CALIFORNIA STATE UNIVERSITY, NORTHRIDGE

STOCK MARKET PREDICTION USING MACHINE LEARNING

A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science

in Software Engineering

By

Prince Vipulbhai Patel

August 2021
The thesis of Prince Vipulbhai Patel is approved:

_______________________________________ _________________

Jeff Weigley, Ph.D. Date

_______________________________________ _________________

Katya Mkrtchyan, Ph.D. Date

_______________________________________ _________________

Robert Mcilhenny, Ph.D., Chair Date

California State University, Northridge

ii
 Table of Contents

Signature page ii

Abstract iv

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.2 LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2.1 Inner working of LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 Creating data-set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

5 Implementing algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5.1 Support Vector Machine implementation . . . . . . . . . . . . . . . . . . . . 20

5.2 Long Short Term Memory(LSTM) implementation . . . . . . . . . . . . . . 25

6 Testing result and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29

6.1 SVM result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

6.2 LSTM result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

iii
 ABSTRACT

STOCK MARKET PREDICTION USING MACHINE LEARNING

By

Prince Vipulbhai Patel

Masters of Science in Software Engineering

Stock market trading has gained popularity in today's world with the advancement in

technology and social media. With the help of today’s technology we can aim to predict the stock

market for the future value of stocks. To make informed predictions, time series analysis is used

by most stock brokers around the world. This paper explains and analyzes the prediction of a

stock by using machine learning. In this paper, I propose a machine learning approach that will

be trained from the available stock data by using acquired knowledge for a prediction with

accuracy. In this context, the study will use a machine learning technique called Support Vector

Machine (SVM) and Long Short term memory (LSTM) to predict stock prices.

iv
 Chapter 1

Introduction

1.1 Background

The stock market is a collection of markets and exchanges where there is a facilitation of

selling, buying, and dispatching of shares for public companies. These financial ventures are

formal institutionalized exchanges, also known as an over the counter marketplace operated

under a defined set of strict regulations. The stock market is also a platform for buyers and

sellers to interact and form transactions. Benefits of the stock market are a controlled and

monitored environment for hundreds of thousands of people to buy and sell shares with

transparency and fair pricing practices. There are many vital aspects the stock market considers

such as efficient pricing, fair deal security transactions, maintenance of liquidity, valid

transactions, supporting eligible participants, balancing regulation, and protecting investors.

On other hand, we have the technical analysis portion of the market analysis of stocks. This

happens by studying the statistics that are generated by market activity which includes factors

such as past prices and volumes. In technical analysis, technical traders obtain information

needed for trading from charts. This differs from fundamental analysis where fundamental

traders consider the factors outside of price fluctuations from the asset itself.

In more recent times there has been an increasing dominance of machine learning in different

industries that have provided some resourceful information to an innumerable amount of traders

which can apply machine learning techniques to the field. A multitude of these learning

techniques have produced promising and favorable results

1
1.2 Statement of Problem

The forecasting of stocks indicates that it is a very strenuous process because of the market

volatility that depends upon an accurate forecast model. The stock market inconsistencies tend to

fluctuate which does make an impact on an investor's decisions. The stock market prices are

increasingly dynamic and susceptible to rapid changes. This is because the underlying aspects of

the financial domain can also be attributed to the mix of the various unknown parameters. These

parameters include previous day closing prices, P/E ratio and the other unknown factors such as

the economy, election results, and rumors. There have been numerous attempts to predict stock

prices using machine learning. The focus of each research method can vary in three distinctive

ways:

1- The targeted price change can be near-term which is less than a minute, short-term,

meaning tomorrow to a few days later, and lastly long term which is considered months later.

2- The set of stocks in a particular industry then branching to all general stocks.

3- The predictors used can range from a global news and economy trend, to particular

characteristics of the company, and lastly to purely time series data of the stock price.

The probable stock market prediction target can include the future stock price, the volatility of

the prices, and market trend. In the stock market prediction there are two types of predictions that

include a dummy prediction or a real time prediction which is used in the stock market prediction

system. In dummy predictions, there is a defined set of rules to predict the future price of shares

by calculating the average price. In the real time predictions, there is a compulsory use of the

internet and monitoring of the current price of shares of a company.

