TABLE OF CONTENTS
1.1 General Introduction
1.2 Problem Statement
1.3.3 Scikit Learn
1.3.6 Compiler Option
2 LITERATURE SURVEY
3 DATA & TOOLS
3.1.1 Choosing the data-set
3.1.2 Gathering the data-set
4 PREVIOUS ANALYSIS
4.1 Technical Analysis Methods
4.2 Fundamental Analysis Techniques
4.3 Traditional Time Series Prediction
4.4 Machine Learning Methods
4.5 Deep Learning
4.5.1 Artificial Neural Network
188.8.131.52 Artificial Neural Network in Stock Market Prediction
4.5.2 Convolution Neural Network(CNN)
184.108.40.206 Convolution Neural Network in Stock Market Prediction
5 THE PROPOSED MODEL
5.1 Recurrent Neural Networks
5.2 Long Short Term Memory(LSTM)
5.3 Advantages of LSTM
6 SYSTEM DESIGN
6.1 System Architecture
6.1.1 Collect data set
6.1.2 Import Training data
6.1.3 Applying Scaling Features
6.1.4 Creating a neural network
6.1.5 Train the Model
6.1.6 Import test data
6.1.7 Visualize Result
6.1.8 Calculate Efficiency
6.2 LSTM Architectural Diagram
7 SYSTEM REQUIREMENTS
8.1 Data Preprocessing
8.1.1 Libraries Import
8.1.2 Importing the training set
8.1.3 Feature Scaling
8.1.4 Inputs and Outputs
8.2 Building the Recurrent Neural Network (LSTM)
8.2.1 Libraries Import
8.2.2 LSTM Construction
8.2.3 Model Fitting
8.3.1 Importing the Test Data
8.3.2 Scaling and Reshaping Test Data
8.3.3 Predicting Test Data
8.4 Visualization and Results
CONCLUSION AND FUTURE WORK
Forecasting stock market prices have always been challenging task for many business analyst and researchers. In fact, stock market price prediction is an interesting area of research for investors. For successful investment, many investors are interested in knowing about the future situation of the market. Effective prediction systems indirectly help traders by providing supportive information such as the future market direction. Data mining techniques are effective for forecasting future by applying various algorithms to data.
This project aims at predicting stock market by using financial news, Analyst opinions and quotes in order to improve quality of output. It proposes a novel method for the prediction of the stock market closing price. Many researchers have contributed in this area of chaotic forecast in their ways. Fundamental and technical analyses are the traditional approaches so far. ANN is another popular way to identify unknown and hidden patterns in data is used for share market prediction.
In this project, we study the problem of stock market forecasting using Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM). The purpose of this project is to examine the feasibility and performance of LSTM in stock market forecasting. We optimize the LSTM model by testing different configurations, i.e., a multi-layered feed-forward neural network is built by using a combination of data mining. The Neural Network is trained on the stock quotes using the Backpropagation Algorithm which is used to predict share market closing price. The Accuracy of the performance of the neural network is compared using various out of sample performance measures. Modeling techniques and the architecture of the Recurrent Neural Network will also reported in the paper
LIST OF FIGURES
4.1 Machine Learning Analysis Curve
4.2 Artificial Neural Network(ANN)
4.3 Stock Visualization Curve
4.4 Training Error in ANN
4.5 Training ANN based Systems
4.6 Convolution Neural Networks(CNN)
4.7 CNN Input Visualization
4.8 Training and validation loss
5.1 Recurrent Neural Network Architecture
5.2 LSTM Architecture
5.3 Comparison of Sigmoid curves
5.4 Data flow through the memory cell
6.1 System Architecture
6.2 LSTM Architectural Diagram
8.1 Dataset Dataframe
8.2 Training set Dataframe
8.3 Scaled values
8.4 X-train and Y-train dataframe
8.5 Data Structure
8.6 X train and Y train dataframe
8.7 Prediction (Scaled)
8.8 Prediction (Reverse Scaled)
8.9 Visualizing using the Mathplot library
8.10 Visualizing using the Plotly library
8.11 Visualizing using the Plotly library (Zoomed in)
8.12 RMSE Calculation Code
4.2 Comparative Results
LIST OF ABBREVIATIONS
Abbildung in dieser Leseprobe nicht enthalten
CHAPTER 1 INTRODUCTION
1.1 GENERAL INTRODUCTION
Modeling and Forecasting of the financial market have been an attractive topic to scholars and researchers from various academic fields. The financial market is an abstract concept where financial commodities such as stocks, bonds, and precious metals transactions happen between buyers and sellers. In the present scenario of the financial market world, especially in the stock market, forecasting the trend or the price of stocks using machine learning techniques and artificial neural networks are the most attractive issue to be investigated. As Giles explained, financial forecasting is an instance of signal processing problem which is difficult because of high noise, small sample size, non-stationary, and non-linearity. The noisy characteristics mean the incomplete information gap between past stock trading price and volume with a future price. The stock market is sensitive