What’s Trending? Stock-Level Investor Sentiment and Returns

Abstract

We study a direct, firm-level measure of investor sentiment derived from social media (BTSS sentiment). While related to firm fundamentals, BTSS sentiment contains a substantial non-fundamental component. We decompose sentiment into fundamental and pure sentiment and show that return predictability and reversal are primarily driven by the latter. Sentiment is persistent and systematic in the short term. High sentiment predicts elevated concurrent returns and subsequent reversal within a year. The effect is strongest in hard-to-value stocks, such as small and young firms, where limits to arbitrage are more binding.

1. Introduction

Episodes such as the dot-com bubble and the 2008 financial crisis have underscored the role of investor sentiment in driving asset prices away from fundamentals. When optimism or pessimism dominates markets, prices may deviate substantially from intrinsic values, and rational investors may be unable to fully arbitrage away the mispricing. Investor sentiment plays a vital role in shaping asset prices. Theoretical work by authors such as De Long et al. (1990) has used investor sentiment to explain stylized facts in finance such as excess volatility, closed-end fund discount, and equity premium.

However, to empirically test the implications of investor sentiment, we face the challenge of measuring investors’ elusive non-fundamental expectations. An exogenous shock to investor sentiment may lead to a chain of events, beginning with some market participants forming a belief and subsequently translating to observable patterns of securities trades. The demand pressure combined with limits to arbitrage causes mispricing, which may then trigger responses by informed insiders. The previous literature has tried to track the traces of sentiment in every part of the chain. Proxies of sentiment include (ordered from origins in investor psychology to responses by insiders) the following: sports outcomes, investor survey, consumer confidence index, trading volume, retail investor trading, mutual fund flows, closed-end fund premium, dividend premium, option implied volatility, IPO first day return, equity issuance, and insider trading (see, e.g., Lee et al., 1991; Chen et al., 1993; Neal & Wheatley, 1998; Brown & Cliff, 2004, 2005; Baker & Wurgler, 2006, 2007; Edmans et al., 2007; Frazzini & Lamont, 2008; Yu & Yuan, 2011; Doukas et al., 2010; Ben-Rephael et al., 2012; Chung et al., 2012; Baker et al., 2012; W. Wang et al., 2021). Studies by authors such as Baker and Wurgler (2006) and Huang et al. (2015) use the principal component of six of the proxies1 to capture the major episodes of investor sentiment. However, all of these are indirect measures subject to confounding influences.2

The recent literature has turned to social media as a more direct and timely source of sentiment. Studies show that aggregate sentiment from platforms like Twitter can predict market-level returns (Bollen et al., 2011; Mao et al., 2015). Gan et al. (2020) demonstrate that, in the U.S., social media has become the dominant sentiment source relative to traditional media. Liang et al. (2020) show a similar pattern in China, where online sentiment better explains market volatility than newspaper sentiment. Lachana and Schröder (2024) find that investor sentiment from Seeking Alpha predicts stock returns more accurately than sentiment from the Wall Street Journal. Complementing these findings, Chari et al. (2023) show that aggregate news sentiment meaningfully impacts market returns on days with extreme movements in India.

At the firm level, Gu and Kurov (2020) use Bloomberg’s Twitter sentiment data to show that daily sentiment predicts same-day returns in the Russell 3000. Dong and Gil-Bazo (2020) find that short-term returns in the Chinese stock market are temporarily influenced by social media sentiment. During the COVID-19 pandemic, Duan et al. (2021) show that firm-level sentiment from Sina Weibo and official news media predicts daily returns, turnover, and speculative activity, with social media sentiment having a more reactive and stronger effect. X. Wang et al. (2022), using data from EastMoney Guba, experimentally show a same-day sentiment effect on returns. Other high-frequency studies examine intraday effects of social media sentiment on individual stocks (Broadstock & Zhang, 2019) and ETFs (Sun et al., 2016; Renault, 2017).

In this paper, we utilize the Bloomberg Twitter Sentiment Score (BTSS), a direct and high-frequency measure of investor sentiment toward individual stocks, constructed by Bloomberg from social media content, and extend its application to examine longer-horizon return predictability and sentiment-driven mispricing. Since 20123, Bloomberg has used a proprietary algorithm to classify stock-related posts from Twitter and StockTwits as positive or negative. These signals are aggregated at the firm-day level to produce a daily sentiment score, which we then average to the monthly level for use in our analysis.

BTSS sentiment offers several advantages over traditional sentiment proxies. It captures investor beliefs directly from real-time discourse, with minimal institutional filtering; it allows for stock-level variation suited to cross-sectional return analysis; and it reflects broad investor mood due to Twitter’s large and diverse user base. Compared to studies that extract sentiment from message boards (e.g., Antweiler & Frank, 2004; X. Wang et al., 2022) or financial blogs like Seeking Alpha (e.g., Lachana & Schröder, 2024), BTSS offers a standardized, scalable approach based on real-time discourse from a broad investor base. Unlike user-curated platforms that may reflect niche or biased audiences, BTSS sentiment reflects widespread investor mood and benefits from machine-learning-based tone classification rather than simple keyword counts, reducing noise and enhancing replicability. These features allow for more precise stock-level variation in sentiment and facilitate cross-sectional asset pricing analysis. Our final dataset includes 3955 firms over the period January 2015 to December 2017.

We begin our analysis by examining the characteristics and determinants of the sentiment measure. We find that the BTSS sentiment measure is positively related to the sentiment of press release and news media and earnings surprises. Investor sentiment tends to be more optimistic about large and growth firms. Overall, our results suggest that investor sentiment is partly driven by firms’ fundamental information, but a large portion of it remains unjustified by firm information.

Existing findings suggest that sentiment, one of the driving forces of retail investors’ trading activities, seems to be persistent and systematic. Barber et al. (2009) find that the order imbalances of retail investors usually persist over several weeks, and Kumar and Lee (2006) show that the trades of individual investors are systematically correlated. Therefore, as a second step, we examine whether the BTSS investor sentiment measure exhibits such properties that one would expect from a good sentiment measure. We construct a monthly contingency table of sentiment decile portfolio assignments and find evidence of persistence in the individual sentiment values, particularly for the optimistic sentiment. We also document that stock-level sentiment is negatively related to industry and market sentiment, suggesting that investor attention is allocated relatively and stocks receive less positive sentiment when broader sentiment is high.

Having validated the sentiment measure, as our third step we test its predictive power on stock returns. The expectation is that, in times of irrational investor optimism, stocks become overvalued and should subsequently experience negative returns, while negative sentiment would drive undervaluation and subsequent positive returns. To test the hypothesis, we sort stocks into deciles in each month and examine the abnormal returns in the concurrent month and the following 12 months. We find that higher investor sentiment is associated with higher concurrent return and lower subsequent returns from months 7 to 12. The result suggests that, indeed, higher sentiment causes temporary overvaluation which subsequently reverses as the fundamentals are revealed. To control for other risk factors that may affect stock prices, we also run a Fama-Macbeth regression by including commonly used controls. The results are qualitatively similar.

Baker and Wurgler (2006) argue that hard-to-value stocks are more sensitive to speculative demand and more costly and risky to arbitrage. So, we should expect investor sentiment to have a stronger effect on hard-to-value stocks such as young firms and small firms. As the fourth step, we follow Baker and Wurgler (2006) and compare the cross-section of predictability of investor sentiment on the returns. Indeed, we find that the predictability of sentiment is stronger for small and young firms.

One may argue that the BTSS sentiment proxy reflects economic fundamentals to some extent, and the concurrent relationship between the BTSS sentiment proxy and return simply reflects investors’ rational reaction to fundamental news. We have two responses to this skeptical view. One, the long-term reversal that we document is not a natural implication of the view. Rather, it suggests that the positive relationship between the BTSS sentiment proxy and return is at least partly driven by investors’ unjustified sentiment. Two, we remove the influences of fundamentals, at least partially, by regressing the BTSS sentiment proxy on a set of fundamental indicators—sentiment in official news media, stock price, firm size, and book-to-market. The predicted value is then defined as Fundamental Sentiment and the residuals as Pure Sentiment. We find that, although both measures are positively related to concurrent stock returns, the magnitude is much larger for Pure Sentiment. In addition, we find the reversal for Pure Sentiment as early as the 5th month in the future.

Gu and Kurov (2020) also utilize the Bloomberg sentiment measure; however, our study distinguishes itself in several key ways. First, while they examine the predictive power of Twitter sentiment for daily stock returns and find evidence of influence without subsequent reversals, suggesting it conveys information quickly incorporated into prices, we focus on the relationship between firm-level social media sentiment and subsequent monthly returns, specifically analyzing for evidence of price reversal within the following year. Second, we propose and utilize a novel decomposition of firm-level social media sentiment into Fundamental Sentiment and Pure Sentiment. This decomposition allows us to demonstrate that the observed subsequent price reversal is particularly driven by the Pure Sentiment component, supporting the idea that sentiment causes temporary misvaluation due to irrational trading, a phenomenon we find is more pronounced in hard-to-value stocks subject to limits to arbitrage.

This paper contributes to the growing literature on the importance of sentiment in asset pricing, by adding an important stock-level dimension with a long-term focus. Our primary contribution is proposing a new direct stock-level measure of sentiment which shows valid and expected characteristics. We further decompose this measure into a fundamentals-based and a pure sentiment component, allowing us to isolate the behavioral effects of investor sentiment on returns. We also test the theoretical predictions regarding the impact of sentiment on the cross-section of concurrent and future monthly stock returns.

These findings have practical relevance for investors, asset managers, and corporate decision-makers. For investors, firm-level social media sentiment offers a timely signal that can inform trading strategies, particularly contrarian positions in hard-to-value stocks prone to sentiment-driven mispricing. The decomposition into Fundamental and Pure Sentiment helps distinguish between information-driven and behavioral return components. For firms, the results highlight the tangible impact of social media discourse on stock prices, underscoring the importance of monitoring and managing investor sentiment in digital channels.

The rest of the paper is organized as follows. Section 2 describes the data used for this research, Section 3 examines the characteristics of the investor sentiment measure, Section 4 presents analysis of the relationship between sentiment and abnormal returns, and Section 5 concludes.

2. Data and Sample

We use the Bloomberg Twitter Sentiment Score (BTSS) as a direct measure of investor sentiment toward individual publicly traded companies. BTSS is derived from social media posts, specifically, tweets and StockTwits messages, that mention a given company using identifiers such as the cashtag (e.g., $AAPL), ticker symbol, or company name.4 Bloomberg integrates this content through its proprietary natural language processing pipeline, which assigns a daily sentiment score ranging from −1 (strongly negative) to +1 (strongly positive). Each score reflects the tone of the aggregated posts and incorporates a confidence metric based on the consistency and volume of classified content.

Importantly, the algorithm is designed to filter for signals relevant to financial and business information, excluding sentiment about unrelated content such as marketing, entertainment, or customer service discussions. According to Bloomberg representatives, the sentiment classification process is powered by supervised machine learning techniques,5 including support vector machines, decision trees, and regression-based models trained on annotated datasets.

We obtain daily BTSS sentiment data for 3955 firms over the period January 2015 to December 2017 and compute monthly averages at the firm level for use in our empirical analysis. The initial dataset contains over 6.6 million raw sentiment observations, which are reduced to 1.5 million after dropping firms with missing values or data errors. Merging with CRSP identifiers further narrows the sample to 1.34 million observations, and after monthly aggregation and filtering for complete data, the final analysis dataset includes 64,302 firm-month observations. We multiply the monthly BTSS variable by 100 to enhance the interpretability of the regression coefficients.

