1

Financial Market Sentiment Analysis Using LLM

and RAG
Vivaan Echambadi Issaquah High School, Issaquah, USA
vivaanechambadi@[Link]

Abstract—This research paper investigates the methodologies alone explains only about 1% of the variance in stock price
and applications of market sentiment analysis in financial trading movements, indicating the influence of other factors such as
and investment strategies. Leveraging fine-tuned language models market conditions, sector-specific trends, and macroeconomic
like FinBERT within a Retrieval-Augmented Generation (RAG)
pipeline, the study examines the relationship between public indicators.
sentiment and short-term stock price movements. Sentiment This paper aims to assess whether sentiment can serve
scores derived from financial news articles are aligned with as a reliable indicator for short-term stock price movements
historical stock price data to build predictive models, providing and how it can enhance trading strategies. The research
actionable insights into market behavior. underscores the value of integrating sentiment analysis with
The results show that sentiment analysis, while statistically
significant, has a limited explanatory power in predicting next- other financial indicators, offering a comprehensive approach
day stock price movements, with an R2 value of 0.010. Negative to understanding market dynamics. For the purposes of this
sentiment exhibits a stronger immediate effect compared to study, short-term stock price change is defined as the price
positive sentiment. While sentiment alone cannot accurately movement from the market close on the day of the news
predict stock prices, its integration with other financial indicators publication to the market close on the following trading day.
enhances decision-making processes. Future considerations in-
clude incorporating real-time data processing and refining models This timeframe was chosen due to the availability of data and
to ensure greater robustness in practical applications. to capture immediate investor reactions. Future considerations
include expanding data sources, incorporating real-time data
Index Terms—Market Sentiment Analysis, Large Language
Models, Retrieval-Augmented Generation, Financial Trading, processing, and addressing challenges such as model bias and
Investment Strategies, Natural Language Processing, Machine data privacy to improve the robustness of sentiment analysis
Learning, Sentiment Scoring, Stock Market Predictions systems.

I. I NTRODUCTION II. R ELATED W ORK

The stock market is a complex, non-linear, and time-variant Market sentiment analysis has gained prominence due to
system heavily influenced by public sentiment. Understanding its critical role in predicting stock market movements, with
the sentiment behind market movements can offer invaluable the integration of machine learning (ML) and natural lan-
insights for investors, hedge funds, and financial analysts. guage processing (NLP) approaches serving as the backbone
Market sentiment analysis involves using machine learning of these efforts. Traditional approaches relied on sentiment
techniques to interpret and quantify the emotions and opinions lexicons and task-specific NLP models, often using supervised
expressed in textual data, such as news articles and financial classification techniques to categorize sentiment as positive,
reports. negative, or neutral. These methods demonstrated limitations
Recent advancements in machine learning and natural lan- in scalability and the ability to process complex, domain-
guage processing (NLP) have enabled the use of large lan- specific text data [1].
guage models (LLMs) to perform sentiment analysis more Recent advancements in AI have seen the application of
effectively. Specifically, fine-tuned models like FinBERT are large language models (LLMs) such as FinBERT, which
well-suited for financial sentiment analysis due to their ability leverage vast pre-trained knowledge bases, enabling robust
to interpret domain-specific language. These models, com- sentiment classification with minimal fine-tuning. Studies
bined with a Retrieval-Augmented Generation (RAG) pipeline, have highlighted FinBERT’s effectiveness in financial con-
provide a robust framework to retrieve relevant financial texts, where domain-specific language plays a significant role.
news articles, process them for sentiment classification, and Compared to traditional ML techniques like support vector
correlate sentiment scores with historical stock prices. machines (SVMs) and random forests, LLMs have shown
The proposed methodology includes extensive data prepro- superior adaptability to context-rich financial data, including
cessing, such as vector embedding generation using Sentence- news articles and social media posts [2], [3]. This adaptability
BERT, to ensure the quality of the dataset. Sentiment scores allows LLMs to interpret investor sentiment more effectively,
from FinBERT are used to predict next-day stock price move- even when faced with nuances in financial jargon or market-
ments, with findings showing that negative sentiment has a specific terminology.
stronger immediate effect on prices compared to positive sen- The Retrieval-Augmented Generation (RAG) pipeline fur-
timent. However, the regression analysis reveals that sentiment ther advances sentiment analysis by integrating vector retrieval
 2

