main.tex

%! Author = caerulescens
%! Advisor = Mirek Mystkowski
%! Date = 3/29/17
%! Title = Artificial Neural Networks for Forecasting Daily Closing Values

% Preamble
\documentclass{ncjms}

% Packages

% Document
\begin{document}
    % Metadata
    \title{Artificial Neural Networks for Forecasting Daily Closing Values}
	\titlerunning{Forecasting with Neural Networks}
	\author[A. Linzie]{Andrew Linzie}
	\address[A. Linzie]{Department of Mathematical Sciences, P.O. Box 7261, Boiling Springs, NC 28017}
	\email[A. Linzie]{alinzie@gardner-webb.edu}
	\authorsrunning{A. Linzie}
	\subjclass[2010]{ 60G25; 62M20}
	\keywords{Artificial Neural Networks; Back-Propagation; Extreme Learning Machine; Forecasting; Stock Market.}
	\date{March 31, 2017}

    % Abstract
    \begin{abstract}
		Over the past decades, machine learning has become an essential area of research with relevant applications in classification and regression.
		Artificial intelligence techniques can be used for statistical analysis of stock markets which is one example of a time series.
		Within this research, Artificial Neural Network models were trained for the purpose of forecasting daily closing prices of five arbitrarily chosen stocks: Apple inc., Walmart, Bank of America Company, Ford Motor Company, and Coca-Cola.
		The accuracy of Back-Propagation models optimized using stochastic gradient descent were compared with an algorithm originating from Nanyang Technical University called Extreme Learning Machines.
		The results indicate that the Extreme Learning Machine model trained faster, classified the up and down of a stock more accurately, and had closer predictions when compared with the Back-Propagation Neural Network model.
	\end{abstract}

    % Title
    \maketitle

	% Introduction
	\section{Introduction}\label{sec:introduction}
	Seldom reward is absent from risk, and stock markets are a prime example.
	Stock markets across the world are viewed as profitable and risky, which attracts companies to forecast these systems.
	Floor trading has progressed towards high-frequency trading with supercomputers that exist within the exchange.
	The computer's computational ability, length of the cable to the exchange, and more have been standardized in favor of no single company having an advantage with exception to algorithms.

	Today, computers are delegated the buying and selling of stocks in the New York Stock exchange, and machine learning theory can be applied towards classification and level estimation.
	Classification refers to the labeling of unseen data as a finite number of categories, while level estimation refers to predicting the price of a stock.
	Artificial Neural Networks have been popular for forecasting noisy systems because of their universal approximation capability.
	\citet{Cybenko:1989} was one of the first to prove the universal approximation theorem stating that a feed-forward network with a single hidden layer containing a finite number of neurons can approximate continuous functions on compact subsets of $R^n$.
	Within this article, Artificial Neural Networks are applied for non-linear classification and level estimation of daily closing prices for blue chip stocks.

	Within section 2, the Efficient Market and Random Walk Hypotheses are introduced, while different weight optimization methods for Artificial Neural Networks are compared.
	The theory of Back-Propagation and Extreme Learning Machines is presented within section 3, and the application of those methods is discussed in section 4.
	Section 5 contains the tabled experimental results, section 6 analyzes the results, and section 7 provides concluding remarks.

	% Literature Review
	\section{Literature Review}\label{sec:literature-review}
	\input{literature-review}

	% Weight Optimization
	\section{Weight Optimization}\label{sec:weight-optimization}
	\input{weight-optimization}

	% Methods
	\section{Methods}\label{sec:methods}
	\input{methods}

	% Results
	\section{Results}\label{sec:results}
	\input{results}

	% Discussion
	\section{Discussion}\label{sec:discussion}
	\input{discussion}

	% Conclusion
	\section{Conclusion}\label{sec:conclusion}
	The mathematical theories and methods of training neural networks for forecasting stock markets were presented within this article.
	While the findings of others and this paper have very positive results on independent data, the academic playground mentality might not transfer to an actual financial market.
	If authors modify their testing parameters to perform well on a testing set, then the model might not generalize well to the real world.
	The testing parameters for all of the neural networks were held constant within this experiment with the goal of finding a general model for forecasting daily closing prices.
	The analysis method used only represents half of a larger solution which would scrape new articles, tag parts of speech, rate the article's affect, and factor that new information into forecasts.

	An Artificial Neural Network model represents an important component of a system for forecasting daily stock change.
	Extreme Learning Machines and Back-Propagation with stochastic gradient descent were used and evaluated against each other.
	Of the five different stocks both networks were tested on, the Extreme Learning Machine model had better results.
	The Extreme Learning Machine model had lower training time, lower MSE, and higher classification success.
	The hit or miss ratio provided within this paper implies that Extreme Learning Machines would perform well with forecasting the direction of change in a market.
	Extreme Learning Machines are a relatively new method for optimizing neural networks.
	Another important question to consider is whether Extreme Learning Machines are able to outperform deep learning algorithms.
	More research should be published to further evaluate using Extreme Learning Machines for regression.

	% Acknowledgments
	\section*{Acknowledgments}
	Special thanks to Dr. Miroslaw Mystkowski from the Department of Mathematical Sciences at Gardner-Webb University for his advising and support.

	% References
	\bibliographystyle{apa}
	\bibliography{references}

\end{document}