To measure sentiment about individual firms in traditional media and corporate press releases, we use RavenPack News Analytics, a leading provider of structured news sentiment data. RavenPack collects and analyzes news from thousands of global sources, including major publishers, newswires, and financial content aggregators. Using a proprietary natural language processing algorithm, RavenPack assigns a sentiment score to each piece of news content by evaluating the tone (positive, negative, or neutral), the type of event reported (e.g., earnings announcement, litigation, and executive turnover) and a relevance score, which captures how central the firm is to the story. For our analysis, we include only news articles with a relevance score above 75,6 ensuring that the firm is a primary subject of the news. We then compute the monthly average sentiment score for each firm by aggregating the daily sentiment values across all qualifying news items. This yields a time series of firm-level news sentiment that reflects the tone of credible, finance-relevant media coverage.

We use three distinct RavenPack sentiment measures in our analysis. The first is the overall Average Event Sentiment (AES), which aggregates sentiment from all qualifying news sources, including both third-party media and company-issued press releases. The second is AES-news, which isolates sentiment derived exclusively from independent third-party news articles. This measure serves as a proxy for objective, externally generated stock-level sentiment. The third is AES-press release, which captures the tone of firm-issued communications, such as earnings announcements and corporate updates. This allows us to distinguish between externally driven sentiment and sentiment that may be strategically managed by the firm itself.

Daily and monthly stock data come from the Center for Research on Security Prices (CRSP), accounting data from Compustat and analyst forecast data from I/B/E/S. The sample includes common stocks (CRSP codes 10 and 11) listed on NYSE, AMEX and NASDAQ. To mitigate market microstructure noise, we restrict the sample to stocks with prices between $5 and $1000, applied consistently throughout the sample period.

We include a broad set of stock characteristics that can potentially impact monthly returns: logarithm of book-to-market ratio (lnbm); momentum (mom); idiosyncratic stock volatility (ivol); market beta (beta); Amihud illiquidity measure (illiq); firm size—logarithm of market value (lnme); earnings surprise; analyst earnings forecast dispersion (disp); stocks co-skewness (coskew); extreme positive stock return (max); and abnormal dollar volume (voldu). These controls are standard in the asset pricing and behavioral finance literature (e.g., Amihud, 2002; Bali et al., 2014). Detailed descriptions of variable construction are available in Online Appendix A.

Table 1. Summary statistics and correlation matrix. The table presents characteristics of monthly variables used in the analysis as well as the correlation matrix. BTSS sentiment is the monthly aggregated sentiment measure derived from social platforms; AES is the aggregate sentiment from all qualifying news sources (including both third-party media and company-issued press releases); AES-news is sentiment derived exclusively from independent third-party news articles; AES-pr captures the tone of firm-issued communications, such as earnings announcements and corporate updates; excess return is stock return less the risk-free rate in percentage points; beta is market Beta; size is the natural logarithm of firm’s market value; lnbm is natural logarithm of stock’s book-to-market ratio; mom is 11-month momentum; coskew is stock’s monthly co-skewness; ivol is idiosyncratic volatility; max is the stock’s extreme positive return; disp is analyst forecast dispersion; earn. surprise is the measure of unexpected earnings; illiq is Amihud illiquidity factor, and voldu is the abnormal dollar volume.

Variable					Obs				Mean				Std. Dev.
BTSS sentiment					64,302				1.978				4.358
AES					63,378				68.279				19.414
AES-news					62,979				67.987				20.439
AES-pr					54,642				69.789				20.117
Excess return					64,301				1.303				13.800
Beta					60,228				1.226				0.962
Size (lnme)					64,302				7.158				1.814
Ln B/M					61,924				−0.857				0.891
Mom					64,302				6.188				38.904
Coskew					60,228				−0.003				0.128
Ivol					64,302				1.724				1.678
Max					64,302				4.932				7.071
Disp					64,302				0.130				1.707
Earn. surprise					64,302				0.000				0.009
Illiq					64,302				0.889				20.176
Voldu					64,302				0.238				1.614
	BTSS	AES	AES News	AES pr	Ex. ret	Beta	Size	Lnbm	Mom	Cosk	Ivol	Max	Disp	Earn. Surpr	Illiq	Vold
BTSS sent	1.00
AES	0.03	1.00
AES-news	0.03	0.98	1.00
AES-pr	0.02	0.89	0.82	1.00
Excess return	0.13	0.04	0.05	0.03	1.00
Beta	0.00	−0.18	−0.18	−0.14	−0.01	1.00
Size (lnme)	0.18	−0.10	−0.12	−0.02	−0.01	−0.02	1.00
Ln B/M	−0.10	−0.10	−0.10	−0.09	0.02	−0.05	−0.31	1.00
Mom	0.13	0.17	0.19	0.11	−0.07	−0.16	0.07	−0.01	1.00
Coskew	0.02	−0.06	−0.05	−0.04	0.00	−0.37	0.06	0.03	0.04	1.00
Ivol	−0.03	−0.15	−0.15	−0.12	0.32	0.14	−0.24	−0.02	−0.11	−0.05	1.00
Max	0.03	−0.10	−0.10	−0.08	0.52	0.12	−0.14	−0.02	−0.10	−0.04	0.89	1.00
Disp	0.00	−0.09	−0.09	−0.08	0.00	0.04	−0.01	−0.02	−0.05	0.00	0.03	0.02	1.00
Earn. surprise	0.03	0.05	0.05	0.04	0.03	0.00	0.01	0.00	0.02	0.00	−0.01	0.01	−0.01	1.00
Illiq	−0.02	0.07	0.07	0.04	0.02	−0.02	−0.08	0.04	−0.01	−0.01	0.06	0.04	−0.01	0.00	1.00
Voldu	0.11	0.08	0.09	0.05	0.26	−0.06	0.02	0.01	0.23	0.01	0.21	0.22	−0.02	0.02	−0.01	1.00

presents summary statistics and the correlation matrix. On average, BTSS sentiment scores are positive, consistent with the tendency of investors to be optimistic. Sentiment is most strongly positively correlated with firm size and abnormal trading volume and most negatively correlated with book-to-market ratio, suggesting greater optimism about large and growth-oriented firms. Correlations with news-based sentiment measures are positive but modest, reinforcing the distinct informational content captured by social media sentiment.

3. Characteristics of Stock-Level Sentiment

We begin our analysis by examining the determinants of BTSS. Our objective is to assess how much of this sentiment can be attributed to firms’ fundamental information, their own communications, the tone of media coverage, and other firm-level characteristics. News sentiment can be argued to be a more objective and information-based measure than the elusive investor sentiment, and along with the fundamental information, can help us identify the irrational portion of the stock sentiment measure.

Table 2. Determinants of the investor sentiment. The table presents the results of the stock-level OLS regression of BTSS investor sentiment on a variety of determinants; the media sentiment measures (AES, AES-news, and AES-pr) and well as stock characteristics (size measured as natural logarithm of market value, natural logarithm of trading volume, logarithm of end of month price, logarithm of book-to-market ratio, lagged end of month return, earnings surprise, and earnings announcement day). The regression includes time and industry fixed effects and heteroscedasticity consistent standard errors. * p < 0.1; ** p < 0.05; *** p < 0.01.

	Dependent Variable: BTSS Investor Sentiment
	1	2	3	4	5	6	7	8
AES	0.020 ***				0.025 ***	0.025 ***	0.023 ***	0.021 ***
	[12.76]				[13.87]	[14.19]	[13.98]	[13.31]
AES-news		0.020 ***		0.038 ***
		[12.33]		[18.02]
AES-pr				0.014 ***	−0.017 ***
			[10.81]	[−9.60]
Earnings announcement day					0.412 ***
					[5.76]
Earnings surprise						12.536 ***	12.523 **
						[2.97]	[2.71]
Lag (excess return)					−0.183	−0.173	−0.246
					[−0.62]	[−0.59]	[−0.88]
Ln trading volume					0.373 ***	0.379 ***	0.214 ***
					[10.90]	[11.29]	[6.22]
Ln price							0.005 ***	0.002 ***
							[8.88]	[4.81]
Size (lnme)							0.166 ***	0.385 ***
							[4.76]	[12.08]
Ln B/M							−0.148 ***	−0.153 ***
							[−2.98]	[−3.14]
Constant	1.413 ***	1.422 ***	2.088 ***	1.577 ***	−3.010 ***	−2.924 ***	−2.370 ***	−1.445 ***
	[3.87]	[3.84]	[4.61]	[3.22]	[−5.39]	[−5.38]	[−4.10]	[−3.07]
Time and industry fixed effects	YES	YES	YES	YES	YES	YES	YES	YES
N	61,692	61,324	53,243	53,228	59,065	59,065	56,958	59,480
R-sq	0.065	0.066	0.059	0.068	0.090	0.089	0.097	0.095

reports results from an OLS regression of monthly BTSS sentiment on various concurrent variables measuring fundamental information about the company and news sentiment measures. All models include time and industry fixed effects and use heteroscedasticity-consistent standard errors.

Across all specifications, we find that, overall, news sentiment AES and AES-news are strongly positively related to BTSS sentiment. However, in column 4, AES-press releases is negatively associated with BTSS sentiment, suggesting that investors discount or react more cautiously to firm-issued communications compared to independent media coverage.

In columns 5 and 6, we find that the arrival of new fundamental information (earnings announcement) is positively related to the stock-level investor sentiment. We observe that the sentiment goes up when the firms announce positive earnings surprises or have high trading volume (expressed as logarithm of trading volume), which can point to higher attention and interest from the investors. In columns 7 and 8, we introduce the fundamental firm characteristics, such as the logarithm of the end-of-month price, size (logarithm of end-of-month market value), and logarithm of book-to-market ratio. We find that the fundamental information about the firm has a significant relation to the investor sentiment, while the news sentiment maintains its significance. In particular, investors tend to be more optimistic about large and growth firms. These results suggest that the BTSS sentiment measure picks up both the fundamental information and the sentiment of the press and media.

While many predictors are statistically significant, the explanatory power of the regression model in

Table 3. Monthly 3 factor Fama–French returns in portfolios based on the stock-level sentiment measure. Each month we sort stocks into 10 decile portfolios based on their BTSS investor sentiment level. The table in Panel A reports the average values of the stock characteristics, such as size in millions, book-to-market ratio, and market share (proportion of total market capitalization represented by the firms in a given portfolio relative to the entire market in that month). Panel B reports the equal average values (and Panel C value-weighted average values) of the monthly 3 factor Fama–French returns, starting from the portfolio formation month (month zero) to 12 months into the future. The last row reports the difference in returns between high (positive sentiment) and low (negative sentiment) sentiment portfolios. T statistics are reported in parentheses.