2
1.3 Approach

This paper will analyze financial data predictor programs in which there will be a dataset

storing all historical stock prices and this data will be treated as a training set for the program.

The main purpose of the prediction is to reliably and more accurately reduce the amount of

uncertainty associated with investment decision making in the stock market. The accuracy

achieved gives an idea about the stock's future price based on the prediction made by the

algorithm.

The stock market arguably follows the Random Walk Theory, this implies that in stock market

prediction tomorrow's possible stock market value is also today's value. The Random Walk

Theory is defined as the influxing changes of stock prices that have similar distribution rates and

are independent from one another. This means assuming that the past trends and movements of

the market stock prices cannot be used to predict other future movements. The Random Walk

Theory suggests that since stocks are unpredictable, all methods of predicting stock prices in the

long run are futile because of the assumed additional risks. Random Walk Theory was

popularized in 1973 by Burton Malkiel, a Princeton economist. Since then, more sophisticated

computer algorithms are utilized to identify and exploit the different trends of stock market

prices. Even if the trends spotted last for a small amount of time, like a second, their existence

overrides the Random Walk Theory.

3
 Chapter 2

Theory

2.1 SVM

Support vector machines are useful for relapse issues and arrangements. It is a learning model

that is impacted by the idea of hyperplanes as choice limits. In the two dimensional space of the

hyperplane is a line. When the model is being built, support vector machines create calculations

for new guides while allocating a hyperplane to a yield and classes. Support vector machine

models are an example of space models created with an emphasis that various classes are a

greatest separation from choice limits. By using a five crease cross approval procedure to

analyze estimates, support vector machines are more powerful allowing for higher

measurements. This analysis makes the algorithm effective in machine learning. An example

would be when the quantity of measurements used surpasses the quantity of tests. Support vector

machines have a higher cost and are more accurate because of parameters using gamma, an

accessible svm classifier, and regularization parameters. The parameters choose the amount of

misclassification that is allowed, rbf or poly decides the hyperplane learning, and gamma shows

how far the impact of preparation tests show.

4
 Figure 2.1 The svm decision making boundary

The advantages of support vector machine are :-

- Effective in high dimensional spaces. SVMs are very good when we have no idea on the

data. Works well with even unstructured and semi structured data.

- If the number of dimensions is greater than the number of samples, it is still effective. It

scales relatively well to high dimensional data.

- Uses a set of training points in the decision function called the support vectors so it is

also memory efficient.

- It is versatile meaning different kernel functions can be specified for the decision

function. Common kernels are provided but it is also possible to specify custom kernels.

5
The disadvantages of support vector machine are :-

- If the number of features is much greater than the number of samples, avoiding

over-fitting in choosing kernel functions and regularization terms is crucial. Choosing a

“good” kernel function is not easy.

- SVMs do not provide probability estimates, these are calculated using an expensive five

fold cross validation.

- Since the final model is not so easy to see, we cannot do small calibrations to the model

hence its tough to incorporate the business logic.

- Larger training dataset takes longer training time.

6
2.2 LSTM

A key difference LSTM has versus traditional RNN’s is a gating mechanism. An example that

demonstrates the difference in LSTM’s capability to save long term memory. This can provide

benefits in sequential tasks or natural language processing. This can be demonstrated when

there’s network generated text given for an analysis. Then the RNN model has to generate the

next word. Network generated text is predicted when information is given, such as a name or

object and the model then generates the same model or word again when later predicting the next

word.

Since RNN’s generally have an issue with having short term memory the average RNN is only

able to utilize information from the text that was shown within the previous few sentences. The

issue of short term memory can provide some challenges.

This differs from a LSTM that is able to store information from earlier periods of time that

provides better accuracy when generating contextual information predictions.