methods with generative LLMs. This approach facilitates to stock tickers using entity recognition techniques, focusing
precise information retrieval from vast financial text corpora, on headlines and key company mentions within the article text.
such as the Surge AI Stock Sentiment Dataset, before passing Table I provides a summary of the key variables in the
the retrieved context to LLMs for sentiment classification. dataset, their respective sources, and a brief description of
This hybrid pipeline ensures that generative responses are each.
contextually grounded, thereby enhancing prediction accuracy
and reliability. Research underscores that combining sentiment B. Text Preprocessing
scores derived from RAG pipelines with historical stock data
can yield actionable insights into market trends, as demon- Preprocessing of the collected news articles was performed
strated in both developed and developing markets [2], [3]. to ensure the textual data was in a format suitable for analysis.
The interplay between sentiment and market behavior has Standard preprocessing techniques were applied, including to-
been extensively studied. Empirical evidence suggests that kenization, removal of stop words, normalization, and lemma-
negative sentiment has a stronger immediate impact on stock tization. Tokenization splits the text into individual words,
prices than positive sentiment, aligning with behavioral finance while stop words were removed to eliminate non-informative
theories. For instance, research in Ghana’s stock market has words. Normalization ensured consistency in terminology
demonstrated that combining multiple data sources—news, (e.g., “AI” and “artificial intelligence”), and lemmatization
social media, and search trends—significantly improves pre- reduced words to their base forms. These steps collectively
diction accuracy, highlighting the importance of diversified reduced noise in the data and prepared it for feature extraction.
data inputs [2]. Similarly, deep learning methods such as Additionally, to prepare the data for RAG integration, the
artificial neural networks (ANNs) and recurrent neural net- textual content was converted into vector embeddings for
works (RNNs) have been shown to outperform traditional ML efficient retrieval.
approaches when applied to sentiment-laden financial datasets, The collected news headlines were passed through a series
particularly in capturing long-term dependencies between sen- of preprocessing steps to ensure consistency and remove
timent and market movements [3]. noise. These steps included tokenization, lemmatization, and
Finally, hybrid approaches that blend sentiment analysis removal of stop words. After preprocessing, each headline
with technical indicators (e.g., stock volume, moving averages) was transformed into a dense vector representation using the
have emerged as powerful tools for prediction. By combining Sentence-BERT model. These vectors were then stored in a
these data modalities, researchers have achieved enhanced vector database for efficient retrieval during the sentiment
prediction accuracy and robustness, as demonstrated in studies analysis process.
utilizing data from global indices and regional markets. This
research builds upon such work, leveraging the latest advance- C. Feature Representation
ments in RAG and LLM-based sentiment analysis to offer To capture the semantic meaning of the news articles, vector
insights into the nuanced relationship between sentiment and embeddings were generated using a pre-trained Sentence-
stock price movements. BERT model. Each news headline was converted into a
dense vector representation, enabling the use of semantic
III. M ETHODOLOGY similarity in downstream tasks. These embeddings were stored
A. Data Collection and Preparation in a vector database (using FAISS) for efficient retrieval.
Additionally, bag-of-words (BoW) and TF-IDF representations
The first step in this study involves collecting and prepro- were employed to enhance the retrieval process by structuring
cessing financial news data alongside historical stock price financial news articles in a way that allows the language model
information. Using a web scraper implemented in Python, to extract and process relevant sentiment-related terms more
relevant financial news articles were gathered from trusted effectively. This approach ensured that key financial terms
sources, ensuring the inclusion of timestamps for accurate were preserved and contextualized before being analyzed by
correlation with stock price movements. The scraper focused the sentiment classification model.
on high-impact articles associated with companies in the tech
sector of the S&P 500. Financial news articles were primarily
scraped through Google News API ensuring coverage from D. Retrieval-Augmented Generation (RAG) Pipeline
multiple reputable financial websites and sources. A total of The Retrieval-Augmented Generation (RAG) pipeline is
34,000 articles covering multiple years were collected for central to this research, enabling the model to retrieve contex-
analysis. tually relevant financial news articles and enhance sentiment
Moreover, historical stock price data was collected from classification accuracy. The pipeline begins with a query about
2021 to 2024 using the Yahoo Finance API, covering top 15 stock performance, processed by a retriever that searches for
major technology sector stocks in the S&P 500 such as Apple relevant documents in the vector store. Fig. 1 illustrates the
(AAPL), Microsoft (MSFT), Nvidia (NVDA) and others. complete workflow, from raw data inputs to the final sentiment
Social sentiment data was initially explored from platforms prediction output, highlighting the key stages of data retrieval,
such as StockTwits. However, due to data limitations and language model processing, and sentiment analysis.
inconsistent labeling, it was ultimately excluded from the final Once the relevant documents are retrieved, they are passed
analysis. To ensure relevance, each news article was mapped to a pre-trained large language model (FinBERT) configured
 3

