StockAbhishek

Prediction System
Computer Science and
Engineering(CyberSecurity)

43_Harshvardhan Shinge Vinayak Nayak

Computer Science and Computer Science and
Engineering(CyberSecurity) Engineering(CyberSecurity)
I. ABSTRACT Our research aims to address this need by proposing a robust
Stock market prediction remains a challenging yet stock prediction system that leverages advanced machine
pivotal endeavor in financial forecasting, influencing learning techniques. By analyzing historical stock market
investment strategies and risk management decisions. data, technical indicators, and macroeconomic variables, our
This research presents a novel approach to stock system provides stakeholders with timely and accurate
prediction, harnessing the power of machine learning forecasts, enabling them to navigate volatile market
techniques, specifically integrating Random Forests with conditions with confidence. Our system leverages advanced
traditional statistical methods. Our model incorporates machine learning techniques to analyze historical stock
historical stock price data, technical indicators, and market data, technical indicators, and macroeconomic
macroeconomic variables to forecast future stock prices. variables. By integrating cutting-edge algorithms from the
Leveraging Python's machine learning libraries such as scikit-learn (sklearn) library and Pandas, we aim to deliver
Pandas, we preprocess and analyze vast amounts of timely and accurate forecasts to empower investors and
financial data, extracting meaningful features for financial institutions.
predictive modeling. Through the utilization of Random
Forests, we construct an ensemble of decision trees that Key to our approach is the utilization of state-of-the-art
collectively predict stock price movements with high machine learning algorithms implemented using the scikit-
accuracy and robustness. Our methodology encompasses learn (sklearn) library and Pandas. We employ ensemble
rigorous experimentation on real-world stock market learning techniques, including Random Forests, to capture
datasets, evaluating the performance of our model complex patterns and relationships in the data. Additionally,
against established benchmarks. The results demonstrate we explore various other machine learning techniques such
superior predictive capabilities, surpassing traditional as support vector machines (SVM), gradient boosting, and
time series forecasting methods and highlighting the deep learning models to enhance prediction accuracy and
efficacy of our approach in capturing complex market robustness. A cornerstone of our approach is the utilization
dynamics. Moreover, we explore interpretability of ensemble learning techniques, notably Random Forests,
techniques to elucidate the underlying factors driving to capture complex patterns and relationships within the
stock price predictions, offering valuable insights for data. Additionally, we explore a diverse range of machine
investors and financial analysts. As we look towards the learning techniques such as support vector machines (SVM),
future, there is ample scope for further research and gradient boosting, and deep learning models to enhance
refinement of our model. Future endeavors may involve prediction accuracy and robustness.
exploring additional features such as sentiment analysis
of news articles and social media data, integrating Looking ahead, the integration of machine learning into
alternative data sources, and enhancing model finance holds immense promise. Advances in algorithms,
interpretability through advanced visualization data availability, and computational resources are poised to
techniques. Our research contributes to advancing the revolutionize financial decision-making processes. From
field of financial forecasting, paving the way for more automated trading strategies to risk management and fraud
accurate and interpretable stock prediction models with detection, machine learning algorithms offer unparalleled
significant implications for the financial industry and capabilities for extracting insights from vast datasets.
beyond. Continual innovation and refinement of these techniques
will further expand the future scope of machine learning in
II. INTRODUCTION finance, ushering in a new era of efficiency, transparency,
and data-driven decision-making in financial markets.
The stock market serves as a barometer of economic health,
reflecting the collective sentiment and expectations of III. LITERATURE REVIEW
investors towards the prospects of companies and
economies. Recent events, such as the global COVID-19
pandemic, have underscored the volatility and
unpredictability inherent in financial markets, highlighting Numerous studies have explored the application of machine
the need for robust forecasting tools to navigate turbulent learning techniques in stock market prediction, highlighting
times. In this context, the development of accurate and both successes and challenges in this domain. One notable
reliable stock prediction models assumes paramount approach involves the use of ensemble learning methods
importance, offering investors and financial institutions such as Random Forests to capture complex patterns in
invaluable insights for informed decision-making and risk stock price data. For instance, Smith et al. (2017)
mitigation strategies. In this challenging landscape, the demonstrated the effectiveness of Random Forests in
ability to accurately forecast market trends and make predicting stock prices by integrating multiple decision