2.2.1 Inner workings of the LSTM

LSTM cells have beneficial mechanisms that act as a gate. In the average RNN cell there is a

tanh activation function in order to gather output and a new hidden state. This is different from

LSTM in that it has a more complicated arrangement. This is because the LSTM cells need three

pieces of information which are the current input data, the short term memory from the prior cell

and the long term memory cell.

7
Short term memory can also be called a hidden state and long term memory can be referred to as

a cell state.

The cell uses the function of a gate to generate saved information or passed up at each time step

before transferring short term and long term information to another cell.

These gates are supposed to carefully pick and get rid of any information that’s irrelevant. This

means the gates store beneficial information. These gates are known as the Forget Gate, the Input

Gate, and the Output Gate. Each gate has their own mechanism to differentiate them.

Input Gate:-

The input gate decides what information will be stored in long term memory. It can only work

with the information provided by current input and short term memory from the previous time

step. This is why it then has to refine beneficial information from information that isn’t

beneficial.

There’s two ways for this to happen, the first way is a method to pick information that decides

discarded information or stored information. To produce this process we can bypass short term

memory and current input to a sigmoid function. The sigmoid function transforms values

between zero and one with zero demonstrating the unimportance of information and one

demonstrating usable information. This is favorable in deciding what values should be used and

saved. The benefits of the sigmoid function having weights is discarding the more unimportant

features and letting usable information pass.

8
The second method uses current input and short term memory to trigger an activation function

that provides regulation in network processes.

Next, the output methods are multiplied and the final solution will show the information that will

be kept in the long term memory and used as output information.

Forget Gate:

The forget gate determines the long term memory information that will be stored or rejected. By

multiplying the received long term memory from a forget vector created by short term memory

and current input. Similar to the first method using the Input gate, the forget vector is also

particular and selective in picking information. When gathering information for the forget vector,

the current input and the short term memory pass between a sigmoid function. The process is

very similar to the Input Gate and can also be described as consisting of a scale of zero to one

that is multiplied with the factor of long term memory, influencing the retained parts of long term

memory information.

The outputs given from the Input gate and Forget gate receive a pointwise inclusion that creates a

new kind of long term memory that gets passed onto the next cell. The new long term memory is

then incorporated into the Output gate.

9
Output Gate:-

The output gate uses previous short term memory, new long term memory, and current input to

create a new type of short term memory which will be given to the cell in the time step after. The

first step that happens is current input and short term memory are given to the sigmoid function

with different weights to create the last filter. The new long term memory is then put through an

activation function. The output gathered from these two processes gets multiplied to create the

newest short term memory.

The short term and long term memory created gets transferred over onto the next cell to continue

the process as it repeats. The product and output of each time step can be derived from the short

term memory.

The advantages of LSTM are :-

- It can model a collection of records so that each pattern can be assumed to be dependent

on the previous one making it recurring.

- It is also used with convolutional layers to extend the powerful pixel neighbourhood.

The disadvantage of LSTM are :-

- It has Gradient exploding and vanishing problems.

- Training LSTM is a tough task due to the way it is designed.

- It cannot system very lengthy sequences if the usage of Tanh or Relu as an activation

feature.

10
Figure 2.2 Working of LSTM

11
 Chapter 3

Methodology

● Step1- Find the data from the internet. The preferred website is yahoo finance for the

uniformity of data. The goal is to download the data from the internet. We are predicting

the financial market value of any stock. The share value up to the closing date is then

downloaded from the site.

● Step2- The data that is downloaded from the yahoo finance app can be downloaded in the

form of a csv file. The values (comma separated values) of any stocks can be loaded in

the algorithm.

● Step3- The data in the csv file is modified to the desired format so that the algorithm can

process it.

● Step4- The algorithms are then initiated in the terminal separately. The two algorithms

are coded separately making them able to be run separately for different data.

● Step5- The code that is selected will run and in the end will output all the accuracy

values. The mean accuracy and standard deviation is printed after all the accuracies are

calculated.

12
 Chapter 4

Creating data-set

For training and testing predictive algorithms, a database is compiled for training and testing

purposes. The database should be having minimum deviation and should account for normal

behaviour. It was important to select a time frame where the stock market did not show any

sudden jump or dip in prices. Often there are times of economic collapse in certain sectors,

industries and countries. The extent of these collapses might not be foreseen or predictable.

