Table of Contents

1. Introduction............................................................................1
2. Literature review....................................................................3
2.1 Progress of stock price prediction......................................3
2.2 Time series model..............................................................4
2.3 Deep learning.....................................................................6
3. Methodology...........................................................................9
3.1 Data source........................................................................9
3.2 Methodology....................................................................10
4. Results and discussion..........................................................13
4.1 Arima process..................................................................13
4.2 GARCH process..............................................................14
4.3 LSTM process..................................................................15
4.4 Mixed model process.......................................................16
5. Conclusion and Implication..................................................19
5.1 Conclusion.......................................................................19
5.2 Implication.......................................................................20
Reference..................................................................................22
Appendix..................................................................................24
1. Introduction
As a high-risk and high-return market, the stock market has always been closely
watched by investors(Daubechies, I. 1992), and stock forecasting has always been a
research topic of great concern to researchers. In addition, the stock market is an
important part of my country's financial market, it reflects the operation of the national
economy, and the operation of the stock market has an important impact on the
operation of the national economy. Although the issue of predictability of stocks has
always been controversial, the study of stock forecasts still helps us understand the
laws of some market changes and development(Fama, E. F., & Blume, M. F. (1966)).
With the advancement of science and technology, a large amount of financial data has
been retained(Fama, E. F., & French, K. R. (1988)), providing a solid data foundation
for the analysis of the stock market; at the same time, the continuous development and
updating of algorithms has provided a powerful tool for people to analyze the stock
market(Faria, G., & Verona, F. (2018), Ferreira, M. I., & SantaClara, P. (2011),
Gençay, R., Selçuk, F., & Whitcher, B. (2002)).
Stock data is a classic time series. Many researchers have used time series models
for forecasting, such as ARIMA or GARCH models, but the assumptions of classic
time series models are relatively high. For example, the series needs to be stable and
linear. However, there are many factors that affect the stock price of stock data, which
makes the stock data itself not stable and linear. Although the difference method can be
used to smooth the sequence, the difference operation also causes data loss, which
makes the traditional time series model have greater limitations in forecasting. With
the development of computer science and artificial intelligence, more and more
researchers choose machine learning models for prediction, such as support vector
machines, perceptron models, etc. Because they can handle nonlinear data, especially
support vector machines. This model has a non-linear kernel function, so it has been
used by the majority of people in the industry for a period of time.
Stock market forecasting is the act of trying to determine the future value of shares of
companies listed on exchanges or other investment targets. The stock market has multi-

1
scale properties. The so-called multi-scale refers to the existence of multiple data at
different time intervals in the stock market. For example, taking stock price data as an
example, there are not only short-term real-time price data per second, every minute,
and hour, but also daily, weekly and even real-time price data. Monthly mid-term
average data, these prices may have different patterns of change at different scales.
Stock market forecasting research is usually based on stock market data under a certain
scale. After analysis, some patterns that recur in the data are extracted under a certain
scale, so as to predict the movement trend of the stock market under this scale.
Although data at different scales may have different changing laws, there is also a close
interaction between them. If the data at different scales can be considered
comprehensively, the state of the stock market can be described more accurately, and
thus better forecasting the stock market. Traditionally, stock market forecasts need to
be given by stock market researchers with deep knowledge and extensive analytical
experience. They make predictions on the future development direction of the stock
market and the degree of fluctuations based on multi-source heterogeneous data such as
foreign exchange, policies, events, and stock prices in the global economic data.
Stock data is a classic time series, and many researchers have used time series
models for forecasting, such as ARIMA or GARCH models(Goyal, A., & Welch,
I. ,2003, Goyal, A., & Welch, I. 2008, Brock, W., Lakonishok, J., & LeBaron, B.
1992), but the assumptions of classic time series models are relatively high, such as the
need for the series to be stationary and linear.