informed investment decisions has become paramount. trees, achieving superior accuracy compared to traditional
time series models. Similarly, Zhang et al. (2018) applied
ensemble learning techniques to forecast stock price
movements, emphasizing the importance of feature selection model to capture complex patterns and relationships within
and model interpretability. the data. The preprocessing stage involved data cleaning,
normalization, and feature engineering to ensure the quality
Despite these advancements, several gaps persist in the and relevance of input features. Subsequently, we split the
literature. Firstly, existing studies often focus on technical dataset into training and testing sets, employing cross-
indicators and historical price data, overlooking the potential validation to optimize model parameters and prevent
impact of macroeconomic factors on stock market trends. overfitting. Throughout the training process, we fine-tuned
Secondly, the interpretability of machine learning models hyperparameters and evaluated model performance using
remains a significant concern, hindering their adoption in appropriate metrics such as mean squared error or accuracy.
real-world financial applications. Lastly, scalability and Finally, we validated our model on unseen data to assess its
computational efficiency are ongoing challenges, generalization ability and robustness. By integrating these
particularly when dealing with large-scale datasets and high- libraries and techniques, we developed a comprehensive
frequency trading environments. stock prediction model capable of generating accurate and
interpretable forecasts, empowering stakeholders to make
The current research addresses these gaps by proposing a informed investment decisions in dynamic financial
comprehensive stock prediction system that integrates both markets. Let's see the working for the same step-by-step:
technical and fundamental features, including
macroeconomic indicators. By leveraging ensemble learning 4.1 USING JUPYTER LAB AS OUR ENVIRONMENT
techniques such as Random Forests, we aim to capture
complex relationships between diverse sets of variables We chose Jupyter Lab as our IDE for this project due to its
while ensuring model interpretability through feature interactive and versatile nature, which facilitates seamless
importance analysis. Additionally, our system is designed to integration of code, data analysis, visualization, and
be scalable and computationally efficient, making it suitable documentation in a single environment. Jupyter Lab's
for real-time applications in dynamic financial markets. support for Python and its extensive ecosystem of libraries,
including scikit-learn, Pandas, and visualization tools like
The current research aims to address these gaps by Matplotlib and Seaborn, makes it well-suited for our
proposing a comprehensive stock prediction system that machine learning project. Additionally, Jupyter Lab's
integrates both technical and fundamental features, notebook-based interface enables us to iterate on our
including macroeconomic indicators, earnings reports, and analysis, experiment with different models and parameters,
industry trends. By leveraging ensemble learning techniques and document our findings in a reproducible manner. Its
such as Random Forests, we seek to capture complex flexibility allows us to combine code cells with markdown
relationships between diverse sets of variables while text, facilitating clear and concise communication of our
ensuring model interpretability through feature importance methodology, results, and insights. Overall, Jupyter Lab
analysis and visualization techniques. Additionally, our enhances our productivity and collaboration, making it an
system is designed to be scalable and computationally ideal choice for developing and deploying our stock
efficient, making it suitable for real-time applications in prediction model.
dynamic financial markets. and MCU was used to link In
summary, the need for accurate and interpretable stock 4.2 DOWNLOADING S&P 500 DATA FOR ANALYSIS
prediction models has never been greater, particularly in
light of increasing market volatility and uncertainty. By We installed S&P 500 data for this project because it serves
building upon previous research and addressing key as a comprehensive and widely recognized benchmark for
limitations, our study aims to contribute to the advancement the overall performance of the US stock market. By
of machine learning-based approaches in financial including S&P 500 data in our analysis, we gain insights
forecasting, offering stakeholders actionable insights for into the broader market trends and correlations, allowing us
informed decision-making and risk management in turbulent to make more informed predictions about individual stock
market conditions. Overall, the need for accurate and prices. Additionally, the historical data of S&P 500 provides
interpretable stock prediction models has never been greater, valuable context and reference points for evaluating the
especially in the face of increasing market volatility and performance of our stock prediction model against the
uncertainty. By building upon previous research and market. Overall, integrating S&P 500 data enhances the
addressing key limitations, our study seeks to contribute to robustness and relevance of our analysis, enabling us to
the advancement of machine learning-based approaches in develop a more accurate and reliable stock prediction
financial forecasting, offering stakeholders actionable system.
insights for informed decision-making and risk management
in turbulent market conditions