Panel A Portfolio Summary Statistics
Sentiment Decile		Sentiment			Firm Size (M)			Book/Market			Market Share
1		−0.04			10,520.54			0.59			0.10
2		−0.01			8765.81			0.60			0.11
3		0.00			5903.95			0.64			0.10
4		0.01			5575.08			0.62			0.09
5		0.02			6183.12			0.57			0.09
6		0.03			7430.98			0.54			0.09
7		0.04			8539.84			0.51			0.10
8		0.05			8601.65			0.49			0.10
9		0.07			10,145.90			0.46			0.11
10		0.12			10,864.09			0.43			0.11
Panel B Equal-Weighted Monthly Abnormal Stock Returns in Sentiment Portfolios
Sentiment Decile	0	1	2	3	4	5	6	7	8	9	10	11	12
1	−4.92	−0.78	−1.03	−2.06	−1.19	−0.25	0.57	1.23	1.01	0.50	0.79	1.93	3.53
	(−17.29)	(−1.11)	(−1.63)	(−2.73)	(−0.76)	(−0.14)	(0.64)	(1.32)	(0.82)	(0.35)	(0.74)	(3.11)	(5.37)
2	−2.20	−0.69	−1.20	−1.57	−0.40	0.02	1.12	0.82	1.10	0.88	0.20	0.89	2.20
	(−5.01)	(−1.00)	(−1.14)	(−2.79)	(−0.49)	(0.01)	(1.46)	(1.12)	(1.15)	(0.62)	(0.20)	(1.74)	(4.03)
3	−0.51	−0.05	−0.42	−1.27	−0.89	−0.25	1.26	1.10	0.38	0.20	0.90	1.69	2.33
	(−4.63)	(−0.09)	(−0.63)	(−2.49)	(−1.19)	(−0.19)	(1.82)	(1.75)	(0.41)	(0.13)	(0.88)	(3.88)	(4.29)
4	0.74	−0.33	−0.29	−1.15	−1.01	−0.97	0.74	1.09	0.64	0.44	0.44	2.65	2.45
	(3.91)	(−0.60)	(−0.50)	(−1.85)	(−0.94)	(−0.88)	(1.22)	(1.53)	(0.61)	(0.28)	(0.41)	(6.14)	(3.85)
5	1.27	−0.45	−0.83	−1.33	−1.20	−0.59	0.70	1.18	0.57	0.61	0.75	1.77	2.67
	(4.72)	(−0.66)	(−0.99)	(−2.13)	(−1.28)	(−0.48)	(1.14)	(1.67)	(0.57)	(0.45)	(0.65)	(3.42)	(4.82)
6	1.58	−0.61	−0.48	−1.68	−0.98	−1.14	0.85	1.12	0.67	0.82	0.33	1.98	2.07
	(9.91)	(−0.90)	(−0.53)	(−2.22)	(−0.78)	(−0.88)	(1.64)	(1.63)	(0.62)	(0.51)	(0.32)	(3.73)	(3.09)
7	1.54	−0.32	−0.75	−1.45	−1.30	−0.40	1.00	1.18	0.38	0.61	0.69	1.67	2.31
	(18.07)	(−0.63)	(−0.73)	(−2.18)	(−1.60)	(−0.28)	(1.15)	(1.52)	(0.44)	(0.43)	(0.56)	(3.33)	(3.65)
8	2.40	−0.27	−1.19	−1.28	−0.96	−0.68	0.74	0.41	0.45	0.73	0.84	2.03	2.35
	(26.47)	(−0.51)	(−1.68)	(−1.86)	(−1.13)	(−0.56)	(1.11)	(0.51)	(0.42)	(0.49)	(0.77)	(3.48)	(3.78)
9	2.64	−0.85	−0.75	−1.62	−0.76	−0.65	0.73	0.48	0.23	0.68	0.70	1.63	2.03
	(13.23)	(−1.05)	(−1.06)	(−2.28)	(−0.81)	(−0.45)	(0.95)	(0.60)	(0.21)	(0.42)	(0.65)	(2.74)	(3.60)
10	4.08	−0.50	−0.56	−1.29	−0.85	−0.73	0.47	0.67	0.46	0.38	0.03	1.59	2.19
	(11.94)	(−0.61)	(−0.80)	(−1.97)	(−0.88)	(−0.57)	(0.72)	(0.99)	(0.39)	(0.24)	(0.03)	(1.96)	(3.14)
DIF(10−1)	9.00	0.28	0.47	0.77	0.34	−0.48	−0.11	−0.56	−0.55	−0.13	−0.76	−0.33	−1.34
	(59.26)	(0.58)	(1.30)	(1.48)	(0.53)	(−0.86)	(−0.22)	(−1.39)	(−2.66)	(−0.47)	(−5.44)	(−0.86)	(−3.77)
Panel C Value-Weighted Monthly Abnormal Stock Returns in Sentiment Portfolios
Sentiment Decile	0	1	2	3	4	5	6	7	8	9	10	11	12
1	−2.46	−0.25	−0.95	−1.72	−0.73	0.35	0.77	1.42	1.59	1.07	1.08	1.02	2.07
	(−8.16)	(−0.48)	(−1.74)	(−3.73)	(−0.57)	(0.27)	(1.19)	(2.51)	(3.36)	(1.25)	(1.60)	(2.11)	(6.62)
2	−0.50	0.76	0.04	−1.11	0.35	0.65	1.45	1.27	0.47	0.82	−0.26	1.91	0.95
	(−2.81)	(1.83)	(0.08)	(−2.24)	(0.38)	(0.66)	(2.45)	(2.65)	(0.74)	(0.95)	(−0.55)	(3.23)	(2.22)
3	−0.06	1.10	0.04	−0.91	−0.24	−0.16	1.60	1.76	0.46	0.73	0.96	1.44	2.02
	(−0.13)	(2.23)	(0.09)	(−1.80)	(−0.43)	(−0.21)	(2.32)	(3.91)	(0.83)	(0.79)	(1.18)	(5.38)	(6.68)
4	1.04	0.09	0.16	−0.66	−0.21	−0.99	0.50	1.17	1.10	1.54	0.62	2.30	1.26
	(6.23)	(0.15)	(0.26)	(−1.37)	(−0.17)	(−1.66)	(0.80)	(1.59)	(1.15)	(1.39)	(0.76)	(6.34)	(2.01)
5	0.70	−0.10	−0.26	−1.04	−0.33	−0.33	2.06	1.19	0.55	0.75	0.73	1.02	1.23
	(5.35)	(−0.18)	(−0.40)	(−1.88)	(−0.53)	(−0.45)	(4.32)	(2.20)	(0.72)	(0.78)	(1.18)	(1.92)	(3.17)
6	0.64	−0.05	0.54	−1.18	−0.47	−0.54	0.51	0.77	0.62	0.38	0.17	1.08	1.39
	(2.32)	(−0.09)	(1.09)	(−1.97)	(−0.47)	(−0.58)	(1.05)	(1.21)	(0.98)	(0.34)	(0.18)	(2.13)	(2.93)
7	−0.02	−0.15	−0.24	−0.72	−0.63	0.13	0.78	0.99	0.36	0.28	0.14	1.48	1.51
	(−0.20)	(−0.27)	(−0.30)	(−1.17)	(−0.75)	(0.11)	(1.38)	(1.84)	(0.51)	(0.22)	(0.17)	(3.54)	(2.50)
8	0.82	0.41	−0.62	−1.00	−0.46	−0.48	0.82	0.51	0.99	0.16	0.09	0.89	1.79
	(6.75)	(0.90)	(−1.09)	(−2.01)	(−0.66)	(−0.45)	(1.56)	(0.84)	(1.11)	(0.16)	(0.11)	(1.05)	(3.40)
9	1.68	−0.03	−0.08	−0.90	−0.34	−0.28	0.86	0.72	0.51	0.69	0.51	1.49	1.15
	(17.27)	(−0.08)	(−0.20)	(−1.59)	(−0.35)	(−0.28)	(1.53)	(0.91)	(0.86)	(0.57)	(0.76)	(2.65)	(1.47)
10	2.24	0.00	−0.44	−1.31	−0.39	−0.53	0.81	0.84	0.18	0.45	0.08	1.25	1.39
	(14.94)	(−0.01)	(−0.84)	(−2.45)	(−0.46)	(−0.55)	(1.79)	(1.21)	(0.16)	(0.39)	(0.13)	(1.80)	(2.10)
DIF (10−1)	4.70	0.24	0.52	0.41	0.34	−0.88	0.04	−0.58	−1.42	−0.62	−1.00	0.23	−0.69
	(23.81)	(0.72)	(2.02)	(1.75)	(0.67)	(−2.18)	(0.09)	(−2.74)	(−1.56)	(−1.70)	(−4.64)	(0.58)	(−1.30)

remains modest, with an R² of approximately 9.5%. This is consistent with the behavioral nature of investor sentiment and suggests that much of the variation in BTSS sentiment reflects investor beliefs not fully explained by observable firm characteristics or media tone. Rather than undermining our approach, this limited explanatory power reinforces the notion that BTSS captures a distinct, behaviorally driven signal. At the same time, we acknowledge that the unexplained variation may also reflect model limitations or omitted variables. To further investigate this, we use the specification from column 8 to decompose BTSS sentiment into two components: the predicted values reflect fundamentals-based sentiment, while the residuals represent “pure” sentiment not accounted for by firm-level fundamentals or news sentiment. In the next section, we examine how each component relates to future stock returns.

To further understand the sentiment measure, we examine its persistence by constructing a contingency graph. Each month we sort stocks based on their BTSS sentiment measure into 10 deciles. Figure 1 shows a heatmap of the initial decile assignment and subsequent decile assignment in month 1, represented as Pr(portfolio assignment in time t = 1|portfolio assignment in time t = 0). We can see that there is certain consistency in sentiment measure over time. Stocks in the highest sentiment decile have a 32% probability of remaining in that decile the following month, while those in the lowest decile show a 23% chance of persistence. This suggests that optimistic and pessimistic investor sentiment exhibits short-term inertia. On the other hand, it is apparent that there is still turnover in portfolio assignment, with 70–80% stocks changing their subsequent sentiment ranks. In the middle portfolios, the persistence of assignment seems weaker. Finally, 7% of past extreme negative sentiment stocks become extreme positive and vice versa. Please note that the exceptionally high persistence of portfolio 3 is because both 3 and 4 contain zero breakpoints (neutral sentiment), and simply by construction, all zero observations are assigned to portfolio 3.

To sum up, while we observe that the sentiment measure may persist over time, especially for optimistic sentiment, there seems to be significant variation and changes in sentiment over time. As an additional test of persistence, we construct an average monthly sentiment measure for the whole market and examine its autocorrelation properties. We find that it follows an autoregressive model of order 1 (y_(t) = 0.018 + 0.53 y_(t−1) + u_(t)), again pointing to certain autocorrelation of the sentiment.

4. Return Consequences of Sentiment

We now turn our focus toward the dynamic relationship between stock-level investor sentiment and returns, focusing on the idiosyncratic price effects of sentiment. Each month, we sort stocks into deciles based on their average BTSS sentiment and track portfolio returns using the Fama–French three-factor model, which adjusts for market, size, and value exposures. Table 3 Panel A summarizes firm characteristics across sentiment deciles. Firms in the most extreme deciles—those with very high or very low sentiment—tend to be larger and represent a greater share of total market capitalization. Notably, eight out of ten deciles exhibit positive average sentiment, reflecting the general optimism observed in investor discourse. Negative sentiment stocks tend to be value firms, while high sentiment stocks are more growth-oriented.

Panels B and C report equal- and value-weighted portfolio returns, respectively. In the concurrent month (month 0), firms in the highest sentiment decile earn strongly positive abnormal returns, while those in the lowest decile experience large negative returns. The spread between the top and bottom deciles reaches 9.00% for equal-weighted returns and remains statistically significant. The pattern is similar but less pronounced in value-weighted returns, indicating that sentiment has a stronger contemporaneous impact on smaller firms. Over the following six months, we observe continued return momentum, followed by a reversal beginning around month 7. We interpret the 6-month delay in the price correction as a sign of the market’s inefficiency that can be partially related to the sentiment’s relative persistence.

To validate the univariate portfolio findings, we conduct multivariate Fama-Macbeth regressions of stock-level excess returns on BTSS sentiment and a comprehensive set of control variables, with industry fixed effects. These include well-established return predictors from the asset pricing literature: market beta, firm size, book-to-market ratio, momentum (Jegadeesh & Titman, 1993), coskewness (Harvey & Siddique, 2000), idiosyncratic volatility (Ang et al., 2006), maximum return, analyst dispersion (Diether et al., 2002), earnings surprise, Amihud illiquidity, and abnormal dollar volume (see Bali et al., 2014 for discussion of return determinants).