with the political and macroeconomic environment. However, these two kinds of information are too complex and unstable to gather. The above information that cannot be included in features are considered as noise. The sample size of financial data is determined by real-world transaction records. On one hand, a larger sample size refers a longer period of transaction records; on the other hand, large sample size increases the uncertainty of financial environment during the 2 sample period. In this project, we use stock data instead of daily data in order to reduce the probability of uncertain noise, and relatively increase the sample size within a certain period of time. By non-stationarity, one means that the distribution of stock data is various during time changing. Non-linearity implies that feature correlation of different individual stocks is various. Efficient Market Hypothesis was developed by Burton G. Malkiel in 1991. In Burton’s hypothesis, he indicates that predicting or forecasting the financial market is unrealistic, because price changes in the real world are unpredictable. All the changes in prices of the financial market are based on immediate economic events or news. Investors are profit-oriented, their buying or selling decisions are made according to most recent events regardless past analysis or plans. The argument about this Efficient Market Hypothesis has never been ended. So far, there is no strong proof that can verify if the efficient market hypothesis is proper or not. However, as Yaser claims, financial markets are predictable to a certain extent. The past experience of many price changes over a certain period of time in the financial market and the undiscounted serial correlations among vital economic events affecting the future financial market are two main pieces of evidence opposing the Efficient Market Hypothesis. In recent years, machine learning methods have been extensively researched for their potentials in forecasting and prediction of the financial market. Multi-layer feed forward neural networks, SVM, reinforcement learning, relevance vector machines, and recurrent neural networks are the hottest topics of many approaches in financial market prediction field. Among all the machine learning methods, neural networks are well studied and have been successfully used for forecasting and modeling financial market. “Unlike traditional machine learning models, the network learns from the examples by constructing an input-output mapping for the problem at hand. Such an approach brings to mind the study of nonparametric statistical inference; the term “nonparametric” is used here to signify the fact that no prior assumptions are made on a statistical model for the input data”, according to Simon. As Francis E.H. Tay and Lijuan Cao explained in their studies, Neural networks are more noise tolerant and more flexible compared with traditional statistical models. By noise tolerance, one means neural networks have the ability to be trained by incomplete and overlapped data. Flexibility refers to that neural networks have the capability to learn dynamic systems through a retraining process using new data patterns. Long short-term memory is a recurrent neural network introduced by Sepp Hochreite and Jurgen Schmidhuber in 1997. LSTM is designed to forecast, predict and classify time series data even long time lags between vital events happened before. LSTMs have been applied to solve several of problems; among those, handwriting Recognition and speech recognition made LSTM famous. LSTM has copious advantages compared with traditional back-propagation neural networks and normal recurrent neural networks. The constant error back propagation inside memory blocks enables in LSTM ability to overcome long time lags in case of problems similar to those discussed above; LSTM can handle noise, distributed representations, and continuous values; LSTM requires no need for parameter fine-tuning, it works well over a broad range of parameters such as learning rate, input gate bias, and output gate bias. The objective of our project can be generalized into two main parts. We examine the feasibility of LSTM in stock market forecasting by testing the model with various configurations.
1.2 PROBLEM STATEMENT
The stock market appears in the news every day. You hear about it every time it reaches a new high or a new low. The rate of investment and business opportunities in the Stock market can increase if an efficient algorithm could be devised to predict the short term price of an individual stock.
Previous methods of stock predictions involve the use of Artificial Neural Networks and Convolution Neural Networks which has an error loss at an average of 20%.
In this report, we will see if there is a possibility of devising a model using Recurrent Neural Network which will predict stock price with a less percentage of error. And if the answer turns to be YES, we will also see how reliable and efficient will this model be.