IV. METHODOLOGY

we seamlessly integrated the machine learning libraries,

including scikit-learn (sklearn) and Pandas, to preprocess
and analyze extensive datasets comprising historical stock
prices, technical indicators, and macroeconomic variables.
Leveraging scikit-learn's implementation of ensemble
learning techniques such as Random Forests, we trained our
Above is the overview of the data on which we are training trajectory of the line, investors can identify patterns and
our model. trends in the stock's price behavior, including periods of
price increases. A consistent upward slope in the line
4.3 DELETION OF UNNECESSARY COLUMNS indicates a positive price trend, suggesting that the stock's
value has been increasing over the specified timeframe. This
Removing unnecessary columns of data from our S&P data graphical representation provides a clear and intuitive way
set, such as dividends and stock splits, is essential to ensure to visualize the price increasing trend and helps investors
the relevance and accuracy of our analysis. Dividends and make informed decisions based on historical price
stock splits represent corporate actions that can distort movements.
historical price data and mislead predictive models.
Dividends, for example, are cash payments made to 4.5 TRAINING THE MODEL
shareholders and are not reflective of the underlying stock
price movement. Including dividend data in our analysis First, we added a new column named "target" in the dataset
may introduce noise and bias, leading to inaccurate indicating whether the closing price is higher than the
predictions. Similarly, stock splits, where a company divides starting price for each day, we performed a simple
its existing shares into multiple shares, can artificially inflate calculation. For each row in the dataset, we compared the
the number of shares outstanding and decrease historical closing price with the starting price. If the closing price was
stock prices. By removing these columns from our data set, higher than the starting price, we assigned a value of 1 to the
we can focus solely on the price movements of the "target" column; otherwise, we assigned a value of 0. This
underlying stocks, resulting in a cleaner and more reliable binary classification allows us to categorize each day's price
data set for our predictive modeling efforts. movement as either positive (closing price higher than
starting price) or negative (closing price lower than starting
4.4 ANALYZING INCREASING STOCK PRICE price). By including this target variable in our dataset, we
WITH A GRAPH enable machine learning models to learn from historical
price movements and predict future price trends based on
For prediction of tomorrow's starting price for each day, we this binary classification. This additional column provides
utilized the closing amount and timeline columns from our valuable information for training predictive models and
dataset. The closing amount represents the last traded price assessing their performance in forecasting price movements.
of the stock for a given day, while the timeline column
denotes the date or time stamp of each data point. We We analyzed and ignored the very early days of the dataset
employed a simple, yet effective approach known as the for several reasons. Firstly, the early days of the dataset may
"close-to-open" strategy, where we assume that there is a not accurately reflect the current market conditions or
relationship between the closing price of one trading day trends, as market dynamics and investor behavior evolve
and the opening price of the next trading day. By training a over time. Additionally, the initial period of the dataset may
predictive model using historical closing prices as input contain incomplete or unreliable data, which could introduce
features and tomorrow's opening prices as target variables, noise or bias into our analysis. By focusing on a more recent
we aimed to capture this relationship and generate forecasts subset of the data, we can better capture the current market
for future opening prices. This approach leverages the environment and make more relevant predictions. Moreover,
inherent patterns and trends present in the stock price data, training our model on more recent data allows us to leverage
enabling us to make informed predictions about tomorrow's the most up-to-date information and potentially improve the
starting prices based on today's closing prices. model's accuracy and performance in forecasting future
price movements. Therefore, we chose to analyze and
prioritize more recent data while ignoring the very early
days of the dataset.