As shown in

Table 4. Cross-sectional regressions of excess raw returns on BTSS sentiment and controls. This table reports the results of stock-level Fama and MacBeth (1973) regressions of monthly excess returns at various future horizons (from month 0 to month 12) on the BTSS investor sentiment measure and a comprehensive set of control variables. The controls include market beta, firm size, book-to-market ratio, 11-month momentum, coskewness, idiosyncratic volatility, maximum daily return in the prior month, analyst forecast dispersion, earnings surprise, Amihud illiquidity, and abnormal dollar volume. Regressions also include RavenPack sentiment metrics from news and press releases to isolate the incremental impact of social media sentiment. T statistics are based on Newey–West standard errors with 2 lags. * p < 0.1; ** p < 0.05; *** p < 0.01.

	exret0	exret1	exret2	exret3	exret4	exret5	exret6
BTSS sentiment	0.258 ***	−0.006	−0.013	−0.007	0.007	−0.011	0.011
	[7.472]	[−0.467]	[−1.185]	[−0.498]	[0.545]	[−0.700]	[0.883]
AES-news	0.025 ***	0.010	0.005	0.006	0.006	0.006	0.006
	[8.279]	[1.458]	[0.826]	[1.096]	[0.988]	[1.069]	[1.140]
AES-pr	−0.012 ***	−0.002	−0.000	−0.002	−0.002	−0.001	−0.004
	[−3.201]	[−0.653]	[−0.035]	[−0.381]	[−0.415]	[−0.109]	[−0.710]
Beta	−0.565	−0.057	−0.185	−0.182	−0.187	−0.239	0.006
	[−1.513]	[−0.233]	[−0.802]	[−0.719]	[−0.838]	[−1.184]	[0.024]
Size (lnme)	−0.278 **	−0.101	−0.101	−0.015	−0.095	−0.153	−0.091
	[−2.539]	[−1.598]	[−1.442]	[−0.192]	[−1.366]	[−1.651]	[−1.263]
Ln B/M	0.215	0.052	−0.026	0.061	0.128	0.023	0.124
	[1.432]	[0.242]	[−0.116]	[0.281]	[0.544]	[0.100]	[0.549]
Momentum	−0.027 ***	−0.009	−0.007	−0.010 **	−0.007	−0.009 *	−0.008
	[−4.194]	[−1.546]	[−1.298]	[−2.371]	[−1.633]	[−1.841]	[−1.726]
Coskew	0.438	0.169	−0.191	0.110	−0.134	−0.659	0.285
	[0.315]	[0.146]	[−0.226]	[0.106]	[−0.129]	[−0.515]	[0.221]
Ivol	−5.862 ***	−0.100	−0.208	−0.100	−0.174	−0.317	−0.333
	[−17.076]	[−0.415]	[−0.802]	[−0.450]	[−0.684]	[−1.138]	[−1.026]
Max	2.248 ***	−0.032	−0.004	−0.004	0.007	0.018	0.029
	[29.382]	[−0.809]	[−0.097]	[−0.079]	[0.179]	[0.490]	[0.588]
Disp	0.045	−0.112	0.075	0.014	0.071	0.100	−0.025
	[0.796]	[−0.959]	[0.444]	[0.086]	[0.342]	[0.531]	[−0.264]
Earnings surprise	87.177 ***	15.343	32.498	−3.909	−51.543	14.530	26.023
	[3.363]	[0.863]	[1.356]	[−0.156]	[−1.486]	[0.449]	[0.823]
Illiq	0.022 *	−0.016	−0.013	0.006	0.019	−0.015	−0.028
	[1.911]	[−1.170]	[−1.255]	[0.456]	[1.206]	[−1.106]	[−1.480]
Voldu	1.121 ***	0.064	0.034	0.095	0.060	0.076	0.037
	[8.721]	[0.832]	[0.653]	[1.383]	[0.923]	[1.143]	[0.778]
Constant	1.194	4.928 *	−5.699	0.558	1.805	2.160	3.053
	[0.437]	[1.844]	[−1.657]	[0.193]	[0.524]	[0.562]	[0.889]
Industry fixed effects	YES	YES	YES	YES	YES	YES	YES
N	48,404	46,562	44,283	41,939	39,868	37,675	35,408
R-sq	0.603	0.137	0.143	0.138	0.136	0.149	0.138
	exret7	exret8	exret9	exret10	exret11	exret12
BTSS sentiment	−0.028 **	−0.030 **	−0.005	0.003	0.016	−0.030 **
	[−2.205]	[−2.715]	[−0.319]	[0.142]	[1.140]	[−2.464]
AES-news	0.002	0.016 ***	0.001	−0.014 **	−0.003	−0.006
	[0.345]	[3.025]	[0.206]	[−2.225]	[−0.508]	[−1.453]
AES-pr	−0.001	−0.012 **	0.001	0.012 **	−0.001	0.002
	[−0.212]	[−2.944]	[0.407]	[2.406]	[−0.407]	[0.700]
Beta	0.077	−0.226	−0.144	0.113	0.089	0.339
	[0.314]	[−0.949]	[−0.555]	[0.377]	[0.306]	[1.286]
Size (lnme)	−0.120	−0.104	−0.037	−0.170 *	−0.127	−0.107
	[−1.178]	[−1.062]	[−0.335]	[−1.821]	[−1.370]	[−0.940]
Ln B/M	0.275	0.200	0.311	0.252	0.182	0.426
	[1.149]	[0.948]	[1.316]	[0.928]	[0.673]	[1.324]
Momentum	−0.005	−0.006 *	−0.006	−0.006	−0.009 *	−0.005
	[−1.126]	[−1.930]	[−1.744]	[−1.439]	[−2.171]	[−1.028]
Coskew	−0.283	−0.778	−0.440	0.325	0.110	2.069 *
	[−0.201]	[−0.691]	[−0.302]	[0.244]	[0.078]	[1.857]
Ivol	−0.379 *	−0.141	0.138	−0.566 **	−0.110	0.105
	[−1.763]	[−0.526]	[0.711]	[−2.467]	[−0.498]	[0.543]
Max	0.027	0.001	−0.009	0.106 ***	−0.053 **	−0.076
	[0.770]	[0.014]	[−0.201]	[3.347]	[−2.745]	[−1.404]
Disp	0.232	0.027	−0.374 *	−0.313	−0.105	−0.111
	[1.680]	[0.150]	[−2.042]	[−1.382]	[−0.913]	[−0.921]
Earnings surprise	−10.807	−8.390	2.995	18.801	5.414	−30.082
	[−0.535]	[−0.586]	[0.116]	[0.484]	[0.156]	[−1.639]
Illiq	−0.002	−0.020 **	−0.003	−0.002	−0.013	−0.014
	[−0.277]	[−2.244]	[−0.384]	[−0.130]	[−0.749]	[−1.634]
Voldu	0.026	0.043	0.004	−0.005	0.055	−0.075
	[0.411]	[0.720]	[0.050]	[−0.067]	[0.795]	[−1.645]
Constant	7.020	5.187	1.014	5.762	4.777 *	3.438
	[1.736]	[1.232]	[0.357]	[1.078]	[1.848]	[1.075]
Industry fixed effects	YES	YES	YES	YES	YES	YES
N	33,448	31,297	29,067	27,133	25,082	23,041
R-sq	0.135	0.139	0.131	0.137	0.140	0.133

, BTSS sentiment is strongly and significantly associated with concurrent returns (month 0), even after accounting for the full set of controls. Given that the standard deviation of BTSS sentiment is 4.358, the coefficient of 0.258 (t = 7.47) implies that a one standard deviation increase in sentiment is associated with a 1.12 percentage point rise in monthly excess return, underscoring the economically meaningful impact of social media sentiment on stock prices. Importantly, this contemporaneous effect is not persistent. Beginning around month 7, the coefficient on BTSS sentiment becomes negative and statistically significant, indicating a reversal in returns. This delayed correction is consistent with the behavioral interpretation that optimistic sentiment leads to short-term overpricing, which gradually unwinds as a more fundamental information is incorporated into prices.

Controlling for AES-news and AES-pr helps isolate the informational content of traditional media and press releases relative to social media sentiment. AES-news enters with a positive and highly significant coefficient, suggesting that news-based sentiment contributes meaningfully to contemporaneous returns. In contrast, AES-pr is negative and significant, indicating that markets may discount firm-issued communications, possibly perceiving them as less objective or strategically biased. Importantly, BTSS sentiment remains stronger, indicating that news sentiment complements the predictive power of social media sentiment. In later months, AES-news sentiment gains modest significance again, particularly during the return reversal period, as investors appear to shift focus back to fundamentals.

We see that the earnings surprise variable coefficient is highly significant and large in value, but our sentiment measure maintains its strong impact nevertheless. Interestingly, the momentum coefficient is negative and statistically significant in several horizons, particularly in the short term. This suggests that sentiment-driven price pressure may initially dominate but reverses sooner than standard momentum patterns would predict, consistent with investor overreaction followed by correction. The intercepts represent average risk-adjusted excess returns conditional on included predictors; while some appear economically large, they are generally not statistically significant and likely reflect sample-specific effects from the strong market environment of 2015–2017.

Overall, the results suggest that social media sentiment exerts a contemporaneous influence on stock prices and that this influence gradually unwinds, with the return reversal becoming statistically significant around month 7. This pattern reinforces the interpretation that high sentiment contributes to temporary mispricing, which corrects over time as fundamentals reassert themselves.

To further explore the role of investor attention in shaping the return impact of social media sentiment,

Table 5. Cross-sectional regressions of excess raw returns on BTSS sentiment, controlling for investor attention (Twitter spikes). This table reports the results of stock-level Fama and MacBeth (1973) regressions of monthly excess returns at various future horizons (from month 0 to month 12) on the BTSS investor sentiment measure, a proxy for investor attention based on abnormal Twitter activity (Tweet spike), BTSS and Twitter spike interaction term, and a comprehensive set of control variables. The controls include market beta, firm size (log market value), book-to-market ratio, 11-month momentum, coskewness, idiosyncratic volatility, maximum daily return in the prior month, analyst forecast dispersion, earnings surprise, Amihud illiquidity, and abnormal dollar volume. Regressions also include RavenPack sentiment metrics from news and press releases to isolate the incremental impact of social media sentiment. T statistics are based on Newey–West standard errors with 2 lags. * p < 0.1; ** p < 0.05; *** p < 0.01.