TABLE I
S UMMARY OF DATA S OURCES AND VARIABLES

Variable Source Description

Ticker Yahoo Finance API Unique identifier for the stock or security being analyzed.
Title Financial News Articles The headline of the financial news article, gathered from trusted sources.
Date Financial News Articles The date the news article was published, used for time-series alignment.
Closing Price Yahoo Finance API The stock’s closing price on the same day the article was published.
Previous Day Price Yahoo Finance API The stock’s price at the close of the previous trading day.
Next Day Price Yahoo Finance API The stock’s price at the close of the next trading day.
Cleaned Title Preprocessing Step Processed and tokenized version of the news title, cleaned for sentiment analysis.
Sentiment Label FinBERT Model The sentiment classification of the news article as Positive, Negative, or Neutral.
Sentiment Score FinBERT Model Numerical sentiment score generated by the FinBERT model.

to analyze sentiment and provide predictions. The sentiment This alignment enabled the creation of a dataset that combined
output is then used to predict next-day stock price move- financial sentiment with price movements, serving as the basis
ments. The FinBERT model assigns sentiment labels based on for predictive modeling.
a confidence-weighted classification mechanism, categorizing
each article as positive, negative, or neutral using pre-trained
weights on financial text data. G. Data Outputs
The workflow of the RAG pipeline is illustrated in Fig. 1, The final output of the analysis was structured in a CSV file
which highlights the key stages of the process from input text containing key variables, including the ticker, publication date,
data to sentiment prediction. closing price, previous day price, next day price, sentiment
label, and sentiment score. This output file served as the
primary dataset for evaluating the relationship between senti-
ment and stock price movements. The alignment of sentiment
scores with next-day stock prices allowed for a comprehensive
analysis of how public sentiment influences short-term market
trends.

H. Predictive Modeling
To predict stock price directions, the combined dataset of
sentiment scores and historical price data was used as input to
machine learning models. Neural networks and Naive Bayes
classifiers were employed to model the relationship between
sentiment and stock price movements. The RAG framework
Fig. 1. Workflow of the LLM and Retrieval-Augmented Generation (RAG) enhanced the robustness of predictions by ensuring that only
model for financial sentiment analysis. The process includes input text data
retrieval, LLM processing, and sentiment prediction. relevant and sentiment-rich data informed the model.

