FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media

arXiv:2605.09469 · cs.CL · Submitted 2026-05-10 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media".

Jane: Here's the gist of the paper. The gist: Emojis alone can achieve F1 ≈ 0.75, capturing most of the predictive signal of the combined models while being substantially more computationally efficient How it works The study explores whether emojis alone can serve as reliable proxies for financial sentiment and compares their performance with traditional text-based analysis.

Tom: The research question posed is What is the role of emojis in financial sentiment analysis, with respect to performance, computational efficiency, data requirements, patterns of use; and what are the differences between financial and generic use.

Jane: This investigation seeks to answer four specific sub-questions regarding public sentiment proxy (Q1), computational speed comparison (Q2), data needs for training models (Q3), and the significance of emoji usage differences between financial and non-financial social media contexts (Q4).

Tom: Data and Methodology The researchers utilized a balanced dataset of about 528,000 emoji-containing StockTwits posts to conduct their experiments.

Paper summary: Tom: So, what they found is that while combining text and emojis gets you a better score—an F1 score of approximately zero point eight eight—the emoji-only models are still pretty useful, hitting an F1 of about zero point seven five <ref:2605.09469#pg1>.

Jane: That’s interesting because it shows that even when you strip away all the text, the emojis alone capture a good chunk of that predictive signal, which is what they call substantial sentiment information in financial communication <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Lu: The paper points out that even though the emoji-only model has a lower accuracy, it needs much less data to train effectively, which is a big deal when you think about training models <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Meng: That’s what I care about—if we can get competitive accuracy with significantly less data, that opens up possibilities for deployment where resources are tight <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Lalam: They also found that certain specific emojis and emoji pairs have a really strong predictive power, showing over ninety percent accuracy when predicting whether a trend will be bullish or bearish <ref:2605.09469#pg1>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Tom: And the paper mentions that specific emojis, like the rocket emoji, show a correlation with things like Bitcoin and Nasdaq one hundred index returns <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Jane: That’s a concrete piece of evidence that connects the symbols directly to actual market movements, which is why the authors think treating emojis as standalone indicators is important <ref:2605.09469#pg2>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Lu: They also showed that simple models like logistic regression can achieve accuracy comparable to more complex transformer models when you’re only using emoji-only data <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Meng: So, the implication here is that we might not always need these huge, heavy transformer architectures to get good results if we just focus on these compact emoji indicators <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Lalam: The paper also highlights that there are big statistical differences in how financial and general social media contexts use emojis, which suggests we need models that are tailored specifically to financial talk <ref:2605.09469#pg2>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Conclusion: Tom: So, looking at "FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media," the authors are essentially arguing that emojis aren't just noise; they can be a distinct and valuable signal when analyzing financial chatter on StockTwits.

Jane: The core of the paper is showing that these symbols, when analyzed alone, can provide a more intuitive, faster, and language-independent tool for financial analysis <ref:2605.09469#pg2>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Lu: The study confirms that these emojis encode actual sentiment information, even if the emoji-only model isn't the absolute most accurate one when you have the text mixed in <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Meng: From a deployment standpoint, it means we can build very efficient systems that don't need massive datasets to train them, which is crucial if you want something running on real traffic <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Lalam: The paper suggests that this approach is transparent because the signals are language-agnostic and computationally lightweight, which makes it practical for things like high-frequency trading <ref:2605.09469#pg3>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Tom: So, the main point is that we should be paying attention to these emojis not just as decoration, but as compact indicators of investor feeling <ref:2605.09469#pg1>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

Jane: It really shifts the focus from just reading the words to understanding the non-textual signals that investors are sending through their social media posts <ref:2605.09469#pg2>—FinMoji: A Framework for Emoji-driven Sentiment Analysis in Financial Social Media.