	exret0	exret1	exret2	exret3	exret4	exret5
BTSS sentiment	0.263 ***	−0.006	−0.014	−0.007	0.008	−0.013
	[7.719]	[−0.426]	[−1.243]	[−0.518]	[0.551]	[−0.825]
Tweet spike	−0.268	0.859	0.598	−0.048	−0.597	−0.568
	[−0.423]	[0.978]	[0.837]	[−0.125]	[−1.290]	[−0.505]
BTSS x Tweet spike	−0.241	−0.254 **	−0.089	0.097	0.384 **	0.013
	[−1.194]	[−2.107]	[−0.876]	[0.334]	[2.175]	[0.053]
AES-news	0.025 ***	0.010	0.005	0.006	0.006	0.007
	[8.172]	[1.475]	[0.804]	[1.066]	[0.977]	[1.103]
AES-pr	−0.012 ***	−0.002	−0.000	−0.002	−0.002	−0.001
	[−3.154]	[−0.701]	[−0.012]	[−0.356]	[−0.373]	[−0.140]
Beta	−0.566	−0.056	−0.187	−0.181	−0.188	−0.239
	[−1.514]	[−0.230]	[−0.811]	[−0.718]	[−0.847]	[−1.185]
Size (lnme)	−0.278 **	−0.104	−0.100	−0.015	−0.094	−0.153
	[−2.547]	[−1.661]	[−1.436]	[−0.195]	[−1.358]	[−1.657]
Ln B/M	0.215	0.047	−0.028	0.059	0.129	0.020
	[1.431]	[0.218]	[−0.128]	[0.268]	[0.545]	[0.087]
Momentum	−0.027 ***	−0.009	−0.007	−0.010 **	−0.007	−0.009 *
	[−4.193]	[−1.543]	[−1.296]	[−2.362]	[−1.632]	[−1.852]
Coskew	0.443	0.160	−0.202	0.119	−0.144	−0.669
	[0.317]	[0.139]	[−0.240]	[0.114]	[−0.138]	[−0.526]
Ivol	−5.857 ***	−0.106	−0.208	−0.104	−0.176	−0.320
	[−17.040]	[−0.441]	[−0.797]	[−0.469]	[−0.688]	[−1.147]
Max	2.247 ***	−0.030	−0.004	−0.002	0.007	0.019
	[29.300]	[−0.774]	[−0.086]	[−0.053]	[0.191]	[0.501]
Disp	0.044	−0.111	0.074	0.015	0.071	0.101
	[0.759]	[−0.945]	[0.442]	[0.091]	[0.342]	[0.534]
Earnings surprise	87.405 ***	15.335	32.794	−4.510	−52.327	14.368
	[3.339]	[0.857]	[1.367]	[−0.180]	[−1.505]	[0.439]
Illiq	0.021 *	−0.016	−0.013	0.006	0.019	−0.015
	[1.891]	[−1.173]	[−1.262]	[0.460]	[1.207]	[−1.100]
Voldu	1.121 ***	0.063	0.034	0.095	0.060	0.077
	[8.672]	[0.827]	[0.643]	[1.391]	[0.931]	[1.155]
Constant	1.251	5.008 *	−5.720	0.564	1.773	2.175
	[0.457]	[1.860]	[−1.665]	[0.196]	[0.515]	[0.565]
Industry fixed effects	YES	YES	YES	YES	YES	YES
N	48,404	46,562	44,283	41,939	39,868	37,675
R-sq	0.603	0.138	0.144	0.138	0.136	0.149
	exret6	exret7	exret8	exret9	exret10	exret11	exret12
BTSS sentiment	0.010	−0.027 **	−0.027 **	−0.005	0.003	0.013	−0.028 **
	[0.727]	[−2.148]	[−2.269]	[−0.293]	[0.168]	[0.855]	[−2.525]
Tweet spike	0.441	0.464	−1.590	−0.282	1.400	−0.995	−0.801
	[0.526]	[0.591]	[−1.235]	[−0.619]	[1.425]	[−0.964]	[−1.140]
BTSS x Tweet spike	−0.119	−0.048	0.129	0.035	−0.284	0.096	0.196
	[−1.342]	[−0.159]	[0.486]	[0.113]	[−1.270]	[0.598]	[1.696]
AES-news	0.006	0.002	0.016 ***	0.001	−0.014 **	−0.003	−0.006
	[1.129]	[0.358]	[3.111]	[0.203]	[−2.203]	[−0.536]	[−1.405]
AES-pr	−0.004	−0.001	−0.012 ***	0.001	0.012 **	−0.001	0.002
	[−0.709]	[−0.211]	[−2.952]	[0.404]	[2.347]	[−0.370]	[0.628]
Beta	0.006	0.078	−0.223	−0.145	0.113	0.088	0.335
	[0.025]	[0.316]	[−0.945]	[−0.557]	[0.377]	[0.304]	[1.275]
Size (lnme)	−0.094	−0.120	−0.108	−0.037	−0.171 *	−0.126	−0.110
	[−1.278]	[−1.184]	[−1.128]	[−0.334]	[−1.828]	[−1.380]	[−0.974]
Ln B/M	0.122	0.275	0.202	0.312	0.253	0.182	0.426
	[0.541]	[1.148]	[0.967]	[1.322]	[0.936]	[0.673]	[1.330]
Momentum	−0.008 *	−0.005	−0.006 *	−0.006	−0.006	−0.009 *	−0.005
	[−1.744]	[−1.107]	[−1.916]	[−1.761]	[−1.444]	[−2.176]	[−1.024]
Coskew	0.289	−0.291	−0.773	−0.439	0.312	0.087	2.040 *
	[0.222]	[−0.207]	[−0.695]	[−0.302]	[0.234]	[0.062]	[1.838]
Ivol	−0.338	−0.380 *	−0.134	0.141	−0.573 **	−0.110	0.104
	[−1.036]	[−1.772]	[−0.507]	[0.729]	[−2.500]	[−0.499]	[0.539]
Max	0.030	0.027	−0.001	−0.010	0.107 ***	−0.052 **	−0.077
	[0.602]	[0.781]	[−0.026]	[−0.220]	[3.373]	[−2.707]	[−1.403]
Disp	−0.023	0.233	0.028	−0.375 *	−0.311	−0.108	−0.112
	[−0.241]	[1.683]	[0.155]	[−2.033]	[−1.383]	[−0.929]	[−0.925]
Earnings surprise	26.253	−11.339	−8.984	2.444	19.733	4.988	−29.759
	[0.828]	[−0.567]	[−0.623]	[0.091]	[0.512]	[0.147]	[−1.625]
Illiq	−0.029	−0.003	−0.020 **	−0.003	−0.002	−0.013	−0.014
	[−1.503]	[−0.283]	[−2.228]	[−0.379]	[−0.131]	[−0.772]	[−1.600]
Voldu	0.040	0.025	0.042	0.002	−0.003	0.054	−0.075
	[0.835]	[0.392]	[0.703]	[0.034]	[−0.040]	[0.782]	[−1.648]
Constant	3.097	7.011	5.181	1.011	5.785	4.760 *	3.453
	[0.902]	[1.735]	[1.227]	[0.354]	[1.080]	[1.843]	[1.077]
Industry fixed effects	YES	YES	YES	YES	YES	YES	YES
N	35,408	33,448	31,297	29,067	27,133	25,082	23,041
R-sq	0.138	0.135	0.140	0.132	0.138	0.140	0.134

extends the baseline regression by incorporating a measure of abnormal Twitter activity and its interaction with BTSS sentiment. Specifically, we introduce a Tweet spike indicator that equals one if the number of Bloomberg-tracked tweets about a firm exceeds its firm-specific 95th percentile in a given month, and zero otherwise. This allows us to test whether the price impact of sentiment is amplified during periods of abnormally high investor engagement.

The results confirm a strong contemporaneous effect of BTSS sentiment on returns, consistent with Table 4. Notably, the interaction between sentiment and Tweet spikes becomes negative and statistically significant in months 7 and 8, suggesting that the return reversal is more pronounced when initial sentiment coincides with heightened investor attention. This finding is consistent with the view that attention amplifies the short-term price impact of sentiment-driven mispricing but also accelerates its correction as overreaction becomes more visible to the market.

4.1. Hard-to-Value Stocks and Sentiment

Next, we want to explore which stocks’ returns are mostly affected by sentiment. Baker and Wurgler (2007) argue that hard-to-value and difficult-to-arbitrage stocks, such as smaller, younger, volatile firms with low or no profitability are more likely to be subject to sentiment swings. They are less transparent, and their valuation is uncertain. Thus, irrational beliefs and biases may play a more significant role in investors’ transaction decisions, when more solid information sources are not available.

We address this theoretical prediction by dividing our sample into top 50% biggest and bottom 50% smallest firms, based on their end of year market value.

Table 6. Sentiment—return relationship in hard-to-value stocks—portfolio sorts. Each month we sort stocks into 10 decile portfolios based on their BTSS investor sentiment level. The table reports the difference in returns between high (positive sentiment) and low (negative sentiment) sentiment portfolios measured as either equal-weighted or value-weighted 3 factor Fama–French returns for different subsamples. Stocks are divided into the large and small firm category based on their logarithm of market value at a given time, where top 50% is assigned as large and bottom 50% as small. Similarly, stocks are divided into young and old category, based on their age, computed as the number of years since the firm’s first appearance on CRSP.

Sample	Diff Dec 10-Dec 1	0	1	2	3	4	5	6	7	8	9	10	11	12
Large firms	Equal weighted	6.76	−0.21	−0.12	−0.30	0.30	−0.32	−0.02	−0.88	−0.81	−0.81	−1.27	−0.10	−0.26
		(26.99)	(−0.76)	(−0.18)	(−0.64)	(0.70)	(−0.71)	(−0.04)	(−3.11)	(−1.79)	(−1.55)	(−7.33)	(−0.47)	(−0.57)
	Value weighted	4.58	−0.31	−0.28	−0.20	0.22	0.05	−0.58	−0.87	−1.50	−0.86	−1.38	0.19	−0.38
		(18.76)	(−1.45)	(−0.66)	(−0.55)	(0.57)	(0.11)	(−0.96)	(−4.36)	(−2.88)	(−1.87)	(−6.14)	(0.38)	(−0.52)
Small firms	Equal weighted	11.64	−0.63	−0.46	−0.22	0.34	−0.19	0.01	−0.88	−0.54	0.24	−0.59	−0.86	−1.87
		(22.93)	(−1.38)	(−0.70)	(−0.38)	(0.55)	(−0.39)	(0.05)	(−3.08)	(−2.53)	(0.70)	(−2.43)	(−4.16)	(−7.18)
	Value weighted	10.35	−0.72	−0.37	−0.09	0.57	−0.53	−0.27	−1.25	−0.60	0.23	−0.69	−0.74	−1.74
		(24.62)	(−1.69)	(−0.51)	(−0.14)	(1.14)	(−1.05)	(−0.57)	(−6.56)	(−2.49)	(0.60)	(−2.30)	(−3.10)	(−2.67)
Old firms	Equal weighted	9.06	−0.31	−0.06	−0.31	0.25	−0.12	−0.02	−0.68	−0.76	−0.72	−0.74	−0.31	−1.20
		(20.00)	(−0.86)	(−0.10)	(−0.78)	(0.46)	(−0.23)	(−0.04)	(−2.13)	(−2.37)	(−3.07)	(−1.76)	(−0.42)	(−3.51)
	Value weighted	4.48	−0.11	−0.25	−0.42	0.32	−0.08	−0.50	−0.67	−1.71	−0.84	−1.08	0.06	−0.56
		(11.14)	(−0.44)	(−0.56)	(−2.28)	(0.71)	(−0.25)	(−0.73)	(−2.93)	(−3.64)	(−2.49)	(−2.53)	(0.12)	(−1.54)
Young firms	Equal weighted	10.12	−0.40	0.32	0.07	0.07	−0.08	0.09	−0.84	−0.76	−0.79	−0.15	0.47	−0.89
		(21.92)	(−0.76)	(0.43)	(0.15)	(0.12)	(−0.13)	(0.10)	(−3.53)	(−2.58)	(−4.08)	(−0.29)	(0.61)	(−1.88)
	Value weighted	5.35	−0.97	0.03	0.04	0.05	0.69	−0.45	−0.57	−1.09	−1.10	−1.06	1.13	0.02
		(9.93)	(−1.82)	(0.05)	(0.06)	(0.10)	(0.97)	(−0.71)	(−1.27)	(−1.84)	(−2.02)	(−1.97)	(1.54)	(0.04)

summarizes the differences between the high and low sentiment monthly decile portfolios. We notice that, indeed, small firms have a 11.64% equally weighted contemporaneous return differential between high and low sentiment deciles compared to 6.76% in the large firm sample. The difference for value-weighted returns is even more striking, with more than double the same-month return differential for small firms.

We also assign firms the young and old subsamples, based on their age defined as the number of years since the firm’s first appearance on CRSP. It seems that younger firms tend to have a slightly stronger contemporaneous relationship with stock returns; however the difference is not big.