I. Evaluation and Validation

E. Sentiment Analysis
The performance of the proposed methodology was evalu-
The sentiment analysis process categorized each retrieved
ated using metrics such as accuracy, precision, recall, and F1
news article as positive, negative, or neutral depending on its
score. Comparative analysis was conducted against traditional
tone and relevance to stock performance. This categorization
sentiment analysis techniques (e.g., lexicon-based or TF-IDF
was performed by the pre-trained FinBERT model to improve
approaches) to validate the effectiveness of the RAG pipeline.
accuracy. The language model’s ability to interpret and analyze
The results demonstrated significant improvements in predic-
contextually rich data, supported by the retrieved embeddings,
tive accuracy, confirming the viability of integrating retrieval-
allowed it to produce accurate sentiment classifications. These
augmented generation with sentiment analysis for financial
sentiment scores were then associated with the corresponding
market predictions. Additionally, an LLM-based evaluation
stock price data, enabling a comprehensive evaluation of the
mechanism was employed to judge sentiment classification
relationship between news sentiment and stock price move-
results where large-scale labeled data was not readily available.
ments.

F. Integration of Stock Price Data J. System Design

Historical stock prices were integrated into the analysis The system design of this research project integrates var-
to quantify the impact of sentiment on market behavior. ious components to enable effective sentiment analysis and
Closing prices, previous-day prices, and next-day prices were stock price prediction using a Retrieval-Augmented Gener-
aligned with the sentiment scores of related news articles. ation (RAG) approach. The design includes data collection,
 4

preprocessing, feature extraction, sentiment analysis, and pre- IV. D ISCUSSION : R ESULTS AND C ONCLUSIONS
diction modeling, culminating in a recommendation engine for A. Volatility and Sentiment
financial decision-making.
Negative sentiment is often associated with higher volatility
The flowchart in Fig. 2 outlines the end-to-end process from
compared to neutral and positive sentiments.
data input to output generation. The workflow begins with
the collection of historical stock prices and financial news
TABLE II
data, followed by preprocessing steps to clean and format the S ENTIMENT LABELS AND THEIR CORRESPONDING VOLATILITY
data for analysis. Sentiment analysis is then conducted using PERCENTAGES .
machine learning models, and the results are passed through a
Sentiment Label Volatility %
prediction model and recommendation engine. The final output Negative 5.47
is a user interface that presents actionable insights to users. Neutral 3.02
Positive 2.76

Stocks with higher volatility tend to have more extreme

sentiment scores. For example, stocks like Tesla (TSLA) and
Nvidia (NVDA) show a wider range of sentiment scores,
reflecting the high volatility and sentiment swings.

TABLE III
S TOCK VOLATILITY AND STANDARD DEVIATIONS .

Ticker Stock Name Volatility Std. Dev.

TSLA Tesla High 15.23
NVDA Nvidia High 12.87
AAPL Apple Medium 8.45
MSFT Microsoft Medium 7.92
AMZN Amazon Medium 9.34