Python was the language of choice for this project. This was an easy decision for the multiple reasons.
1. Python as a language has an enormous community behind it. Any problems that might be encountered can be easily solved with a trip to Stack Overflow. Python is among the most popular languages on the site which makes it very likely there will be a direct answer to any query.
2. Python has an abundance of powerful tools ready for scientific computing. Packages such as Numpy, Pandas, and SciPy are freely available and well documented. Packages such as these can dramatically reduce, and simplify the code needed to write a given program. This makes iteration quick.
3. Python as a language is forgiving and allows for programs that look like pseudo code. This is useful when pseudocode given in academic papers needs to be implemented and tested. Using Python, this step is usually reasonably trivial.
However, Python is not without its flaws. The language is dynamically typed and packages are notorious for Duck Typing. This can be frustrating when a package method returns something that, for example, looks like an array rather than being an actual array. Coupled with the fact that standard Python documentation does not explicitly state the return type of a method, this can lead to a lot of trials and error testing that would not otherwise happen in a strongly typed language. This is an issue that makes learning to use a new Python package or library more difficult than it otherwise could be.
Numpy is python modules which provide scientific and higher level mathematical abstractions wrapped in python. In most of the programming languages, we can’t use mathematical abstractions such as f(x) as it would affect the semantics and the syntax of the code. But by using Numpy we can exploit such functions in our code.
Numpy’s array type augments the Python language with an efficient data structure used for numerical work, e.g., manipulating matrices. Numpy also provides basic numerical routines, such as tools for finding Eigenvectors.
1.3.3 Scikit Learn
Scikit-learn is a free software machine learning library for the Python programming language. It features various classification, regression and clustering algorithms including support vector machine, random forest, gradient boosting, k-means etc. It is mainly designed to interoperate with the Python numerical and scientific libraries NumPy and SciPy.
Scikit-learn is largely written in Python, with some core algorithms written in Cython to achieve performance. Support vector machines are implemented by a Cython wrapper around LIBSVM .i.e., logistic regression and linear support vector machines by a similar wrapper around LIBLINEAR.
TensorFlow is an open source software library for numerical computation using data flow graphs. Nodes in the graph represent mathematical operations, while the graph edges represent the multidimensional data arrays (tensors) communicated between them. The flexible architecture allows you to deploy computation to one or more CPUs or GPUs in a desktop, server, or mobile device with a single API. TensorFlow was originally developed by researchers and engineers working on the Google Brain Team within Google's Machine Intelligence research organization for the purposes of conducting machine learning and deep neural networks research, but the system is general enough to be applicable in a wide variety of other domains as well.
TensorFlow is Google Brain's second-generation system. While the reference implementation runs on single devices, TensorFlow can run on multiple CPUs and GPUs (with optional CUDA and SYCL extensions for general-purpose computing on graphics processing units). TensorFlow is available on 64-bit Linux, macOS, Windows, and mobile computing platforms including Android and iOS.
Keras is a high-level neural networks API, written in Python and capable of running on top of TensorFlow, CNTK, or Theano. It was developed with a focus on enabling fast experimentation. Being able to go from idea to result with the least possible delay is key to doing good research.
Keras allows for easy and fast prototyping (through user friendliness, modularity, and extensibility). Supports both convolutional networks and recurrent networks, as well as combinations of the two. Runs seamlessly on CPU and GPU.
The library contains numerous implementations of commonly used neural network building blocks such as layers, objectives, activation functions, optimizers, and a host of tools to make working with image and text data easier. The code is hosted on GitHub, and community support forums include the GitHub issues page, a Gitter channel and a Slack channel.
1.3.6 Compiler Option
Anaconda is a freemium open source distribution of the Python and R programming languages for large-scale data processing, predictive analytics, and scientific computing, that aims to simplify package management and deployment. Package versions are managed by the package management system conda.
The following papers were studied in order to get an overview of the techniques that were applied earlier to predict the stock market.
LSTM Fully Convolutional Networks for Time Series Classification
-Fazle Karim, Somshubra Majumdar, Houshang Darabi and Shun Chen 
With the proposed models, we achieve a potent improvement in the current state-of-the-art for time series classification using deep neural networks. Our baseline models, with and without fine-tuning, are trainable end-to-end with nominal preprocessing and are able to achieve significantly improved performance.