To ensure these findings are not driven by industry composition or other firm characteristics, we estimate stock-level Fama-Macbeth regressions similar to those in Table 4 but with added interactions between BTSS sentiment and firm-type indicators. Panel A of

Table 7. Sentiment—return relationship in hard-to-value stocks—cross-sectional regressions. This table reports the results of stock-level Fama and MacBeth (1973) regressions of monthly excess returns at various future horizons (from month 0 to month 12) on the BTSS investor sentiment measure, a comprehensive set of control variables, and interaction terms designed to capture variation in the sentiment–return relationship across firm types. Specifically, we examine how sentiment impacts returns among harder-to-value stocks. Panel A includes indicator variables for firms in the top two (Top20%Size) and bottom two (Bottom20%Size) size deciles based on monthly rankings of log market equity, along with their interactions with BTSS sentiment. Panel B includes indicator variables for firms in the youngest and oldest 20% of the firm age distribution each month (Top20%Age and Bottom20%Age) and interaction terms with BTSS to assess how sentiment effects vary with firm maturity. Control variables include market beta, firm size, book-to-market ratio, 11-month momentum, coskewness, idiosyncratic volatility, maximum daily return in the prior month, analyst forecast dispersion, earnings surprise, Amihud illiquidity, and abnormal dollar volume. Regressions also include RavenPack sentiment metrics from news and press releases to isolate the incremental impact of social media sentiment. T statistics are based on Newey–West standard errors with 2 lags. * p < 0.1; ** p < 0.05; *** p < 0.01.

Panel A SIZE

exret0

exret1

exret2

exret3

exret4

exret5

exret6

exret7

exret8

exret9

exret10

exret11

exret12

BTSS Sentiment

0.304 ***

−0.014

−0.010

0.002

0.011

−0.015

0.022

−0.008

−0.022

−0.001

0.009

0.037

−0.026 **

[8.476]

[−0.627]

[−0.744]

[0.080]

[0.669]

[−0.711]

[1.656]

[−0.468]

[−1.135]

[−0.034]

[0.334]

[1.498]

[−2.488]

Top20%Size * BTSS

−0.159 ***

0.019

0.013

−0.019

−0.006

0.014

−0.036 *

−0.032

−0.019

−0.015

−0.015

−0.048 *

−0.007

[−8.827]

[0.849]

[0.654]

[−0.940]

[−0.251]

[0.583]

[−1.791]

[−1.351]

[−0.686]

[−0.568]

[−0.644]

[−2.040]

[−0.224]

Bottom20%Size * BTSS

0.163 **

0.036

−0.051

0.007

−0.024

−0.062

−0.010

−0.059

−0.033

−0.022

−0.063

−0.054

−0.060

[2.267]

[0.524]

[−0.895]

[0.111]

[−0.481]

[−1.141]

[−0.224]

[−0.961]

[−0.506]

[−0.314]

[−0.651]

[−1.206]

[−1.275]

Top20%size

0.128

0.169

0.144

0.186

0.282

0.029

0.157

0.341

0.068

0.156

−0.018

0.037

−0.224

[0.849]

[0.759]

[0.617]

[0.968]

[1.515]

[0.130]

[0.695]

[1.238]

[0.229]

[0.641]

[−0.073]

[0.114]

[−0.849]

Bottom20%size

0.438 *

−0.142

−0.244

−0.324

−0.454

−0.130

−0.170

−0.712

−0.252

−1.100 *

−0.704

−0.720

−0.535

[2.008]

[−0.386]

[−0.541]

[−0.867]

[−1.118]

[−0.337]

[−0.330]

[−1.260]

[−0.548]

[−2.028]

[−1.493]

[−0.992]

[−1.035]

Controls: AES-news, AES-press release, Beta, Size (lnme), Ln B/M, Momentum, Coskew, Ivol, Max, Disp, Earnings surprise, Illiq, Voldu, 48 Fama–French industry indicators.

N

48,404

46,562

44,283

41,939

39,868

37,675

35,408

33,448

31,297

29,067

27,133

25,082

23,041

R-sq

0.607

0.142

0.148

0.143

0.141

0.154

0.143

0.140

0.145

0.137

0.144

0.145

0.138

Panel B AGE

exret0

exret1

exret2

exret3

exret4

exret5

exret6

exret7

exret8

exret9

exret10

exret11

exret12

BTSS Sentiment

0.275 ***

−0.016

−0.021 *

−0.004

0.013

−0.011

0.011

−0.027

−0.034 **

−0.022

0.002

0.010

−0.046 **

[7.528]

[−1.263]

[−1.729]

[−0.276]

[0.698]

[−0.634]

[0.463]

[−1.562]

[−2.864]

[−1.140]

[0.073]

[0.438]

[−2.754]

Top20%Age * BTSS

−0.087 ***

0.010

0.015

−0.035

−0.007

−0.015

−0.015

−0.018

0.013

0.034 *

−0.025

0.003

0.023

[−5.199]

[1.028]

[0.871]

[−1.613]

[−0.285]

[−0.933]

[−0.468]

[−0.707]

[0.572]

[1.946]

[−0.707]

[0.074]

[0.982]

Bottom20%Age * BTSS

0.041

0.067 *

0.027

0.021

−0.046

0.005

0.006

0.026

0.010

0.052

0.034

0.043 *

0.053

[1.267]

[1.914]

[0.623]

[0.443]

[−0.638]

[0.096]

[0.146]

[1.019]

[0.275]

[1.185]

[0.661]

[1.944]

[1.704]

Top20%Age

−0.026

0.071

0.112

0.129

0.148

0.107

0.101

0.090

0.037

−0.000

−0.040

0.087

−0.011

[−0.272]

[0.408]

[0.779]

[1.178]

[1.244]

[1.220]

[1.032]

[0.771]

[0.355]

[−0.003]

[−0.325]

[0.742]

[−0.073]

Bottom20%Age

−0.191

−0.227

−0.243

−0.154

−0.340

−0.179

−0.224

−0.116

−0.284

−0.077

−0.269

−0.198

−0.032

[−1.258]

[−1.011]

[−1.114]

[−0.671]

[−1.353]

[−0.703]

[−0.857]

[−0.389]

[−1.037]

[−0.266]

[−0.889]

[−0.584]

[−0.093]

Controls: AES-news, AES-press release, Beta, Size (lnme), Ln B/M, Momentum, Coskew, Ivol, Max, Disp, Earnings surprise, Illiq, Voldu, 48 Fama–French industry indicators.

N

48,404

46,562

44,283

41,939

39,868

37,675

35,408

33,448

31,297

29,067

27,133

25,082

23,041

R-sq

0.604

0.140

0.146

0.140

0.139

0.151

0.140

0.137

0.141

0.133

0.139

0.142

0.135

includes dummies for the top and bottom 20% of the monthly size distribution and their interactions with BTSS sentiment. We find that the sentiment–return relationship is significantly stronger among small firms, consistent with their greater exposure to mispricing due to limited information and arbitrage constraints.

Panel B presents analogous results using firm age. Here, we define young and old firms as those in the bottom and top 20% of the monthly age distribution. While the magnitude of the interactions is smaller, we find some evidence that sentiment has a greater short-term effect among younger firms. Overall, these results support the view that investor sentiment has stronger pricing implications in segments of the market characterized by higher valuation uncertainty and limited arbitrage.

4.2. Fundamental and Rational Sides of Sentiment

Investor sentiment expressed as a statement in social media is inherently multidimensional, reflecting both rational assessments of firm fundamentals and irrational beliefs. Which one plays a more important role in shaping stock returns? To answer this question, we follow Baker and Wurgler (2007) and decompose BTSS sentiment into two components: Fundamental Sentiment as “a belief about future cash flows and investment risks that are justified by facts at hand” and Pure Sentiment as “a belief about future cash flows and investment risks that is not justified by the facts at hand”.

Specifically, we estimate the following OLS regression from Table 2, Column 8:

BTSS_i,t = α + β₁·AES _i,t + β₂·lnPrice _i,t + β₃·Size _i,t + β₄·lnBook-to-Market _i,t + ε_i,t,

where BTSS_i,t is the monthly sentiment score for firm i at time t, AES_i,t is RavenPack’s Average Event Sentiment, and the remaining controls reflect firm fundamentals. The regression includes industry and time fixed effects, and standard errors are heteroscedasticity-consistent.

The predicted values from this regression define our Fundamental Sentiment, capturing the portion of social media sentiment explained by firm information and traditional media tone. The residuals ε_i,t are interpreted as Pure Sentiment, the component orthogonal to fundamentals and consistent with irrational investor beliefs. This decomposition allows us to identify the behavioral element of investor sentiment and examine its distinct impact on stock returns.

To further understand the behavior of BTSS and Pure Sentiment, we examine their co-movement with broader sentiment at the industry and market levels. We construct an industry-level sentiment index using market value-weighted averages of BTSS sentiment across firms within the same Fama–French 48 industry (Fama & French, n.d., retrieved 26 January 2020), excluding the firm in question. Similarly, we create a market-level sentiment index as the market value-weighted average BTSS sentiment across all firms outside the focal firm’s industry. We repeat both constructions using the Pure Sentiment measure.

Table 8. Panel A: Comovement of individual sentiment with industry and market. This table reports the correlations between stock-level BTSS sentiment and broader sentiment measures, including industry-level and market-level sentiment indices. Industry sentiment is computed using value-weighted averages within Fama–French 48 industries, excluding the stock. Market sentiment is computed as a value-weighted average across all firms, excluding those in the same industry. We report results for both raw BTSS and Pure Sentiment measures. Panel B: This table reports the results of Fama and MacBeth (1973) cross-sectional regressions of stock-level sentiment on contemporaneous industry-level and market-level sentiment. Column (1) uses BTSS as the dependent variable, while column (2) uses Pure Sentiment. Industry and market sentiment are computed using market value-weighted averages, excluding the stock and its industry, respectively. Coefficients represent time-series averages of the cross-sectional estimates. T statistics are based on Newey–West standard errors with 2 lags. * p < 0.1; ** p < 0.05; *** p < 0.01.

Panel A.
	BTSS	Pure Sentiment	Fundamental Sentiment	BTSS Sentiment Industry MV	Pure Sentiment Industry MV	BTSS Sentiment Market MV	Pure Sentiment Market MV
BTSS	1.000
Pure Sentiment	0.956	1.000
Fundamental Sentiment	0.294	0.000	1.000
BTSS industry MV	0.221	0.063	0.545	1.000
Pure Sent. industry MV	0.103	0.087	0.068	0.723	1.000
BTSS market MV	0.090	−0.058	0.492	0.313	−0.225	1.000
Pure Sent. market MV	0.201	−0.006	0.704	0.683	0.035	0.767	1.000
Panel B.
		1			2
		BTSS			Pure Sentiment
BTSS sentiment industry MV		−0.155 ***
		[−4.39]
BTSS sentiment market MV		−16.666 ***
		[−21.55]
Pure Sentiment industry MV					−0.229 ***
					[−6.27]
Pure Sentiment market MV					−20.621 ***
					[−23.12]
Constant		39.908 ***			0.142
		[18.05]			[0.91]
N		59,461			59,461
R-sq		0.050			0.047

Panel A presents the correlation coefficients between these measures. The correlation between BTSS and Pure Sentiment is very high (0.956). BTSS is only weakly correlated with Fundamental Sentiment (0.294), suggesting that social media sentiment captures distinct, non-fundamental information.

Stock-level BTSS sentiment shows modest positive correlations with both its industry-level (0.221) and market-level (0.090) counterparts, suggesting some degree of co-movement, but leaving large firm-specific variation. Pure Sentiment exhibits weaker or even negative correlations with aggregate sentiment: 0.063 with industry sentiment and −0.058 with market sentiment, reinforcing its idiosyncratic nature after filtering out shared sentiment components.

