 2022).

Behavioral finance theories offer strong support for the Twitter Sentiment influence of stock returns. According to noise trader theory (De Long et al., 1990), retail investors’ actions in response to sentiment can cause temporary mispricing and excess volatility. Apple’s strong price response fits this pattern, whereas institutional stability at Microsoft and Amazon dampens such return shocks but still enables sentiment-driven trades.

Real-world cases support this behavior. For example, activist investor Carl Icahn’s tweets about Apple in 2013 significantly moved the stock price in a matter of minutes. These anecdotal instances underscore how Twitter sentiment—especially from influential accounts—can trigger market reactions, particularly in retail-favored equities like Apple.

Overall, traditional media influence was idiosyncratic to each company, and the effects were not very strong. However, the social media influence of stock market activities tends to be consistent across companies, with some exceptions. Twitter Counts were consistently positively significant for Trading Volume and Stock Volatility, and Twitter Sentiment is consistently positively significant for Stock Returns.

Among the sentiment variables included in the model, such as News Paragraph Counts, News Sentiment, Twitter Counts, and Google Trends index, only Twitter Sentiment consistently demonstrated statistical significance, particularly for Apple’s stock returns and Microsoft and Amazon’s trading volumes. This outcome is likely due to the focused, real-time nature of Twitter Sentiment, which is filtered to reflect financially relevant discourse. In contrast, Twitter Counts often capture general chatter or non-investment-related noise, diluting their relationship with market movements (Bollen et al., 2011; Sprenger et al., 2014). Similarly, News Paragraph Counts and News Sentiment, which are derived from news articles, may lag behind events due to slower editorial processes, making them less timely for capturing intraday sentiment shifts that influence trading decisions (Tetlock, 2007). Meanwhile, Google Trends index data measures public attention but fails to distinguish between interest from casual browsers and active investors (Preis et al., 2013).

Moreover, Twitter Sentiment aligns more directly with trading behavior in high-frequency environments. These tweets seem to capture breaking speculation, earnings chatter, analyst comment, or retail investor opinion that can immediately influence prices and turnover, especially among tech-heavy stocks with heavy public attention. This corresponds with attention-based trading theory (Barber & Odean, 2008) and noise trader risk (De Long et al., 1990) that assert increasing attention, and behavioral biases are likely to lead to excess trading or short-term mispricing. For Apple, whose stock is retail-oriented and much commented upon online, Twitter Sentiment indicated sentiment that translated into immediate price action, while for Microsoft and Amazon—more institutionally owned stocks—they elicited volume reactions as opposed to immediate price movement (Yoon & Oh, 2022). In sum, the unique targeting and immediacy of Twitter Sentiment make them more powerful in capturing short-term sentiment shocks than broader, less-filtered indicators.

Beyond the baseline findings, several broader insights can be drawn. First, firm-specific differences show that social media sentiment does not affect all companies in the same way. Apple’s high baseline trading volume means that additional social media activity is less likely to change liquidity and instead appears more directly in price movements. Amazon, on the other hand, shows the highest realized volatility, making it especially sensitive to sentiment shocks that amplify uncertainty and risk. Microsoft displays the opposite profile: with lower volatility and strong institutional ownership, Twitter sentiment is reflected more in trading volume than in immediate price changes.

Second, comparing traditional news and social media reveals a clear contrast. Traditional news variables, such as paragraph counts and sentiment, generally have weak or inconsistent effects on trading volume, volatility, and returns. This likely reflects the more deliberate reporting cycles of traditional media, which often trail the immediate reactions captured on social platforms. By contrast, social media—especially Twitter sentiment—shows immediate and consistent effects, influencing both trading activity and short-term returns.

Third, the role of Google Trends appears limited. Search intensity has little explanatory power, consistent with earlier studies showing that it reflects broad public curiosity rather than targeted investor attention. Any firm-specific variation was not systematic and does not alter the broader conclusion that Google Trends is a weaker signal compared to social media sentiment.

Fourth, the weak influence of financial control variables should be noted. Measures such as ROA, revenue growth, and free cash flow show limited and inconsistent effects across firms. One reason is methodological, since quarterly data were interpolated into weekly series, which smooths out variation. More importantly, the stable fundamentals of large technology firms mean that short-term fluctuations in trading activity and returns are more closely tied to shifts in attention and sentiment than to incremental changes in fundamentals. This aligns with behavioral finance research, which argues that short-term dynamics are often shaped by investor attention and sentiment rather than fundamentals.

Last but not least, advances in large language models (LLMs) provide a promising direction for future sentiment analysis. While our study used conventional dictionary- and score-based measures, LLMs offer the ability to capture context, subtle tone, irony, and domain-specific meanings that are often missed by traditional methods. Applying LLM-based sentiment measures to news and social media could reduce noise, sharpen the distinction between financial and non-financial discourse, and improve the predictive value of sentiment in explaining market behavior.