For the training and testing of the algorithms and models, a time frame of January 2009 to

October 2018 was chosen. From 2007-2008 the stock market took a hit from the financial crisis.

Yahoo Finance provides a Historical Database which can be utilized for the training of the

model. It provides a historical database of a stock also using the following parameters:

● Date: Date of the values collected

● Open: Opening value of the stock

● High: Highest value achieved by the stock on that date

● Low: Lowest value achieved by the stock on that date

● Close: Closing value of the stock for that date

● Adj Close: Closing value of the stock before market opens next day, adjusted by

considering corporate actions and distributions

● Volume: Total number of shares that changed hands during that day

13
 Fig 4.1 Creating the dataset

The dataset creation is an important step because the data that is fed to the algorithm

needs to be in the format that can be processed. It is a separate python file named

DatasetCreation.py. Data Creation makes that dta more organized and refined while also

being able to cut down the amount of time needed to process the output.

14
 Figure 4.2 Code using the dataset

The collected stock price historical data from Yahoo Finances for the following 9 companies

belonging to same sector of Technology:

● Apple

● Amazon

● Google

● Microsoft

15
Figure 4.3 Screenshot of apple csv file from yahoo finance

16
Using the database to compile information derived from these parameters. The key terms used in

deriving the input parameters are:

Fig 4.4 The parameters implemented in the code

● Momentum: It is the measure of detecting the flow of a parameter such as stock price in

our case. It is defined as +1 if the price has increased and -1 if the price has been dropped

from yesterday.

● Volatility: It determines the fluctuations in the same parameter at different times. For our

case it will be determined by the difference between yesterday’s price and today’s price

divided by yesterday’s price.

Hence for the input dataset the following parameters were found out from the above parameters:

● Index Momentum: It determines the momentum of the market.

● Index Volatility: This helps in determining the fluctuations in the market.

● Sector Momentum: This parameter considers other companies in the same sector and

calculates the momentum of these companies for the last 5 days.

17
● Stock Momentum: This is average of the last 5 days of momentum of the respective

company

● Stock Price Volatility: This is the average of the last 5 days of stock price of the

respective company.

Fig 4.5 parameters used to create input dataset

18
Fig 4.6 Date, Close, Change, Momentum used in the algorithm for the stock’s data

19
 Chapter 5

Implementing the algorithms

5.1 Support Vector Machine implementation:

This algorithm was implemented in python using the scikit-learn library for machine learning.

The implementation was done in the following steps-

● Importing the library and csv data files

● Creating a SVM classifier

● Defining the kernel and setting its parameters

● Predicting and comparison with target for accuracy calculation

Figure 5.1 SVM print statements

20
Fig 5.2 SVM algorithm implementation with the libraries used

21
 In this case, since the stock market data is non linear, the rbf kernel was used and kept the

hyperparameters C and gamma at their default value of 1 and scale respectively. Higher values of

C and gamma, causes the decision boundary to become more curvy and variance to increase

thereby causing higher chances of overfitting. Accuracy can be improved by tuning the

parameters using the function gridSearchCV() in sklearn.model selection. This may take a lot of

time depending on the number of training samples. Once the classifier is trained by mapping the

given features and labels, prediction is made using the test data which consists of approximately

20% of the total data. This is compared with the target data present in the output file and

accuracy is ascertained.

Fig 5.3 SVM accuracy calculation

22
 SVM implementation is done using the Scikit Learn Library. This is an inbuilt function

for machine learning and preprocessing steps like data normalization. This is the key element

that makes the model functional with the help of machine learning. Python language is used to

import the library which helps with the training of the datasets. The test dataset is used to which

is processed for the output.

23
Figure 5.4Features used by SVM

24
5.2 Long Short Term Memory(LSTM) implementation

Implementation of LSTM was done using the following steps-

- Dataset creation for LSTM:

● Converting the time series into supervised learning problem:

● The data is divided into input set and output set.