LSTM-FCNs are able to augment FCN models, appreciably increasing their performance with a nominal increase in the number of parameters. An LSTM-FCNs provide one with the ability to visually inspect the decision process of the LSTM RNN and provide a strong baseline on their own. Fine-tuning can be applied as a general procedure to a model to further elevate its performance.
The strong increase in performance in comparison to the FCN models shows that LSTM RNNs can beneficially supplement the performance of FCN modules for time series classification. An overall analysis of the performance of our model is provided and compared to other techniques.
There is further research to be done on understanding why the attention LSTM cell is unsuccessful in matching the performance of the general LSTM cell on some of the datasets. Furthermore, an extension of the proposed models to multivariate time series is elementary but has not been explored in this work.
Learning Long term Dependencies with Gradient Descent is difficult
-Yoshua bengio, Patrice Simard and Paolo Frasconi 
Recurrent networks are very powerful in their ability to represent context, often outperforming static network. But the factor off gradient descent of an error criterion may be inadequate to train them for a task involving long-term dependencies. It has been found that the system would not be robust to input noise or would not be efficiently trainable by gradient descent when the long-term context is required. The theoretical result presented in this paper holds for any error criterion and not only from mean square error.
it can also be seen that the gradient either vanishes or the system is not robust to input noise. The other imp factor to note is that the related problems of vanishing gradient may occur in deep feed-forward networks. The result presented in this paper does not mean that it is impossible to train a recurrent neural network on a particular task. It says that the gradient becomes increasingly inefficient when the temporal span of the dependencies increases. So at one point in time, it is evident that it becomes obsolete.
Improving N Calculation of the RSI Financial Indicator Using Neural Networks
-Alejandro Rodríguez-González, Fernando Guldris Iglesias, Ricardo Colomo-Palacios Giner Alor-Hernandez, Ruben Posada-Gomez 
There has been growing interest in Trading Decision Support Systems in recent years. In spite of its volatility, it is not entirely random, instead, it is nonlinear and dynamic or highly complicated and volatile. Stock movement is affected by the mixture of two types of factors: determinant (e.g. gradual strength change between buying side and selling side) and random (e.g. emergent affairs or daily operation variations).
There are 3 modules that are talked about in this research paper. The Neural Network Module is responsible for providing the N values that are used to calculate RSI and decide if an investor should invest in a certain company.
Trading system Module analyzes the result given by neural network module. When a query is formulated to the system, it takes the actual values of the market and builds a query to the neural network. If RSI value is higher than 70 the decision that trading system return is a sell signal. If RSI value is lower than 30 the decision that trading system return is a buy signal.
The heuristic module is in charge of managing the different formulas that provide the heuristic used to generate the optimal values for RSI indicator.
Stock Trend Prediction Using Simple Moving Average Supported
by News Classification
-Stefan Lauren Dra. Harlili S., M.Sc. 
The simple moving average is one of many time series analysis technique. Time series analysis is a method of timely structured data processing to find statistics or important characteristics for many reasons. The simple moving average shows stock trend by calculating the average value of stock prices on specific duration. The prices that are used are closing prices at the end of the day. This technique can avoid noises and therefore smooth the trend movement.
The main objective of financial news classification is to classify and calculate each news’ sentiment value. The positive news is marked by sentiment value which is greater than 0, while negative news is marked by less than 0 sentiment value. If there are news having 0 sentiment value, they will be omitted as their neutralism does not affect the stock trend.
Machine learning using artificial neural network algorithm is used to predict a stock trend. The artificial neural network uses three features along with one label. The three features are a simple moving average distance which is a subtraction of long-term and short-term simple moving average, the total value of positive sentiment value for one-day news, and the total value of negative sentiment value for one-day news. Stock trend label is used and classified as uptrend and downtrend. On one hand, learning component is done by a background process. On the other hand, prediction component is foreground process which is seen and interact with the user.
VISUALIZING AND UNDERSTANDING RECURRENT NETWORKS
-Andrej Karpathy, Justin Johnson, Li Fei-Fei 
Character-level language models have been used as an interpretable test bed for analyzing the predictions, representations training dynamics, and error types present in Recurrent Neural Networks. In particular, the qualitative visualization experiments, cell activation statistics and comparisons to finite horizon n-gram models demonstrate that these networks learn powerfully, and often interpretable long-range interactions on real-world data.