B. Impact of Sentiment on Next-Day Stock Prices

A key objective of this study was to determine the extent
to which sentiment analysis can predict next-day stock price
movements. The dataset was analyzed to calculate the percent-
Fig. 2. System Design Flow Chart illustrating the end-to-end data flow from age change between the closing price and the next day’s price,
input collection to output generation using the RAG pipeline. grouped by sentiment label.
The analysis shows a clear relationship between sentiment
labels and the next-day stock price movement:
K. Implementation Details
• Negative sentiment: On average, stocks with a negative
This study utilized a proprietary framework developed using sentiment label experienced a -1.29% decrease in next-
a combination of large language models (LLMs) and retrieval- day price.
augmented generation (RAG) techniques. The implementation • Neutral sentiment: Stocks with a neutral sentiment label
leveraged FinBERT, a fine-tuned language model for financial showed minimal change, averaging -0.08% movement.
sentiment analysis, as the core model to perform sentiment • Positive sentiment: Stocks with positive sentiment saw
classification on financial news articles. a 0.29% increase in next-day price.
The data collection, preprocessing, vectorization, and senti-
These results demonstrate that sentiment analysis provides
ment classification pipelines were custom-built using Python,
meaningful predictive signals for stock movements. While
incorporating libraries such as Pandas, Scikit-learn, FAISS for
positive sentiment can suggest a slight upward movement,
vector storage, and Hugging Face transformers for language
negative sentiment tends to lead to more substantial declines.
models. The workflow also included custom scripts to align
This indicates that negative news has a stronger immediate
sentiment scores with historical stock price data for regression
impact on stock performance compared to positive news,
analysis.
which aligns with behavioral finance theories suggesting that
The proprietary LLM-based model architecture and custom
investors react more strongly to negative information than to
Python code were essential in ensuring accurate sentiment
positive news.
classification and alignment with stock price movements. The
end-to-end system flow chart depicting this workflow is shown
in Fig. 2. C. Supporting Data for Sentiment Analysis
The full proprietary codebase used in this study, includ- The table below provides the aggregated statistics used to
ing the custom LLM implementation and sentiment analysis support the bar chart visualization (Fig. 3) and highlights the
workflow, was developed by the author using Python and is average price movement and the number of occurrences for
available upon request for academic purposes. each sentiment label.
 5

TABLE IV baseline movement for negative sentiment, whereas the table

S ENTIMENT A NALYSIS S UPPORTING DATA captures actual average movements from past occurrences.
Sentiment Label Avg. Price Move- Occurrences This highlights that the regression model provides a simplified
ment (%) linear relationship, which may not perfectly align with the
Negative -1.29 3,479 averages of observed data points.
Neutral -0.08 20,130
Positive 0.29 10,695 The strength of the relationship between sentiment scores
and next-day stock price movements is quantified using the
R2 value, which measures the proportion of variance in the
D. Visual Representation of Sentiment Influence on Stock dependent variable explained by the independent variable:
Prices Pn
The results are further illustrated in Fig. 3, which shows 2 (yi − ŷi )2
R = 1 − Pi=1
n 2
(2)
the average next-day stock price movement for each sentiment i=1 (yi − ȳ)
label. The bar chart highlights that negative sentiment has a
more pronounced effect on stock prices compared to positive In this study, the R2 value was calculated to be 0.010,
sentiment, underscoring the potential value of sentiment anal- indicating that sentiment scores explain approximately 1% of
ysis in financial decision-making. the variance in next-day stock price movements. This suggests
that while sentiment analysis has a measurable impact on stock
prices, other factors such as market conditions, sector trends,
and macroeconomic indicators play a more significant role in
influencing price movements.
The model’s intercept and coefficient values provide ac-
tionable insights for financial analysts. However, the low R2
value highlights the need to incorporate additional variables
to improve the model’s predictive accuracy.