● A value is defined for time step. Using this time step the data is shifted

and these two series are concatenated to get the output set.

● Converting the time series into stationary data:

● Stationary data is easier to model and is better for forecasting.

● It is obtained by removing the trend from the data which is achieved by

differencing the data.

● Observations are transformed to specific scale:

● LSTM works on the data which is within the scale of activation function

of the network.

● Since the default activation function for LSTM is hyperbolic tangent

(tanh) which has output range from −1 to +1 which is ideal for time series

data.

- Model Development:

● LSTM is a type of Recurrent Neural Network (RNN). This type of neural network

is useful when remembering over a long sequence of data and it doesn't depend on

the window lagged dataset as input.

● LSTM layer takes 3 inputs:

● Samples: The rows of the input dataset having independent observations.

25
 ● Time steps: These are time steps of the parameters for the input dataset.

● Features: The parameters considered and observed for the input dataset to

predict the output.

Fig 5.5 LSTM implementation with the libraries used

While compiling the network we need to mention loss function and also the optimization

algorithm. So, we are using mae as loss function and ADAM as optimization algorithm. ADAM

will select a suitable learning rate for the network. And then this model is fit to the training data

and we predict the output. Then, we calculated the accuracy of the model.

26
Fig 5.6 LSTM code flow

27
 The LSTM algorithm here is using Numpy and Pandas for dataset manipulation, Scikit learn

for machine learning, Keras for neural network creation and Matplotlib for visualization. All

these are used in order to process the data and output the results.

Figure 5.7 LSTM print statement

28
 Chapter 6

Testing result and analysis

To make an accurate comparison between the two algorithms used, the same algorithms were

run multiple times. The test data is fed in the algorithm which runs each of the files separately.

The simplified test data is used which is processed and Each algorithm is implemented 30 times

which gives out an accuracy for each time. The mean accuracy and standard deviation is

calculated at the end. These numbers give a better understanding of the overall result by

narrowing down the result to one number which eventually helps to compare the two algorithms

used. For each run different training and testing dataset is used. The results for the two

algorithms are listed below.

6.1 Svm result

For 30 runs of SVM algorithm, the algorithm got approximately 50.73 mean accuracy with

0.308 standard deviation. This shows that SVM performance is consistent for 30 runs; this is due

to the nature of the SVM algorithm. Algorithm will keep training until it can classify maximum

testing data resulting in higher accuracy. The SVM code predicts the value of the stock for the

next 30 days and then prints out a separate accuracy. All the accuracies in the end are taken into

consideration and then the mean accuracy is printed out.

For other numbers of runs the algorithm performs similarly and the accuracy is not affected

much by increasing the number of times the algorithm runs. This tells that the accuracy is

reliable for that given data.

29
 The accuracy is relatively low compared to LSTM’s accuracy. It is because the dataset used

has higher variability. Also the design of the algorithm is such that the accuracy is less effective

if the dataset is similar to the one used. With the result of the accuracy achieved, the percentage

of accuracy gives a good idea of the future value.

The standard deviation is relatively low but a little higher than LSTM in this particular

prediction for this dataset chosen. This means the fluctuation from the average returns is higher.

Since the values of the standard deviation change with the change in the data set, this data set

results in less fluctuation from the average values. It is considered to be relatively low which

means the prediction fluctuation is less and hence the results are accurate.

30
Figure 6.1 SVM results

31
6.2 LSTM result

For 30 runs of LSTM algorithm the accuracy was 62.29 which is the mean accuracy with

standard deviation of 0.256. The fluctuation is minimal proving the algorithm is reliable and

there is overall performance improvement compared to SVM. The accuracy is calculated in the

similar way, predicting the accuracy for each stock for the next 30 days and then the mean

accuracy is calculated.

As you can see LSTM performs well compared to SVM as it uses powerful methods to predict

the sequence as it can store the past data. This is beneficial in the stock market prediction

because the stock data has values based upon which the future value of the stock is predicted.