The error analysis broke down cross entropy loss into several interpretable categories and allowed us to illuminate the sources of remaining limitations and to suggest further areas for study.
In particular, it was found that scaling up the model almost entirely eliminates errors in the n-gram category, which provides some evidence that further architectural innovations may be needed to address the remaining errors.
LSTM: A Search Space Odyssey
-Klaus Greff, Rupesh K. Srivastava, Jan Koutn´ık, Bas R. Steunebrink, Jurgen Schmidhuber 
This paper reports the results of a large-scale study on variants of the LSTM architecture. We conclude that the most commonly used LSTM architecture (vanilla LSTM) performs reasonably well on various datasets. None of the eight investigated modifications significantly improves performance.
The forget gate and the output activation function are the most critical components of the LSTM block. Removing any of them significantly impairs performance. We hypothesize that the output activation function is needed to prevent the unbounded cell state to propagate through the network and destabilize learning. This would explain why the LSTM variant GRU can perform reasonably well without it: its cell state is bounded because of the coupling of input and forget gate.
The analysis of hyperparameter interactions revealed no apparent structure. Furthermore, even the highest measured interaction (between learning rate and network size) is quite small. This implies that for practical purposes the hyperparameters can be treated as approximately independent. In particular, the learning rate can be tuned first using a fairly small network, thus saving a lot of experimentation time.
Neural networks can be tricky to use for many practitioners compared to other methods whose properties are already well understood. This has remained a hurdle for newcomers to the field since a lot of practical choices are based on the intuitions of experts, as well as experiences gained over time. With this study, we have attempted to back some of these intuitions with experimental results. We have also presented new insights, both on architecture selection and hyperparameter tuning for LSTM networks which have emerged as the method of choice for solving complex sequence learning problems. In future work, we plan to explore more complex modifications of the LSTM architecture.
The difficulty of training recurrent neural networks
-Razvan Pascanu, Tomas Mikolov, Yoshua Bengio 
We provided different perspectives through which one can gain more insight into the exploding and vanishing gradients issue. We put forward a hypothesis stating that when gradients explode we have a cliff-like structure in the error surface and devise a simple solution based on this hypothesis, clipping the norm of the exploded gradients.
The effectiveness of our proposed solutions provides some indirect empirical evidence towards the validity of our hypothesis, though further investigations are required. In order to deal with the vanishing gradient problem, we use a regularization term that forces the error signal not to vanish as it travels back in time.
This regularization term forces the Jacobian matrices ∂xi ∂xi−1 to preserve norm only in relevant directions. In practice, we show that these solutions improve the performance of RNNs on the pathological synthetic datasets considered, polyphonic music prediction and language modeling.
Deep Sparse Rectifier Neural Networks
-Xavier Glorot Antoine Bordes Yoshua Bengio 
Sparsity and neurons operating mostly in a linear regime can be brought together in more biologically plausible deep neural networks. Rectifier units help to bridge the gap between unsupervised pre-training and no pre-training, which suggests that they may help in finding better minima during training.
This finding has been verified for four image classification datasets of different scales and all this in spite of their inherent problems, such as zeros in the gradient, or ill-conditioning of the parameterization. Rather sparse networks are obtained (from 50 to 80% sparsity for the best generalizing models, whereas the brain is hypothesized to have 95% to 99% sparsity), which may explain some of the benefits of using rectifiers.
Rectifier activation functions have shown to be remarkably adapted to sentiment analysis, a text-based task with a very large degree of data sparsity. This promising result tends to indicate that deep sparse rectifier networks are not only beneficial to image classification tasks and might yield powerful text mining tools in the future.
Furthermore, rectifier activation functions have shown to be remarkably adapted to sentiment analysis, a text-based task with a very large degree of data sparsity. This promising result tends to indicate that deep sparse rectifier networks are not only beneficial to image classification tasks and might yield powerful text mining tools in the future.
Stock Market Trends Prediction after Earning Release
-Chen Qian, Wenjie Zheng 
As known to the public, the stock market is known as a chaotic system and it has been proved that even model built with empirical key features could still result in low accuracy. In our work, we tried to limit our scope to earnings release day, and it turned out that we could build models achieving around 70% prediction accuracy.