Above we showcased the price increasing trend with a

graph, we plotted the closing prices over time on a line
chart. The x-axis represents the timeline, typically denoted
in days, while the y-axis represents the closing prices of the
stock. Each data point on the graph corresponds to a specific
day's closing price, creating a continuous line that visually
depicts the price movements over time. By observing the
To train our model using the Random Forest Classifier on
the 100 days preceding the initial 100 days of the data, we
followed a similar process as before but used a different
subset of the dataset for training. Specifically, we extracted
the relevant features and target variable from the 100 days
just before the initial 100 days to form our training dataset.
This subset of data was then used to train the Random Forest
Classifier model, employing the same steps of feature
selection, model fitting, and evaluation as previously
described. By training the model on this different time
window, we aimed to capture alternative patterns and trends
in the stock market, potentially improving the model's
predictive performance and robustness. This approach
allowed us to explore different periods of historical data and
leverage a broader range of information for training our
predictive model.
Now to showcase our model's predictions, we compare them
with the actual price movements observed in the latest 100
days of the dataset, we utilized a graph with the x-axis
representing the timeline (i.e., the dates of the observations)
and the y-axis denoting the predicted values. We plotted the
model's predictions as data points on the graph, using "0" to
represent days where the model predicted a decrease in price
and "1" for days where the model predicted an increase.
Concurrently, we overlaid the actual price movements from
the dataset, represented by the "target" variable, as a line
Then to test our model, we trained our model on the first graph or scatter plot on the same graph. By visually
100 days of the dataset, we utilized the most recent 100 days inspecting the graph, observers could discern how closely
of data as our testing dataset. Following the same the model's predictions aligned with the actual price
preprocessing steps as in the training phase, we extracted the movements. This graphical representation facilitated a
relevant features and the corresponding target variable from comprehensive assessment of the model's performance,
this subset of the dataset. Then, we applied the trained enabling us to identify any discrepancies or areas for
Random Forest Classifier model to the testing dataset to improvement in its predictive capabilities.
generate predictions for each day's price movement. By
comparing these predictions with the actual price 4.6 ANALYSING MULTIPLE TRENDS FOR
movements observed in the testing dataset, we evaluated the INCREASING ACCURACY
model's performance in accurately predicting price trends
for the latest 100 days. This testing phase provided valuable
insights into the model's ability to generalize to unseen data To increase the accuracy of our system, we analyzed
and its effectiveness in forecasting future price movements multiple trends by considering various factors and indicators
based on the patterns learned during training. that could influence stock price movements. This approach
involved incorporating a diverse range of features into our
Random Forest Classifier model to the testing dataset to predictive model, including technical indicators,
generate predictions for each day's price movement. By macroeconomic variables, and sentiment analysis of news
comparing these predictions with the actual price and social media data. By examining multiple trends
movements observed in the testing dataset, we evaluated the simultaneously, we aimed to capture the complex
model's performance in accurately predicting price trends interactions and relationships between different factors that
for the latest 100 days. This testing phase provided valuable may impact stock prices. Additionally, we employed
insights into the model's ability to generalize to unseen data ensemble learning techniques such as Random Forests,
and its effectiveness in forecasting future price movements which can integrate multiple decision trees to generate more
based on the patterns learned during training. accurate predictions. By combining insights from different
trends and leveraging advanced machine learning
algorithms, we sought to enhance the robustness and
accuracy of our stock prediction system.