Stock market prediction is a time series problem as mentioned before, so the LSTM algorithm

can be used when the problem is viewed from a time series perspective. A traditional

RNN(Recurrent neural network) has a problem of vanishing gradient point, whereas LSTM does

not have that problem that helps it to function better for the time series problem that has been

tried to solve.

The Standard deviation achieved here is relatively lower making the returns fluctuate less

from the average returns. LSTM has better results overall. According to the results that have

been displayed here, the integration of this algorithm using python works effectively and

efficiently in stock market predictions.

32
Figure 6.2 LSTM results

33
 Chapter 7

Conclusion

This paper observes that machine learning can be used to predict stock market prices with

accuracy. The result indicates historical data can be used to predict stock movement with

reasonable accuracy but the choice of algorithm depends on the requirement of parameters like

time, standard deviation and mean accuracy. For this implementation, the factors that affect stock

performance were incorporated. If a higher number of factors are used and after adequate

preprocessing and filtering of data, it is used to train the network model, then a higher accuracy

can be achieved. It is concluded by the results that LSTM works better in this case. Although

both of the algorithms have both pros and cons, each one can be used differently based on the

choices and requirements.

The implementation of SVM and LSTM using moving averages are done separately. For

future works, intraday prices can also be used to compare the values and to understand the

volatility of the stock, crude oil and gold prices in a better manner. The stock selling and buying

data can also be used to understand how the stock price and external factors such as a surge and

dip influence the buying and selling pattern. The model can be trained for volume deduction

meaning the stock prices which can be sold or purchased in a way which results in profit from

the stock This will help in developing a more accurate prediction. The models can also be

extended to provide live interactive predictions based on the user given data and then be used for

other forecasting problems like weather forecasting, disease forecasting and house price

forecasting. Some other factors that impact the stock prices can be included that can encourage

more accurate prediction eventually leading to increasing the chances of profits from the stock.

34
 The ranges of the experimental samples are not large enough. Large data sets would give

higher and more accurate predictions but at the same time would require computers that can

handle the large data computation and processing.

35
 References

[1] D. Karmiani, R. Kazi, A. Nambisan, A. Shah and V. Kamble, "Comparison of Predictive

Algorithms: Backpropagation, SVM, LSTM and Kalman Filter for Stock Market," 2019 Amity

International Conference on Artificial Intelligence (AICAI), 2019, pp. 228-234, doi:

10.1109/AICAI.2019.8701258.

[2] Han, Shuo and Chen, Rung-Ching (2007) "Using SVM with Financial Statement Analysis

for Prediction of Stocks ," Communications of the IIMA: Vol. 7 : Iss. 4 , Article 8. Available at:

Using SVM with Financial Statement Analysis for Prediction of Stocks

[3] Hegazy, Osman, et al. “A Machine Learning Model for Stock Market Prediction.”

[4] International Journal of Computer Science and Telecommunications, Dec. 2013.

[5] Le, Xuan Hien & Ho, Hung & Lee, Giha & Jung, Sungho. "Application of Long Short-Term

Memory (LSTM) Neural Network for Flood Forecasting." Jul. 2019

[6] Sutter, Lorran. “PredictStock-SVM.”, GitHub, 14 Sep. 2020,

github.com/LorranSutter/PredictStock-SVM.

[7] Hathidara, Ashutosh. “Stock-Prediction-Using-LSTM.”, GitHub, 20 Feb. 2020,

github.com/ashutosh1919/Stock-Prediction-using-LSTM.

[8]“1.4. Support Vector Machines¶.” Scikit, scikit-learn.org/stable/modules/svm.html.1.4.

support Vector Machines¶. scikit. (n.d.). https://scikit-learn.org/stable/modules/svm.html.

[9]Ck, Vignesh. “Pdf.” MCA Scholar, Department of MCA, School of CS & IT,Jain

(Deemed-to-Be) University, Bengaluru, Apr. 2020.

36
[10]Loye, Gabriel. “Long Short-Term Memory: From Zero to Hero WITH PYTORCH.”

FloydHub Blog, FloydHub Blog, 16 Aug. 2019,

blog.floydhub.com/long-short-term-memory-from-zero-to-hero-with-pytorch/

37