To build the model, We take financial statistics collected from company’s quarterly earnings report, market surprise due to consensus expectations in terms of digital data, and sentiment analysis of relevant articles from mainstream media of financial professionals as two sets of input features, and make stock market movements prediction in after-hour period and trend in the day after the release day. SVM and LWLR model outperforms other models as shown by experiments, as they control the correlation among data, which was discussed in Section VI.
However, due to the limited number of company choices, we have small data size (300 samples) which could lead to high bias and overfitting. The stock price is not only affected by certain financial features, consensus news, but also company direction and future business guidance, which are difficult to be digitized.
Predicting Stock Trends through Technical Analysis and Nearest Neighbor Classification
-Lamartine Almeida Teixeira Adriano Lorena Inácio de Oliveira 
Tech Examination is built on the philosophies of the Dow Theory and practices the past of prices to forecast upcoming actions. The method used in tech examination can be enclosed as an outline credit problem, where the ideas are resulting from the history of values and the output is an estimate of the price or an estimate of the prices trend.
The most significant evidence of this type of examination is that the marketplace action reductions everything. It means the specialist believes that anything that can perhaps affect the marketplace is already reflected in the prices, as well as that all the new evidence will be directly reflected in those prices. As an import, all the technician needs is to analyze the past of prices.
The main gears of the tech examination are the capacity and price charts. Based on the data of values and size the tech pointers are built. Tech pointers are math formulations that are applied to the price or volume statistics of a safekeeping for demonstrating some aspect of the association of those amounts.
CHAPTER 3 DATA AND TOOLS
3.1 Data Used
3.1.1 Choosing the Dataset
For this project, we chose the Google stocks. The Google stocks is a large index traded on the New York stock exchange. All companies in the index are large publicly traded companies, leaders in each of their own sectors. The index covers a diverse set of sectors featuring companies such as Microsoft, Visa, Boeing, and Walt Disney. It is important to use a predefined set of companies rather than a custom selected set so that we do leave ourselves open to methodology errors or accusations of fishing expeditions. If we had selected a custom set of companies, it could be argued that the set was tailored specifically to improve our results. Since the aim of the project is to create a model of stock markets in general. Google was chosen because it is well known. The components provided a good balance between available data and computational feasibility.
3.1.2 Gathering the Datasets
A primary dataset will be used throughout the project. The dataset will contain the daily percentage change in stock price. Luckily, daily stock price data is easy to come by. Google and Yahoo both operate websites which offer a facility to download CSV files containing a full 14 daily price history. These are useful for looking at individual companies but cumbersome when accessing large amounts of data across many stocks. For this reason, Quandl was used to gather the data instead of using Google and Yahoo directly. Quandl is a free to use website that hosts and maintains vast amounts of numerical datasets with a focus specifically on economic datasets, including stock market data which is backed by Google and Yahoo. Quandl also provides a small python library that is useful for accessing the database programmatically. The library provides a simple method for calculating the daily percentage change daily in prices.
For instance, data we gather for a Monday will be matched with, and try to predict, Tuesday’s trend. This dataset was then saved in CSV format for simple retrial as needed throughout the project. This dataset containing the daily trends of companies will serve as the core dataset that will be used in most experiments later in the report.
Abbildung in dieser Leseprobe nicht enthalten
TAbbildung in dieser Leseprobe nicht enthalten
The most interesting task is to predict the market. So many methods are used for completing this task. Methods vary from very informal ways to many formal ways a lot. This tech. are categorized as:
- Prediction Methods
- Traditional Time Series
- Technical Analysis Methods
- Machine Learning Methods
- Fundamental Analysis Methods
- Deep Learning
The criteria for this category are the kind of tool and the kind of data that these methods are consuming in order to predict the market. What is mutual to the technique is that they are predicting and hence helping the market's future behavior.
4.1 Technical Analysis Methods
Technical analysis is used to attempt to forecast the price movement of virtually any tradable instrument that is generally subject to forces of supply and demand, including stocks, bonds, futures and currency pairs. In fact, technical analysis can be viewed as simply the study of supply and demand forces as reflected in the market price movements of a security. It is most commonly applied to price changes, but some analysts may additionally track numbers other than just prices, such as trading volume or open interest figures.
Over the years, numerous technical indicators have been developed by analysts in attempts to accurately forecast future price movements. Some indicators are focused primarily on identifying the current market trend, including support and resistance areas, while others are focused on determining the strength of a trend and the likelihood of its continuation. Commonly used technical indicators include trendlines, moving averages and momentum indicators such as the moving average convergence divergence (MACD) indicator.