Panel B presents the results of Fama-Macbeth cross-sectional regressions, where the stock-level sentiment, measured by BTSS in column 1 and Pure Sentiment in column 2, are regressed on industry- and market-level sentiment, computed as above. The estimated coefficients on both industry and market sentiment are negative and highly significant across both specifications. These results imply that stock-level sentiment moves inversely with broader sentiment, both at the industry and market levels. One interpretation is that relative sentiment matters more than absolute sentiment: when industry or market sentiment is high, an individual stock may receive comparatively less investor attention or positive tone, resulting in a lower firm-specific sentiment score. This pattern is consistent with theories of attention allocation, where sentiment is a scarce resource distributed unevenly across firms.

The effect of market sentiment is economically large in both specifications, suggesting that aggregate market sentiment dominates firm-level sentiment variation more than industry sentiment does. This may reflect the influence of broad market narratives or macroeconomic signals on investor perceptions at the firm level.

We further test the pricing implications of each sentiment component using portfolio sorts.

Table 9. Decomposition of sentiment into Pure Sentiment and Fundamental Sentiment. We decompose the BTSS investor sentiment measure into the Fundamental and Pure Sentiment parts. We use the regression in Table 2, Column 8, of the BTSS sentiment measure on Ravenpack Average Event Sentiment of the news, firm’s price, size, and book-to-market ratio. The predicted values from this regression become our Fundamental Sentiment measure, and residuals from this regression are a measure of the Pure Sentiment. Then each month we sort stocks into 10 decile portfolios based on their Pure Sentiment (Panel A and B) and Fundamental Sentiment (Panel C and D). The tables in Panel A and C report the average values of the stock characteristics, such as size in millions, book-to-market ratio, and market share (proportion of total market capitalization represented by the firms in a given portfolio relative to the entire market in that month). Panels B and D report the equal average and value-weighted average values of the monthly 3 factor Fama–French returns, starting from the return concurrent to sentiment measure, until the return 12 months in the future. The last row reports the difference in returns between high (positive sentiment) and low (negative sentiment) sentiment portfolios. T statistics are reported in parentheses.

Panel A Portfolio Summary Statistics for Pure Sentiment
Sentiment Decile		Sentiment			Firm Size (M)			Book/Market			Market Share
1		−0.07			26,062.82			0.53			0.22
2		−0.04			10,075.60			0.57			0.16
3		−0.03			8871.35			0.52			0.10
4		−0.02			4578.96			0.59			0.07
5		−0.01			4326.95			0.62			0.07
6		0.00			4935.30			0.61			0.07
7		0.01			5484.96			0.58			0.07
8		0.02			6256.60			0.54			0.07
9		0.04			7287.77			0.49			0.08
10		0.09			8310.13			0.47			0.09
Panel B Monthly Abnormal Stock Returns in Pure Sentiment Portfolios
Equal-Weighted
Sentiment Decile	0	1	2	3	4	5	6	7	8	9	10	11	12
1	−4.30	−0.32	−0.94	−1.88	−0.77	−0.25	0.62	1.20	0.97	0.67	0.91	2.14	3.10
	(−11.88)	(−0.46)	(−1.68)	(−2.79)	(−0.50)	(−0.16)	(0.89)	(1.44)	(0.90)	(0.50)	(0.89)	(2.95)	(5.01)
2	−0.78	−0.44	−0.69	−1.29	−0.84	−0.64	0.73	0.65	0.29	0.26	0.54	1.74	2.65
	(−7.42)	(−0.62)	(−0.77)	(−2.13)	(−1.01)	(−0.47)	(1.02)	(0.78)	(0.24)	(0.15)	(0.43)	(3.26)	(4.08)
3	−1.27	−0.94	−0.54	−1.20	−0.68	−0.25	1.13	1.38	0.97	0.82	0.49	1.19	2.35
	(−6.48)	(−1.54)	(−0.68)	(−1.96)	(−0.77)	(−0.19)	(1.59)	(1.99)	(1.02)	(0.57)	(0.47)	(3.01)	(3.86)
4	0.16	−0.02	−0.46	−1.49	−1.10	−0.57	1.06	0.88	0.85	0.78	0.49	1.70	2.00
	(0.50)	(−0.03)	(−0.63)	(−2.69)	(−1.48)	(−0.45)	(1.27)	(1.38)	(1.03)	(0.60)	(0.46)	(4.23)	(3.65)
5	0.50	−0.46	−0.63	−1.32	−0.70	−0.47	0.56	0.88	0.56	0.76	0.72	1.92	2.24
	(1.50)	(−0.75)	(−0.81)	(−2.43)	(−0.69)	(−0.45)	(0.88)	(1.33)	(0.58)	(0.60)	(0.77)	(4.78)	(3.50)
6	0.82	−0.19	−0.92	−1.52	−1.18	−0.55	0.85	1.19	0.79	0.56	0.28	2.03	2.31
	(6.19)	(−0.37)	(−1.20)	(−2.30)	(−1.27)	(−0.43)	(1.52)	(1.71)	(0.74)	(0.38)	(0.31)	(3.85)	(3.93)
7	1.57	−0.39	−0.68	−1.64	−1.16	−0.88	0.87	1.08	0.25	0.62	0.61	1.74	2.47
	(15.82)	(−0.65)	(−0.88)	(−2.32)	(−1.28)	(−0.74)	(1.63)	(1.70)	(0.28)	(0.42)	(0.51)	(3.68)	(3.82)
8	2.32	−0.40	−0.67	−1.12	−1.07	−0.64	0.63	0.96	0.74	0.45	0.85	2.01	2.46
	(22.66)	(−0.80)	(−0.73)	(−1.69)	(−1.04)	(−0.46)	(0.68)	(1.18)	(0.70)	(0.31)	(0.75)	(3.43)	(4.51)
9	2.65	−0.83	−1.22	−1.87	−0.79	−0.62	1.00	0.72	0.21	0.49	0.92	1.90	2.12
	(10.36)	(−1.12)	(−1.83)	(−2.91)	(−0.77)	(−0.41)	(1.64)	(0.86)	(0.20)	(0.30)	(0.86)	(3.15)	(3.54)
10	4.19	−0.75	−0.65	−1.37	−0.78	−0.77	0.62	0.53	0.36	0.59	0.22	1.55	2.48
	(10.50)	(−0.93)	(−0.98)	(−2.02)	(−0.82)	(−0.54)	(0.94)	(0.76)	(0.29)	(0.38)	(0.21)	(1.93)	(3.43)
DIF(10−1)	8.50	−0.44	0.29	0.51	−0.02	−0.51	0.01	−0.67	−0.61	−0.07	−0.69	−0.60	−0.62
	(88.80)	(−0.99)	(0.76)	(0.96)	(−0.03)	(−2.45)	(0.04)	(−2.41)	(−2.10)	(−0.22)	(−4.47)	(−2.39)	(−2.86)
Value Weighted
Sentiment Decile	0	1	2	3	4	5	6	7	8	9	10	11	12
1	−0.65	0.67	−0.17	−1.19	−0.07	0.30	1.28	1.41	1.25	0.94	0.53	1.43	1.22
	(−4.96)	(1.40)	(−0.30)	(−2.54)	(−0.07)	(0.38)	(2.14)	(2.80)	(1.80)	(1.17)	(0.87)	(3.74)	(3.89)
2	0.61	0.73	0.16	−0.70	−0.09	0.09	0.98	1.41	0.16	1.07	0.59	1.69	1.45
	(3.88)	(1.59)	(0.30)	(−1.31)	(−0.11)	(0.09)	(1.79)	(2.24)	(0.35)	(1.06)	(0.76)	(4.22)	(3.75)
3	0.15	−0.42	0.17	−0.90	−0.57	−0.36	0.91	1.50	0.97	0.35	0.30	1.01	1.31
	(1.91)	(−0.69)	(0.33)	(−1.82)	(−0.59)	(−0.43)	(1.87)	(2.54)	(1.05)	(0.39)	(0.47)	(3.12)	(3.66)
4	−0.08	0.12	0.07	−1.11	−0.87	−0.94	0.96	0.73	0.33	0.46	−0.16	0.98	1.28
	(−0.66)	(0.31)	(0.13)	(−1.95)	(−1.08)	(−1.20)	(1.76)	(1.28)	(0.64)	(0.41)	(−0.16)	(1.50)	(2.72)
5	−0.24	−0.55	0.11	−1.05	0.03	−0.20	1.01	1.03	0.49	0.25	0.45	1.55	1.84
	(−1.04)	(−0.96)	(0.18)	(−1.56)	(0.03)	(−0.23)	(1.33)	(2.10)	(0.79)	(0.16)	(0.59)	(4.38)	(3.34)
6	0.05	−0.03	−0.68	−1.29	−1.27	0.06	0.58	0.57	0.57	0.64	0.21	1.03	1.39
	(0.28)	(−0.06)	(−0.85)	(−2.58)	(−1.52)	(0.04)	(0.91)	(0.84)	(0.68)	(0.78)	(0.25)	(1.49)	(3.54)
7	0.64	0.07	−0.51	−0.86	−0.34	−0.58	1.08	1.17	0.67	0.47	−0.05	1.66	1.46
	(7.28)	(0.18)	(−1.10)	(−1.69)	(−0.48)	(−0.60)	(2.32)	(2.12)	(0.92)	(0.43)	(−0.07)	(2.67)	(3.06)
8	1.21	−0.09	−0.28	−1.08	−0.43	−0.16	0.72	0.84	0.96	0.27	−0.01	1.68	1.67
	(13.06)	(−0.18)	(−0.54)	(−2.02)	(−0.48)	(−0.14)	(1.21)	(0.92)	(1.47)	(0.22)	(−0.01)	(3.37)	(2.10)
9	1.75	0.20	−0.45	−1.21	−0.37	−0.51	0.97	0.71	0.11	0.63	1.15	1.51	1.30
	(10.63)	(0.40)	(−1.34)	(−2.30)	(−0.36)	(−0.42)	(1.72)	(1.01)	(0.12)	(0.52)	(2.03)	(2.85)	(1.72)
10	2.41	−0.22	−0.41	−1.24	−0.44	−0.37	0.72	0.78	0.42	0.59	−0.07	1.20	1.63
	(20.94)	(−0.36)	(−0.61)	(−2.06)	(−0.51)	(−0.39)	(1.52)	(1.22)	(0.38)	(0.48)	(−0.10)	(1.50)	(2.42)
DIF(10−1)	3.07	−0.89	−0.25	−0.05	−0.37	−0.66	−0.56	−0.63	−0.83	−0.35	−0.60	−0.23	0.41
	(15.40)	(−2.73)	(−0.62)	(−0.18)	(−1.30)	(−2.38)	(−2.91)	(−2.11)	(−1.82)	(−0.75)	(−1.96)	(−0.44)	(0.86)
Panel C Portfolio Summary Statistics for Fundamental Sentiment
Sentiment Decile		Sentiment			Firm Size (M)			Book/Market			Market Share
1		0.01			333.98			0.84			0.01
2		0.02			1643.37			0.68			0.04
3		0.02			1149.03			0.57			0.02
4		0.03			1806.38			0.53			0.02
5		0.03			2784.56			0.47			0.03
6		0.03			4198.52			0.44			0.04
7		0.03			6555.27			0.40			0.06
8		0.03			9816.75			0.36			0.08
9		0.04			18,468.20			0.35			0.14
10		0.04			73,021.91			0.29			0.56
Panel D Monthly Abnormal Stock Returns in Pure Sentiment Portfolios
Equal Weighted
Fundamental Decile	0	1	2	3	4	5	6	7	8	9	10	11	12
1	−0.15	−1.16	−1.29	−1.95	−1.32	−0.77	1.00	0.88	0.63	0.34	0.95	1.86	2.91
	(−0.58)	(−1.75)	(−1.74)	(−3.51)	(−1.19)	(−0.52)	(1.19)	(1.11)	(0.53)	(0.21)	(0.77)	(4.55)	(4.16)
2	0.95	−0.46	−0.51	−1.36	−0.84	−0.66	0.60	0.91	0.69	0.33	0.51	1.94	2.88
	(4.59)	(−0.69)	(−0.60)	(−2.38)	(−0.96)	(−0.46)	(0.96)	(1.25)	(0.65)	(0.20)	(0.36)	(4.12)	(4.18)
3	0.80	−0.35	−0.26	−1.57	−1.28	−0.55	0.67	1.07	0.62	0.61	0.65	2.09	2.72
	(11.07)	(−0.62)	(−0.30)	(−2.54)	(−1.34)	(−0.40)	(0.88)	(1.35)	(0.49)	(0.39)	(0.57)	(3.03)	(4.13)
4	1.21	−0.36	−0.26	−1.49	−0.79	−0.29	0.93	1.27	0.54	0.92	0.61	1.99	2.74
	(15.34)	(−0.65)	(−0.37)	(−2.18)	(−0.91)	(−0.20)	(1.41)	(1.65)	(0.49)	(0.70)	(0.57)	(3.32)	(3.90)
5	0.50	−0.67	−0.91	−1.15	−0.54	−0.37	1.19	0.90	0.54	0.99	0.64	1.89	2.19
	(3.10)	(−0.96)	(−1.05)	(−1.55)	(−0.50)	(−0.31)	(1.78)	(1.26)	(0.59)	(0.68)	(0.61)	(4.16)	(3.38)
6	0.37	−0.46	−1.04	−1.40	−1.03	−0.93	0.40	0.75	0.55	0.69	0.61	1.64	2.24
	(2.43)	(−0.57)	(−1.23)	(−2.19)	(−1.00)	(−0.78)	(0.69)	(1.27)	(0.64)	(0.52)	(0.72)	(2.78)	(4.02)
7	0.57	−0.21	−0.84	−1.30	−0.86	−0.24	0.58	0.62	0.36	0.92	0.52	1.65	1.99
	(3.32)	(−0.37)	(−1.19)	(−2.05)	(−1.00)	(−0.18)	(0.76)	(0.81)	(0.39)	(0.71)	(0.62)	(3.41)	(4.63)
8	0.68	−0.20	−1.22	−1.23	−0.61	−0.55	0.90	0.90	0.78	1.11	0.47	1.74	1.41
	(13.69)	(−0.33)	(−1.53)	(−1.85)	(−0.57)	(−0.48)	(0.96)	(1.14)	(0.90)	(0.80)	(0.60)	(3.16)	(3.12)
9	0.71	−0.09	−0.56	−1.43	−0.95	−0.52	0.97	0.78	0.52	0.44	0.33	1.19	1.46
	(12.04)	(−0.17)	(−1.02)	(−2.07)	(−1.13)	(−0.48)	(1.65)	(1.16)	(0.53)	(0.35)	(0.41)	(2.82)	(4.30)
10	0.71	0.45	0.02	−0.76	−0.22	−0.42	0.91	1.10	0.79	0.54	0.09	1.08	0.98
	(5.71)	(0.95)	(0.05)	(−1.29)	(−0.31)	(−0.58)	(1.80)	(1.93)	(1.22)	(0.51)	(0.13)	(2.43)	(1.99)
DIF(10−1)	0.87	1.61	1.32	1.19	1.09	0.35	−0.09	0.22	0.15	0.19	−0.86	−0.78	−1.93
	(2.62)	(4.31)	(2.92)	(2.74)	(2.32)	(0.39)	(−0.13)	(0.67)	(0.24)	(0.29)	(−1.34)	(−2.39)	(−2.85)
Value Weighted
Fundamental Decile	0	1	2	3	4	5	6	7	8	9	10	11	12
1	−0.37	−1.53	−1.88	−2.62	−1.82	−1.08	1.33	1.14	0.38	0.58	0.84	2.15	3.30
	(−1.02)	(−1.87)	(−1.96)	(−3.31)	(−1.14)	(−0.53)	(1.20)	(1.06)	(0.24)	(0.27)	(0.52)	(4.51)	(4.04)
2	0.53	−0.13	0.04	−1.27	−0.41	−0.09	1.07	1.41	0.68	0.50	0.41	1.58	1.64
	(3.17)	(−0.18)	(0.06)	(−2.33)	(−0.36)	(−0.07)	(1.73)	(2.15)	(0.80)	(0.41)	(0.45)	(2.33)	(2.35)
3	0.50	−0.70	−0.57	−2.18	−1.25	−0.52	1.04	1.56	0.56	0.23	0.50	2.09	2.81
	(2.19)	(−0.97)	(−0.66)	(−3.17)	(−1.09)	(−0.32)	(1.16)	(1.72)	(0.40)	(0.15)	(0.48)	(2.11)	(3.59)
4	0.44	−0.86	−0.40	−1.26	−0.70	0.05	1.05	0.94	0.51	0.75	0.62	1.85	2.16
	(1.25)	(−1.32)	(−0.77)	(−1.73)	(−0.61)	(0.03)	(1.37)	(0.97)	(0.50)	(0.63)	(0.58)	(2.13)	(2.99)
5	−0.04	−1.03	−1.37	−1.54	−0.58	0.08	1.52	0.76	0.61	0.72	1.15	2.14	2.31
	(−0.21)	(−1.42)	(−1.46)	(−2.61)	(−0.43)	(0.06)	(2.50)	(1.04)	(0.66)	(0.51)	(1.18)	(4.27)	(3.37)
6	0.16	−0.56	−1.05	−1.82	−0.78	−0.55	1.01	1.01	0.31	0.41	0.56	1.99	2.89
	(1.57)	(−0.81)	(−1.63)	(−3.55)	(−0.61)	(−0.37)	(1.35)	(2.03)	(0.35)	(0.31)	(0.62)	(2.41)	(3.45)
7	−0.19	−0.65	−0.63	−1.28	−0.51	−0.12	0.32	1.17	0.69	0.93	0.71	1.61	2.30
	(−0.81)	(−1.01)	(−1.06)	(−2.49)	(−0.54)	(−0.10)	(0.57)	(1.64)	(0.88)	(0.84)	(1.00)	(3.45)	(6.01)
8	0.16	−0.22	−1.13	−1.90	−0.43	−0.21	1.22	0.87	0.64	1.21	0.53	1.86	1.67
	(2.76)	(−0.38)	(−1.71)	(−3.36)	(−0.36)	(−0.19)	(1.50)	(1.12)	(0.79)	(0.98)	(0.65)	(3.29)	(3.52)
9	0.25	−0.10	−0.33	−1.10	−0.54	−0.14	1.04	0.94	0.35	0.38	0.54	1.48	1.59
	(2.69)	(−0.21)	(−0.62)	(−2.04)	(−0.61)	(−0.13)	(2.14)	(1.53)	(0.39)	(0.34)	(0.74)	(3.62)	(4.67)
10	0.64	0.70	0.21	−0.69	−0.09	−0.12	0.98	1.18	0.75	0.74	0.19	1.13	1.05
	(4.82)	(1.59)	(0.42)	(−1.36)	(−0.12)	(−0.17)	(1.86)	(2.29)	(1.31)	(0.82)	(0.32)	(2.87)	(2.97)
DIF(10−1)	1.02	2.23	2.09	1.93	1.72	0.96	−0.35	0.04	0.37	0.16	−0.66	−1.02	−2.25
	(2.11)	(3.84)	(2.70)	(2.95)	(1.91)	(0.65)	(−0.34)	(0.06)	(0.34)	(0.12)	(−0.61)	(−3.41)	(−3.28)