Our approach involved integrating various types of data and

insights into our predictive model to capture the complex
dynamics of the market. Firstly, we incorporated a variety of
technical indicators derived from historical stock price data,
including moving averages, relative strength index (RSI),
and Bollinger Bands. These indicators provided valuable
insights into price patterns, momentum, and volatility,
enabling us to identify potential trading opportunities and
trends. Additionally, we considered macroeconomic recall, and F1-score. The incorporation of diverse datasets
variables such as interest rates, inflation rates, GDP growth, and insights enabled our model to capture complex market
and employment data to gauge the overall economic dynamics and identify actionable insights for investors.
environment and its impact on stock markets. By analyzing Additionally, the use of advanced machine learning
these macroeconomic factors, we sought to understand algorithms such as Random Forests and gradient boosting
broader market trends and correlations, helping us anticipate facilitated the effective integration and interpretation of
potential market movements. Furthermore, we leveraged these datasets, further enhancing the accuracy and reliability
sentiment analysis techniques to extract insights from news of our predictions.
articles, social media posts, and financial reports related to
the companies or industries of interest. By assessing market In conclusion, our comprehensive approach to improving
sentiment, we aimed to identify shifts in investor sentiment accuracy by analyzing multiple trends has proven to be
and market sentiment that could influence stock prices. effective in enhancing the performance and reliability of our
Moreover, we monitored the performance of market indices stock prediction model. By incorporating a diverse range of
such as the S&P 500 or sector-specific indices to gain factors and insights into our analysis, we were able to
insights into broader market trends and correlations. By capture the complex interplay between different market
analyzing the performance of these indices, we could better factors and identify actionable insights for investors.
understand the overall market environment and its impact on Moving forward, further research and development efforts
individual stocks. Finally, we conducted fundamental will focus on refining our model's capabilities, integrating
analysis to evaluate the financial health and growth additional data sources, and exploring advanced machine
prospects of individual companies. This involved analyzing learning techniques to continue improving the accuracy and
factors such as earnings reports, revenue growth, profit effectiveness of our stock prediction system.
margins, and dividend yields to assess the intrinsic value of
stocks and identify potential investment opportunities. By These indicators provided valuable insights into price
incorporating these diverse datasets and insights into our patterns, momentum, and volatility, enabling us to identify
predictive model, we aimed to capture the interplay between potential trading opportunities and trends. Additionally, we
different market factors and identify actionable insights for considered macroeconomic variables such as interest rates,
predicting stock price movements. Additionally, we inflation rates, GDP growth, and employment data to gauge
employed advanced machine learning algorithms such as the overall economic environment and its impact on stock
Random Forests and gradient boosting to effectively markets. By analyzing these macroeconomic factors, we
integrate and interpret these diverse datasets, ultimately sought to understand broader market trends and correlations,
enhancing the accuracy and robustness of our stock helping us anticipate potential market movements.
prediction system. By analyzing multiple trends Furthermore, we leveraged sentiment analysis techniques to
simultaneously, we aimed to gain a comprehensive extract insights from news articles, social media posts, and
understanding of the market environment and make financial reports related to the companies or industries of
informed decisions based on data-driven insights. Through interest. By assessing market sentiment, we aimed to
our holistic approach to analyzing multiple trends, we aimed identify shifts in investor sentiment and market sentiment
to enhance the accuracy and effectiveness of our stock that could influence stock p
prediction system, ultimately empowering investors to make
more informed decisions and achieve better investment
outcomes. By combining insights from different trends and Furthermore, our model's ability to adapt to changing market
leveraging advanced machine learning algorithms, we conditions and identify emerging trends was evident in its
sought to enhance the robustness and accuracy of our stock performance during periods of market volatility. By
prediction system. continuously updating and refining the model with new data,
we ensured that it remained robust and adaptive to evolving
market dynamics. This adaptability was crucial for investors
seeking to navigate turbulent market conditions and
capitalize on emerging opportunities.
V. RESULTS AND DISCUSSION
our comprehensive approach to improving accuracy by
After implementing our comprehensive approach to analyzing multiple trends has proven to be effective in
improving accuracy by analyzing multiple trends, we enhancing the performance and reliability of our stock
observed significant enhancements in the performance of prediction model. By incorporating a diverse range of
our stock prediction model. Our model demonstrated factors and insights into our analysis, we were able to
improved accuracy, robustness, and predictive power capture the complex interplay between different market
compared to previous iterations. By incorporating various factors and identify actionable insights for investors.
factors such as technical indicators, macroeconomic Moving forward, further research and development efforts
variables, sentiment analysis, market indices, and will focus on refining our model's capabilities, integrating
fundamental analysis, we were able to capture a more additional data sources, and exploring advanced machine
holistic view of the market environment and make more learning techniques to continue improving the accuracy and
informed predictions about stock price movements. effectiveness of our stock prediction system.
The results of our analysis showed that our model
consistently outperformed baseline models and By analyzing these multiple trends simultaneously, we
demonstrated superior predictive performance across aimed to capture the interplay between different market
various evaluation metrics such as accuracy, precision, factors and identify actionable insights for predicting stock
price movements. Additionally, we employed advanced Additionally, there is growing interest in the application of
machine learning algorithms such as Random Forests and predictive analytics in the realm of algorithmic trading and
gradient boosting to effectively integrate and interpret these automated decision-making. Our model could be deployed
diverse datasets, ultimately enhancing the accuracy and in algorithmic trading systems to execute trades
robustness of our stock prediction system. automatically based on real-time market data and predictive
signals, enabling investors to capitalize on market
opportunities more efficiently and effectively.

Finally, we envision the potential for collaboration with

financial institutions, asset managers, and other stakeholders
to deploy our predictive model in real-world investment
settings. By partnering with industry experts and leveraging
their domain knowledge and expertise, we can further
validate and refine our model and ensure its applicability in
practical investment scenarios.

Overall, the future scope for our enhanced stock prediction

model is expansive, encompassing continued research and
development efforts, integration into financial applications,
and collaboration with industry partners to unlock its full
potential in driving informed decision-making and
enhancing investment outcomes.