Overall, this broader discussion underscores the contributions of our analysis. By examining how social media sentiment relates to trading activity, volatility, and returns across different firms, we provide evidence that media-based measures can capture aspects of investor behavior that are not fully reflected in fundamentals. Our results indicate that, relative to social media sentiment, traditional news and Google Trends provide weaker explanatory power for short-term dynamics, consistent with their broader and less time-sensitive nature. At the same time, our findings suggest that short-term dynamics can be shaped by sentiment indicators, even if financial variables remain central to long-term value. Looking forward, advances in large language models offer the potential for more nuanced sentiment analysis that distinguishes financially relevant discourse from noise. Future research combining refined sentiment measures with fundamentals could provide a clearer picture of how information, attention, and valuation interact in financial markets.

6. Robustness Check

The study maintains the OLS framework as the main estimation approach for several methodological and empirical reasons. First, the OLS model provides a transparent and interpretable benchmark for examining the contemporaneous relationship between sentiment indicators and stock returns. It allows a straightforward assessment of how different sentiment measures derived from news, social media, or search trends affect market performance, while remaining consistent with the methodology widely used in the literature on investor sentiment and asset pricing. Secondly, adopting more sophisticated dynamic specifications (such as VAR or ARDL models) is not feasible given the structure and frequency of the dataset. Our data consists of firm-level daily observations that have been converted into weekly, and expanding the model to a system-based (VAR) or lag-augmented form (ARDL) would require a substantially larger dataset, in order not to compromise degrees of freedom, and/or additional identifying assumptions that are beyond the current scope of the paper.

To ensure the robustness of the OLS results, we will include robustness checks by testing alternative functional forms. That is, this section focuses on assessing the stability of the baseline model to ensure that the estimates are not driven by specific model choices or omitted-variable bias. Thus, a series of robustness checks is conducted for all three companies via re-estimating each of the three main models—returns, trading volume, and volatility—using alternative specifications. These specifications sequentially add or remove control variables such as return on assets, revenue growth, free cash flow, and the price-to-book ratio to test for sensitivity to model composition.

Using Apple as an example, alternative specifications were run in which the control variables were gradually included and excluded to test the sensitivity of the coefficients. Across these variations, the main results remain stable. In every specification, Twitter sentiment continues to show a strong and statistically significant positive effect on Apple’s daily stock returns. The coefficient of Twitter sentiment remains large in magnitude, while news sentiment and paragraph counts keep their expected but weaker signs. None of the additional controls, such as revenue growth, return on assets, or the price-to-book ratio, changes the conclusion that market performance responds faster to social media tone than to conventional news sentiment.

Across all specifications (reported at the Appendix A), the results for stock returns are remarkably consistent. Twitter sentiment remains positive and highly significant for all three companies, suggesting that an optimistic tone in tweets is reliably associated with higher next-week returns. News sentiment maintains the expected positive sign but with weaker and often statistically insignificant effects, supporting the view that social media captures short-term investor mood more effectively than conventional outlets.

The trading volume regression specifications report almost the same results as the stock return. For Apple, Amazon, and Microsoft, Twitter counts are strongly significant at the 1% level across all models, indicating that the intensity of online discussion is a key driver of market activity. On the other hand, news-related variables, both paragraph counts or sentiment, show less systematic effects, implying that social media conversations better explain bursts in trading volume.

The volatility models further confirm the robustness of these relationships. While volatility responds positively to changes in both Twitter and news sentiment, the magnitudes are generally higher for social media indicators, consistent with the argument that investor disagreement and attention amplified through Twitter elevate short-term risk. Importantly, the coefficients across specifications are stable in sign and size, and the R-squared values remain virtually unchanged.

Overall, the near-identical patterns observed across Apple, Amazon, and Microsoft confirm that the baseline models are well specified. Both Twitter and news sentiment retain their signs and significance levels. The stability of coefficients across all robustness checks indicates that the baseline model is well specified and resilient to alternative functional forms. Collectively, the evidence across Apple, Amazon, and Microsoft confirms that the relationship between sentiment and stock performance is robust and not an artifact of model design.

7. Conclusions

This study examined the role of traditional news media, social media, and firm-level financial fundamentals in shaping the stock performance of three leading technology firms, Apple, Microsoft, and Amazon, over the period January 2015 to January 2020. By integrating datasets from The New York Times, Twitter, Google Trends, and firm financial statements with market data from Nasdaq, this paper provides a multifaceted assessment of how information flows influence stock performance in terms of trading volume, volatility, and returns.

The results demonstrate that while both news media and social media contribute significantly to explaining stock performance, social media consistently exerts a stronger influence across all three specifications. In the trading volume regressions, Twitter counts and sentiment emerged as robust predictors of market activity, suggesting that investor attention is increasingly driven by real-time discourse rather than traditional coverage. Similarly, in the volatility models, social media sentiment showed a more pronounced effect, reflecting its role in amplifying market uncertainty during both positive and negative shocks. Finally, for stock returns, the tone and intensity of social media activity explained a greater share of variation compared to conventional news measures, underscoring the growing importance of digital platforms in shaping investor expectations and short-term price movements.

These findings highlight the shifting landscape of financial information. Traditional media remains important, particularly as a baseline source of credibility, yet social media has become the dominant channel for attention and sentiment transmission in capital markets. The study illustrates that even before the pandemic-era acceleration of digital discourse, social media had already surpassed news media in market relevance. For practitioners and regulators, these results emphasize the need to account for the outsized role of online platforms in influencing market behavior, both as a source of opportunity and as a channel of systemic risk.