presents the results of monthly decile sorts based on BTSS, Pure Sentiment, and Fundamental Sentiment, examining Fama–French three-factor adjusted returns. In Panel A, we observe that there is a large dispersion in the sentiment across deciles. Also, interestingly, large firms tend to dominate negative Pure Sentiment deciles. In contrast, Panel C shows that large firms tend to cluster in the most optimistic Fundamental Sentiment decile.

Panel B shows that Pure Sentiment exhibits a strong contemporaneous return spread of 8.5% between high and low deciles (equal-weighted), whereas Panel D shows a much smaller spread, roughly one-tenth the size, for Fundamental Sentiment. Moreover, consistent with its behavioral nature, Pure Sentiment reverses as early as month 5, while Fundamental Sentiment shows delayed and weaker reversals (months 11 and 12). This pattern supports the idea that only the irrational component of sentiment induces temporary mispricing corrected over time.

In summary, BTSS sentiment reflects both information-driven and behavioral elements. Investors react immediately to both: optimistic sentiment drives higher concurrent returns, while pessimism depresses them. However, only the irrational, Pure Sentiment component leads to subsequent reversals, as prices gradually adjust to reflect fundamentals. This suggests that Pure Sentiment contributes to temporary mispricing that can be exploited through timing-based strategies.

5. Conclusions

Motivated by the theoretical and empirical importance of sentiment in asset pricing, this paper contributes to the literature by analyzing a direct, stock-level investor sentiment measure using Bloomberg’s Twitter Sentiment Score (BTSS). While previous studies have often relied on indirect sentiment proxies or focused on aggregate market-level sentiment, and more recent work using social media has concentrated on short-term return predictability, our study fills an important gap by examining medium-term, sentiment-driven mispricing and subsequent return reversals.

We examine the determinants of BTSS and its predictive power for stock-level returns using U.S. equity data from 2015 to 2017. Our analysis shows that, while BTSS is partially driven by firm characteristics and traditional media tone, a substantial portion of its variation remains unexplained, consistent with the view that sentiment reflects investor beliefs not fully grounded in fundamentals. We document short-term persistence in BTSS.

In the cross-section of returns, we find that high BTSS sentiment predicts positive contemporaneous returns, followed by a gradual reversal over the subsequent months suggesting that overly optimistic sentiment leads to temporary mispricing that is later corrected. These effects are most pronounced in stocks that are hard to value, such as small and young firms, and when social media activity is abnormally high, suggesting that elevated attention amplifies short-term mispricing.

Further, we decompose BTSS into fundamentals-based and residual sentiment components and show that return predictability is concentrated in the residual, reinforcing the interpretation of BTSS as a behavioral signal.

Our findings offer practical relevance for investors seeking to exploit sentiment-driven mispricing through contrarian strategies and highlight the financial implications of social media discourse for corporate managers. More broadly, they underscore the rising importance of non-traditional data sources in investment analysis and risk management, particularly as investor sentiment becomes increasingly decentralized and digitally mediated.

Limitations and Future Research

While our findings provide new insights into the price impact of firm-level investor sentiment, several limitations warrant discussion. First, our sentiment measure is constructed from publicly available social media data, which may not fully capture sentiment across all investor types, particularly institutional investors or those trading in thinly covered firms. Second, the length of our sample period is relatively short, which may restrict our ability to detect long-term effects or account for structural changes in market behavior. Third, although we control for a wide range of firm characteristics and employ robust portfolio and regression methodologies, omitted variable bias remains a potential concern.

Our work opens the door to future research exploring the interaction between stock- and aggregate market-level sentiment, the role of sentiment volatility, and the use of sentiment-based strategies across different asset classes or global markets. With increasingly sophisticated data sources and tools for sentiment analysis, the financial community is well positioned to deepen our understanding of behavioral forces in asset pricing.