F. Discussion of Regression Results

The sentiment scores show a correlation with stock price
movement, the regression model indicates that sentiment alone
explains only a small fraction (1%) of the variance. This
Fig. 3. Average Next-Day Stock Price Movement by Sentiment Label. suggests that while sentiment is a contributing factor, it should
Negative sentiment is associated with the largest price decline, while positive be analyzed in conjunction with other market indicators.
sentiment shows a slight upward movement. While this may seem like a weak relationship, it is impor-
tant to recognize that short-term stock price movements are
influenced by a multitude of factors beyond sentiment. These
E. Regression Analysis factors include broader market conditions, investor behavior,
To quantify the relationship between sentiment scores and macroeconomic events, and technical trading patterns [5].
next-day stock price movements, a simple linear regression Even company-specific news, which is already included in
model was applied. The model predicts next-day price move- sentiment data, does not fully determine next-day stock prices.
ment (y) based on the sentiment score (x) using the following In the context of financial trading, even small predictive
linear equation: signals can be valuable. Hedge funds and algorithmic traders
often rely on models that aggregate many small signals to gain
y = β0 + β1 x (1) an edge in the market [6]. Therefore, while sentiment alone
is not sufficient to predict stock prices with high accuracy, it
For this study: remains a useful input when combined with other indicators.
• β0 = −0.87 (intercept), indicating that a stock with a Additionally, behavioral finance research supports the no-
negative sentiment score is expected to decrease by 0.87% tion that sentiment influences stock prices, particularly in the
on average the next day. short term [7], [8]. Negative sentiment tends to have a stronger
• β1 = 0.65 (coefficient), suggesting that for each unit immediate impact on prices, whereas positive sentiment often
increase in the sentiment label (e.g., from negative to has a more gradual effect [9]. This asymmetry further high-
neutral, or neutral to positive), there is an average increase lights the importance of understanding sentiment as a factor
of 0.65% in next-day price movement. in financial decision-making.
It is important to note that the intercept value of −0.87% The low R-squared value also underscores the need to
represents the predicted next-day price movement based on explore additional variables that can improve the explanatory
the regression model, while the −1.29% average shown power of the model. Future work could incorporate broader
in Table IV is derived from observed historical data. The market sentiment, sector-specific trends, and macroeconomic
difference arises because the regression model estimates a indicators to develop a more robust predictive framework.
 6

V. C HALLENGES AND F UTURE C ONSIDERATIONS Future improvements to the model, including incorporating
The deployment of sentiment analysis models in financial real-time data and more advanced natural language processing
markets involves several key challenges that need to be ad- techniques, will further enhance its robustness and accuracy.
dressed to ensure reliable and robust predictions. This research underscores the evolving role of sentiment
analysis in financial markets and its potential to improve
trading strategies and risk management practices.
A. Data Privacy
Financial data often contains sensitive information that must VII. F UTURE W ORK
be handled responsibly to ensure compliance with data privacy Future work will focus on expanding the data sources to
regulations. Ensuring that data is anonymized and securely include real-time sentiment data from various online platforms,
stored is critical to protecting both individuals and institutions which could provide more immediate insights into market
from data breaches and misuse. movements. Additionally, incorporating more advanced natural
language processing (NLP) techniques could further improve
B. Algorithm Bias the accuracy of sentiment classification.
Sentiment analysis models, such as FinBERT, can introduce Another key area of improvement involves integrating
biases in their predictions based on the training data used. broader market sentiment indicators, sector-specific trends, and
Biases can arise from factors such as imbalanced datasets or macroeconomic data to enhance the model’s predictive power.
overrepresentation of certain types of sentiment. It is important The model’s ability to adapt to real-time data processing
to continuously monitor and mitigate these biases to ensure fair will also be prioritized, making it more suitable for practical
and accurate predictions. applications in financial trading strategies.
Furthermore, addressing challenges such as data privacy,
algorithm bias, and continuous model fine-tuning will be
C. Continuous Model Fine-Tuning critical for ensuring the reliability and robustness of sentiment
Financial markets are dynamic, and sentiment analysis mod- analysis models in dynamic market environments.
els need to be updated regularly to maintain their predictive
accuracy. Continuous fine-tuning of the model with new data is ACKNOWLEDGMENTS
essential to adapt to evolving market conditions and emerging
I would like to express my heartfelt gratitude to my
trends.
mentor, Mohith Manohar, who holds an M.S. in Computer
By addressing these challenges, future research can improve Science from Columbia University and has extensive experi-
the reliability and robustness of sentiment analysis models ence in software engineering, high-performance computing,
and ensure they remain valuable tools for financial decision- and machine learning. His professional background spans
making. contributions to AI-driven solutions at companies such as
Simply Fixable and CoreStack, where he developed cutting-
VI. C ONCLUSION edge technologies using large language models, retrieval-
In conclusion, this research demonstrates the significant augmented generation (RAG) pipelines, and vector databases.
potential of advanced machine learning models and retrieval- He continues to be active in these technologies in the biotech
augmented generation (RAG) systems for financial market industry.
sentiment analysis. By using FinBERT as the core language His mentorship throughout this project has been invaluable,
model, this study developed a framework to predict next-day particularly in guiding me through the intricacies of building
stock price movements based on sentiment extracted from and fine-tuning language models and applying sentiment anal-
financial news articles. The data preprocessing and sentiment ysis techniques in a financial context. His insightful feedback
scoring methodologies show that sentiment scores have a on model design, data preprocessing, and RAG implementa-
measurable, albeit limited, impact on stock price fluctuations. tion helped shape this research into a comprehensive frame-
The regression analysis indicates that sentiment explains work for market sentiment analysis.
approximately 1% of the variance in next-day stock price I am deeply grateful for his encouragement, expertise, and
movements, suggesting that while sentiment is an important commitment to helping me refine my technical and analyt-
input, other factors such as market conditions, sector trends, ical skills. I would also like to acknowledge Polygence for
and macroeconomic indicators play a significant role. The providing an exceptional platform that connects students with
findings highlight that negative sentiment tends to have a experienced mentors, enabling young researchers like me to
stronger immediate impact on stock prices compared to posi- explore and pursue their academic passions.
tive sentiment.
Despite challenges like data privacy concerns, algorithm R EFERENCES
biases, and the need for continuous model updates, sentiment [1] H. Kim and H. Oh, “What Stocks Do You Recommend Today? A
analysis remains a valuable tool for investors and analysts. Pioneering Approach to Stock Market Analysis Using Advanced NLP
The integration of sentiment insights with traditional financial Techniques,” SSRN, 2023.
[2] I. K. Nti, A. Adekoya, and J. Weyori, “Predicting Stock Market Price
indicators offers a comprehensive approach to predicting stock Movement Using Sentiment Analysis: Evidence From Ghana,” Heliyon,
movements and making informed investment decisions. vol. 6, no. 5, pp. e03921, 2020. DOI: 10.1016/[Link].2020.e03921.
 7