Technical analysts apply technical indicators to charts of various timeframes. Short-term traders may use charts ranging from one-minute timeframes to hourly or four-hour timeframes, while traders analyzing longer-term price movement scrutinize daily, weekly or monthly charts.
4.2 Fundamental Analysis Techniques
Fundamental analysis uses real, public data in the evaluation a security's value. Although most analysts use fundamental analysis to value stocks, this method of valuation can be used for just about any type of security. For example, an investor can perform fundamental analysis on a bond's value by looking at economic factors such as interest rates and the overall state of the economy. He can also look at information about the bond issuer, such as potential changes in credit ratings.
For stocks and equity instruments, this method uses revenues, earnings, future growth, return on equity, profit margins, and other data to determine a company's underlying value and potential for future growth. In terms of stocks, fundamental analysis focuses on the financial statements of the company being evaluated. One of the most famous and successful fundamental analysts is the so-called "Oracle of Omaha", Warren Buffett, who is well known for successfully employing fundamental analysis to pick securities. His abilities have turned him into a billionaire.
4.3 Traditional Time Series Prediction
Time series analysis can be useful to see how a given asset, security or economic variable changes over time. It can also be used to examine how the changes associated with the chosen data point compare to shifts in other variables over the same time period.
For example, suppose you wanted to analyze a time series of daily closing stock prices for a given stock over a period of one year. You would obtain a list of all the closing prices for the stock from each day for the past year and list them in chronological order. This would be a one-year daily closing price time series for the stock.
Delving a bit deeper, you might be interested to know whether the stock's time series shows any seasonality to determine if it goes through peaks and valleys at regular times each year. The analysis in this area would require taking the observed prices and correlating them to a chosen season. This can include traditional calendar seasons, such as summer and winter, or retail seasons, such as holiday seasons.
Alternatively, you can record a stock's share price changes as it relates to an economic variable, such as the unemployment rate. By correlating the data points with information relating to the selected economic variable, you can observe patterns in situations exhibiting dependency between the data points and the chosen variable.
4.4 Machine Learning Methods
Various sectors of the economy are dealing with huge amounts of data available in different formats from disparate sources. The enormous amount of data, known as Big Data, is becoming easily available and accessible due to the progressive use of technology. Companies and governments realize the huge insights that can be gained from tapping into big data but lack the resources and time required to comb through its wealth of information. In this regard, Artificial Intelligence (AI) measures are being employed by different industries to gather, process, communicate and share useful information from data sets. One method of AI that is increasingly utilized for big data processing is Machine Learning.
The various data applications of machine learning are formed through a complex algorithm or source code built into the machine or computer. This programming code creates a model which identifies the data and builds predictions around the data it identifies. The model uses parameters built into the algorithm to form patterns for its decision-making process. When new or additional data becomes available, the algorithm automatically adjusts the parameters to check for a pattern change, if any. However, the model shouldn’t change.
How machine learning works can be better explained by an illustration in the financial world. Traditionally, investment players in the securities market like financial researchers, analysts, asset managers, individual investors scour through a lot of information from different companies around the world to make profitable investment decisions. However, some pertinent information may not be widely publicized by the media and may be privy to only a select few who have the advantage of being employees of the company or residents of the country where the information stems from. In addition, there’s only so much information humans can collect and process within a given time frame. This is where machine learning comes in.
An asset management firm may employ machine learning in its investment analysis and research area. Say the asset manager only invests in mining stocks. The model built into the system scans the World Wide Web and collects all types of news events from businesses, industries, cities, and countries, and this information gathered comprises the data set. All the information inputted in the data set is information that the asset managers and researchers of the firm would not have been able to get using all their human powers and intellects. The parameters built alongside the model extracts only data about mining companies, regulatory policies on the exploration sector, and political events in select countries from the data set. Say, a mining company XYZ just discovered a diamond mine in a small town in South Africa, the machine learning app would highlight this as relevant data. The model could then use an analytics tool called predictive analytics to make predictions on whether the mining industry will be profitable for a time period, or which mining stocks are likely to increase in value at a certain time. This information is relayed to the asset manager to analyze and make a decision for his portfolio. The asset manager may make a decision to invest millions of dollars into XYZ stock.