AHMED MAHROUS, JENS SCHNEIDER, ROBERTO DI PIETRO

King Abdullah University of Science and Technology (KAUST) · Hamad Bin Khalifa University

cs.CL

Submitted: 2026-05-10

Updated: 2026-05-10

Journal ref: ACM Transactions on the Web 20(3), Article 35, pp. 1-30 (2026)

DOI: 10.1145/3829331

Code: https://github.com/AhmedMahrous00/finmoji_replication

Project page: https://www.similarweb.com/website/stocktwits.com/#geography

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: The gist: Emojis alone can achieve F1 ≈ 0.75, capturing most of the predictive signal of the combined models while being substantially more computationally efficient How it works The study explores

Key concepts

Emoji-only Model: This model analyzes financial posts using only the emojis present and ignores the surrounding text. The research found this approach is computationally very efficient and surprisingly effective at capturing sentiment signals, often matching performance achieved by more complex models.
Text-Emoji Combined Model: This method analyzes posts that include both written words and emojis together. This combination yielded the highest accuracy (F1 of 0.88) in the study, indicating that text provides valuable context alongside the visual cues from emojis.
Computational Efficiency: The study highlights how emoji-only models are significantly faster and require less data to train than text-based models. This efficiency is crucial for time-sensitive applications like high-frequency trading where speed is a major advantage.
Domain-Specific Analysis: The research found that the patterns of emoji use differ greatly between financial social media and general social media. This means sentiment analysis models must be tailored specifically to the unique language and context of financial discussions for better accuracy.

Terminology

Accuracy — 1× in this paper · explained in 1 episode
Transformer — 1× in this paper · explained in 2 episodes
Transformer Model — 1× in this paper · explained in 2 episodes
Transformer Models — 1× in this paper · explained in 2 episodes

Summary

The gist: Emojis alone can achieve F1 ≈ 0.75, capturing most of the predictive signal of the combined models while being substantially more computationally efficient

How it works

The study explores whether emojis alone can serve as reliable proxies for financial sentiment and compares their performance with traditional text-based analysis. The research question posed is What is the role of emojis in financial sentiment analysis, with respect to performance, computational efficiency, data requirements, patterns of use; and what are the differences between financial and generic use. This investigation seeks to answer four specific sub-questions regarding public sentiment proxy (Q1), computational speed comparison (Q2), data needs for training models (Q3), and the significance of emoji usage differences between financial and non-financial social media contexts (Q4).

Data and Methodology

The researchers utilized a balanced dataset of about 528,000 emoji-containing StockTwits posts to conduct their experiments. This dataset was carefully curated, as the initial 18.5 million posts were narrowed down to a final balanced sample of 528,000 posts evenly split between bullish and bearish sentiment. The methodology involved testing three data variants: Text with emojis version, Emoji only version where all text is removed, and Text only version where all emojis are removed. Pre-processing steps included removing cashtags, hashtags, and hyperlinks, lowercasing all text, and creating these three distinct versions of the data.

Model Evaluation and Performance

The study compared several machine learning models across the different data variants to assess their performance. The best text+emoji model achieved an F1 score of 0.88, while the emoji-only variant reached F1 ≈ 0.75. Logistic Regression performed on par with more complex transformer models when analyzing emoji-only data, indicating that emojis, as compact sentiment indicators, reduce the need for sophisticated deep learning architectures in financial sentiment analysis. Furthermore, certain emojis and emoji pairs exhibit strong predictive power for market sentiments, demonstrating over 90% accuracy in predicting bullish or bearish trends.

Efficiency and Contextual Insights

A key finding is the efficiency of emoji-only models, which can achieve accuracy comparable to text-based models but at substantially lower data requirements. With as little as 1,000 posts used for training can be sufficient to achieve competitive accuracy, especially if logistic regression is used. The computational cost is significantly reduced, which is a useful feature where time is at premium, like in High Frequency Trading. Additionally, the research reveals extreme statistical differences in emoji usage between financial and general social media contexts, stressing the need for domainspecific sentiment analysis models.

Key Findings and Implications

The findings suggest that emoji-based sentiment classification is not only effective but also computationally efficient, requiring significantly less data and processing power compared to text-based models. The correlation between the Rocket Index containing the emoji—moves positively with Bitcoin and Nasdaq 100 index returns, and negatively with the CBOE Volatility Index (VIX) index. This pattern confirms intuition: rocket emojis typically express enthusiasm and confidence, which rise during market upswings and decline when volatility or fear (VIX) increases. The study concludes that emoji-based sentiment classification is not only effective but also computationally efficient, requiring significantly less data and processing power compared to text-based models.