[3] T. Sahu, P. H. Shah, and A. Kumar, “Financial Market Prediction Using

Machine Learning Techniques: A Comprehensive Survey,” Journal of
King Saud University – Computer and Information Sciences, 2020. DOI:
10.1016/[Link].2020.03.013.
[4] Journal of King Saud University, “AI Techniques in Financial Trading:
A Systematic Literature Review,” 2023.
[5] A. W. Lo, “The Adaptive Markets Hypothesis: Market Efficiency from
an Evolutionary Perspective,” The Journal of Portfolio Management, vol.
30, no. 5, pp. 15-29, 2004.
[6] R. C. Grinold and R. N. Kahn, Active Portfolio Management: A
Quantitative Approach for Producing Superior Returns and Controlling
Risk. New York: McGraw-Hill, 1999.
[7] R. J. Shiller, “Do Stock Prices Move Too Much to Be Justified by
Subsequent Changes in Dividends?,” The American Economic Review,
vol. 71, no. 3, pp. 421-436, 1981.
[8] M. Baker and J. Wurgler, “Investor Sentiment and the Cross-Section of
Stock Returns,” The Journal of Finance, vol. 61, no. 4, pp. 1645-1680,
2006.
[9] P. C. Tetlock, “Giving Content to Investor Sentiment: The Role of Media
in the Stock Market,” The Journal of Finance, vol. 62, no. 3, pp. 1139-
1168, 2007.
[10] S. Sohangir, D. Wang, A. Pomeranets, and T. M. Khoshgoftaar, “Deep
Learning for Financial Sentiment Analysis,” Journal of Big Data, vol.
5, no. 3, pp. 1–25, 2018. DOI: 10.1186/s40537-017-0111-6.