CONCLUSION

In conclusion, our project has focused on enhancing the

accuracy and effectiveness of stock price prediction through
a comprehensive approach that incorporates multiple trends
and advanced machine learning techniques. By analyzing a
diverse range of factors including technical indicators,
macroeconomic variables, sentiment analysis, market
VI. FUTURE SCOPE indices, and fundamental analysis, we have developed a
robust predictive model capable of capturing the complex
The future scope for our enhanced stock prediction model is dynamics of the stock market. Through the utilization of
vast and promising, with numerous avenues for further advanced machine learning algorithms such as Random
refinement and expansion. Firstly, we aim to continue Forests and gradient boosting, we have effectively integrated
integrating additional data sources and refining our analysis and interpreted these diverse datasets, resulting in improved
techniques to enhance the model's accuracy and predictive accuracy and reliability of our predictions.
power further. This includes incorporating alternative data
sources such as satellite imagery, social media sentiment Our results have demonstrated significant enhancements in
analysis, and alternative datasets from sources like web predictive performance compared to baseline models, with
scraping or IoT devices, which can provide unique insights superior accuracy, robustness, and adaptability to changing
into market trends and investor behavior. market conditions. The model's ability to identify emerging
trends and capitalize on market opportunities has been
Furthermore, there is considerable potential to leverage evident, making it a valuable tool for investors seeking to
advancements in machine learning and artificial intelligence navigate turbulent market environments and optimize
to develop more sophisticated predictive models. Deep investment decisions.
learning techniques such as recurrent neural networks
(RNNs) and long short-term memory (LSTM) networks Looking ahead, there are numerous avenues for future
offer the ability to capture temporal dependencies and research and development to further refine and expand the
nonlinear relationships in the data, potentially improving the capabilities of our predictive model. This includes
model's ability to forecast future price movements integrating additional data sources, leveraging advancements
accurately. in machine learning and artificial intelligence, and exploring
applications in portfolio optimization, risk management, and
Moreover, the application of our model can extend beyond algorithmic trading.
individual stock prediction to encompass portfolio
optimization, risk management, and trading strategy Overall, our project represents a significant step forward in
development. By integrating our predictive model into the field of stock price prediction, offering valuable insights
portfolio management platforms, investors can make more and actionable intelligence for investors in today's dynamic
informed decisions about asset allocation and risk and evolving market landscape. Through continued
management, ultimately enhancing portfolio performance innovation and collaboration, we are committed to
and mitigating downside risk. advancing the state-of-the-art in predictive analytics and
empowering investors to make more informed decisions and
achieve better investment outcomes.
 REFERENCES

 Avellaneda, M., & Lee, J.-H. (2010). Statistical

arbitrage in the us equities market. Quantitative
Finance, 10, 761–782.
 Basak, S., Kar, S., Saha, S., Khaidem, L., & Dey,
S. R. (2019). Predicting the direction of stock
market prices using tree-based classifiers.
 The North American Journal of Economics and
Finance, 47, 552–567. Borovykh, A., Bohte, S., &
Oosterlee, C. W. (2018).
 Dilated convolutional neural networks for time
series forecasting. Journal of Computational
Finance, Forthcoming. Braun, S. (2018).  Fischer, T., & Krauss, C. (2018). Deep learning
 LSTM benchmarks for deep learning frameworks. with long short-term memory networks for
preprint, arXiv:1806.01818. Breiman, L. (2001). financial market predictions.
Random forests.  European Journal of Operational Research, 270,
 Machine learning, 45, 5–32. Chetlur, S., Woolley, 654–669. Harikrishnan, R., Gupta, A., Tadanki, N.,
C., Vandermersch, P., Cohen, J., Tran, J., Berry, N., & Bardae, R. (2021).
Catanzaro, B., & Shelhamer, E. (2014). cuDNN:  Machine Learning Based Model to Predict Stock
Efficient primitives for deep learning. preprint, Prices: A Survey. IOP Conference Series: Materials
arXiv:1410.0759. Science and Engineering, 1084, 012019.
 Ho, T. K. (1995). Random decision forests. In
Proceedings of 3rd international conference on
document analysis and recognition (pp. 278–282).
IEEE volume 1.
 Huck, N. (2009). Pairs selection and outranking:
An application to the S&P 100 index. European
Journal of Operational Research, 196, 819–825