Reproducibility and Future Directions

The code and data needed to replicate the results are publicly available at https://github.com/AhmedMahrous00/finmoji replication. Future work could explore the creation of emoji-based financial sentiment indices, moving beyond binary sentiment classification to multinomial analysis to capture a broader spectrum of emotions relevant to financial decision-making. A formal study to establish that emojis in financial social media are culturally unbiased and thus can be used to help NLP models generalize across different languages is also a very interesting research direction. The research encourages the creation of specialized models that take into consideration the unique lexicon, including emojis, used in financial social media. The correlations reported in Table 11 serve as a pilot demonstration that emojibased sentiment signals align with market dynamics. This interpretability strengthens the practical case for emoji-only approaches: they are transparent, language-agnostic, and computationally lightweight, yet remain relevant to real investor behavior. The research demonstrates that emojis alone can achieve considerable accuracy in sentiment classification. This efficiency makes them particularly attractive for time-sensitive financial applications such as high-frequency trading, where even microseconds matter. The emoji sentiment lexicon developed in this paper also provides insights into why emojis achieve such efficiency—certain emojis and emoji pairs alone can predict sentiment with over 90% accuracy. Another key insight from our analysis is that simpler models, such as logistic regression, achieve comparable accuracy to more complex transformerbased models when trained on emoji-only data. The comparative analysis of emoji usage between financial and general social media contexts confirmed that financial platforms exhibit distinct emoji patterns, underscoring the necessity of domain-specific sentiment analysis models. The code and data needed to replicate the results are publicly available at https://github.com/AhmedMahrous00/finmoji replication. The research reported in this publication was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI - under award number 5940. This study has several significant implications for financial sentiment analysis. The analysis was confined to data sourced from StockTwits, which, while comprehensive and relevant, does not necessarily represent emoji usage across other platforms where financial discussions occur. This demographic concentration may not accurately reflect global financial social media, where cultural differences significantly influence emoji usage. The sentiment analysis conducted was binary, classifying posts as either bullish or bearish. This approach may oversimplify the complex emotions and sentiments expressed in financial discussions, such as uncertainty, caution, and speculation. The emoji sentiment lexicon developed in this paper also provides insights into why emojis achieve such efficiency—certain emojis and emoji pairs alone can predict sentiment with over 90% accuracy. Another key insight from our analysis is that simpler models, such as logistic regression, achieve comparable accuracy to more complex transformerbased models when trained on emoji-only data. The comparative analysis of emoji usage between financial and general social media contexts confirmed that financial platforms exhibit distinct emoji patterns, underscoring the necessity of domain-specific sentiment analysis models. The code and data needed to replicate the results are publicly available at https://github.com/AhmedMahrous00/finmoji replication. The research reported in this publication was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI - under award number 5940. This study has several significant implications for financial sentiment analysis. The analysis was confined to data sourced from StockTwits, which, while comprehensive and relevant, does not necessarily represent emoji usage across other platforms where financial discussions occur. This demographic concentration may not accurately reflect global financial social media, where cultural differences significantly influence emoji usage.

Improvements for AI systems

Bold header: Low-Resource Emoji Sentiment Classification

This system can achieve high accuracy with fewer training samples because models trained on just 1,000 posts can achieve relatively high accuracy when using emojis. This makes sentiment analysis feasible in data-scarce contexts where traditional text-based models require significantly larger datasets.

Bold header: Computational Efficiency for High-Frequency Trading

The improved system can offer a significant competitive advantage in time-sensitive applications because emoji-only models achieve F1 ≈ 0.75, lower than text–emoji combined models (F1 ≈ 0.88) but with far lower computational cost—a useful feature where time is at premium, like in High Frequency Trading.

Bold header: Domain-Specific and Cross-Domain Modeling

The system can be tailored to specific financial contexts by addressing the extreme statistical differences in emoji usage between financial and general social media contexts, thereby ensuring that domain-specific sentiment analysis models are used for high accuracy.