Abbildung in dieser Leseprobe nicht enthalten
Fig[4.1]: Machine Learning Analysis Curve
In the wake of an unfavorable event, such as South African miners going on strike, the computer algorithm adjusts its parameters automatically to create a new pattern. This way, the computational model built into the machine stays current even with changes in world events and without needing a human to tweak its code to reflect the changes. Because the asset manager received this new data on time, he is able to limit his losses by exiting the stock. Machine learning is used in different sectors for various reasons. Trading systems can be calibrated to identify new investment opportunities. Marketing and e-commerce platforms can be tuned to provide accurate and personalized recommendations to their users based on the users’ internet search history or previous transactions. Lending institutions can incorporate machine learning to predict bad loans and build a credit risk model. Information hubs can use machine learning to cover huge amounts of news stories from all corners of the world. Banks can create fraud detection tools from machine learning techniques. The incorporation of machine learning in the digital-savvy era is endless as businesses and governments become more aware of the opportunities that big data presents.
4.5 DEEP LEARNING
An artificial intelligence function that imitates the workings of the human brain in processing data and creating patterns for use in decision making. Deep learning is a subset of machine learning in Artificial Intelligence (AI) that has networks which are capable of learning unsupervised from data that is unstructured or unlabeled. Also known as Deep Neural Learning or Deep Neural Network.
BREAKING DOWN 'Deep Learning'
The digital era has brought about an explosion of data in all forms and from every region of the world. This data, known simply as Big Data, is gotten from sources like social media, internet search engines, e-commerce platforms, online cinemas, etc. This enormous amount of data is readily accessible and can be shared through fine tech applications like cloud computing. However, the data, which normally is unstructured, is so vast that it could take decades for humans to comprehend it and extract relevant information. Companies realize the incredible potential that can result from unraveling this wealth of information and are increasingly adapting to Artificial Intelligence (AI) systems for automated support.
One of the most common AI techniques used for processing Big Data is Machine Learning. Machine learning is a self-adaptive algorithm that gets better and better analysis and patterns with experience or with newly added data. If a digital payments company wanted to detect the occurrence of or potential for fraud in its system, it could employ machine learning tools for this purpose. The computational algorithm built into a computer model will process all transactions happening on the digital platform, find patterns in the data set, and point out any anomaly detected by the pattern.
Deep learning, a subset of machine learning, utilizes a hierarchical level of artificial neural networks to carry out the process of machine learning. The artificial neural networks are built like the human brain, with neuron nodes connected together like a web. While traditional programs build analysis with data in a linear way, the hierarchical function of deep learning systems enables machines to process data with a non-linear approach. A traditional approach to detecting fraud or money laundering might rely on the amount of transaction that ensues, while a deep learning non-linear technique to weeding out a fraudulent transaction would include time, geographic location, IP address, type of retailer, and any other feature that is likely to make up a fraudulent activity. The first layer of the neural network processes a raw data input like the amount of the transaction and passes it on to the next layer as output. The second layer processes the previous layer’s information by including additional information like the user's IP address and passes on its result. The next layer takes the second layer’s information and includes raw data like geographic location and makes the machine’s pattern even better. This continues across all levels of the neuron network until the best and output is determined.
Using the fraud detection system mentioned above with machine learning, we can create a deep learning example. If the machine learning system created a model with parameters built around the amount of dollars a user sends or receives, the deep learning method can start building on the results offered by machine learning. Each layer of its neural network builds on its previous layer with added data like a retailer, sender, user, social media event, credit score, IP address, and a host of other features that may take years to connect together if processed by a human being. Deep learning algorithms are trained to not just create patterns from all transactions, but to also know when a pattern is signaling the need for a fraudulent investigation. The final layer relays a signal to an analyst who may freeze the user’s account until all pending investigations are finalized.
Deep learning is used across all industries for a number of different tasks. Commercial apps that use image recognition, open source platforms with consumer recommendation apps, and medical research tools that explore the possibility of reusing drugs for new ailments are a few of the examples of deep learning incorporation.
4.5.1 Artificial Neural Networks (ANN)
A computing system that is designed to simulate the way the human brain analyzes and process information. Artificial Neural Networks (ANN) is the foundation of Artificial Intelligence (AI) and solves problems that would prove impossible or difficult by human or statistical standards. ANN has self-learning capabilities that enable it to produce better results as more data becomes available.