2
2. Literature review

Research in finance has explored how stock markets are affected by their
multisource and heterogeneous data on some scales. Multi-source heterogeneous
datain the stock market means that the data of the stock market includes data from
different sources such as the stock market, the foreign exchange market and even the
weather system, as well as the structure of stock prices, trading volumes, and stock
news, announcements and social networks. and other unstructured data. In particular,
the efficient market hypothesis believes that information from various sources in the
stock market will have an impact on the stock market, while behavioral
financebelieves that financial markets are explained, studied and predicted from the
individual behaviors of traders and the motivations that produce such behaviors. the
trend and extent of price fluctuations. These studies point out that the internal
mechanism of the stock market is very complex, similar to Brownian motion.
Combining the multi-source heterogeneous data in the stock market can more
accurately classify and predict the stock market state.
2.1 Progress of stock price prediction
The research on stock behavior was first conducted by Bachelier in 1900. He used
random walks to express stock price trends. Fama tested that stock price changes are
characterized by random walks. Malkiel and Fama studied valid market assumptions in
1970 and found that all new information will be reflected in asset prices immediately
without delay. Therefore, changes in future asset prices have nothing to do with past
and present information. From their perspective, predicting future asset prices is
considered impossible. On the other hand, many studies try to prove effective market
hypotheses experimentally, and empirical evidence shows that the stock market can be
predictable in some ways. In traditional time series models, parameter statistical
models are used for forecasting, such as ARMA model, ARIMA model and vector
autoregressive model, etc., to find the best estimate. Virtanen and Yliolli used six
explanatory variables to estimate the Finnish stock market index, including the lagging
index and macroeconomic factors in an econometric model based on ARIMA.
3
Work(Clark, T. E., & West, K. D. 2007) proposed a stock price prediction system
based on ARIMA in 2014, which has been tested in the listed stocks originated from
the Stock Exchange in New York and the Stock Exchange running from the country
Nigeria. Then the ARIMA model is regarded as a high potential model for forecasting
short-term series.
The LSTM neural network can control the transmission of input data through
input gates, forget gates and output gates, and maintain the independence of the output
of the memory storage unit and the result output, so that the sequence can retain
important information during transmission and maintain it for a longer period of time
memory. Therefore, the application of LSTM neural network in the prediction of
financial time series has become more and more extensive.

2.2 Time series model

2.2.1 Stationary time series
Stationary time series are divided into strictly stationary time series and wide
stationary time series. Below we introduce their definitions. Strictly stationary time
series provide important theoretical significance, but it is difficult to obtain the joint
distribution of random sequences in the actual research process. Therefore, in order to
better use in practical applications, researchers have defined a relatively weak wide
stationary time sequence. Researchers choose to use the characteristic statistics of the
sequence to define wide stationarity, which can make the constraint conditions a little
looser. By ensuring the stationarity of the low-order moments of the sequence to
ensure that the sequence can be approximately stationary.
Time series analysis also belongs to the field of statistics. It can also analyze the
population through samples like statistics. And from the statistical theorems, we can
know that the number of random variables is directly proportional to the complexity of
the analysis, and the sample size is inversely proportional to the accuracy of obtaining
the overall information (obviously the sample information obtained when the
population is selected as the sample is Overall information, but such an operation is
obviously unrealistic). But time series data has its peculiarities. For a time series {···,
4
X1, X2, · · ·, Xt, · · · }, its value Xt at any time t is a random variable, and since time
is one-way, it cannot be repeated , So we can only get one sample value in this way,
which leads to too little sample information for statistical analysis. But if we have the
concept of stationarity, this problem will be solved.
2.2.2 Principles of the ARMA model
Autoregressive moving average model (ARMA model) is the most commonly
used time series model. According to different conditions, it can be divided into the
following three types of models: autoregressive (AR) model, moving average (MA)
model and ARMA model.
Most of the data in real life is not stable. We need to smooth the data. Box and
Jenkins proved that the difference method is an effective smoothing method.
Therefore, applying the difference method to the ARMA model will result in the
wellknown ARIMA model. The following article will introduce the general
modelingsteps of the ARIMA model.
(1). Sequence stationarity test: First, as with any data analysis, we should draw a
time series graph and check whether the image has an obvious trend; then, we also
need to draw a correlation graph, by observing whether the ACF image is rapidly
reduced to 0 is used to judge the stationarity; finally, if the data does not have
stationarity, then the difference method is needed to smooth the data.
(2). Determine the order of the model: After an appropriate degree of difference
(d), we then need to determine the order p, q of the model, usually using ACF diagram
and PACF diagram to determine the order, in order to make the result more accurate
Using Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC),
the smaller the value of AIC and BIC, the more correct the choice of model order.
(3). Model checking is mainly divided into two parts. The first part is the
significance test of the model. It is usually chosen to test whether the residuals belong
to the white noise sequence. The second part is the parameter test. We choose the
quotient of the estimated coefficient and its standard deviation as the test statistic,
which is usually compared with the critical value of the T statistic significance of 5%

5
(that is, 1.96). If the absolute value of the test statistic is greater than 1.96, the null
hypothesis is rejected, indicating that the coefficient is not significant, otherwise, the
coefficient is significant and the model meets the requirements.

2.3 Deep learning

Although the various forecasting methods mentioned above can achieve a rough
forecast of future stock price changes, they cannot achieve accurate forecasts. This is
because my country's stock market is still in an immature state of development, and
many factors such as national economic conditions, macroeconomic policies, and
investors' psychological expectations in the short term will affect stock prices to some
extent. Therefore, in future forecasts, various factors should also be considered
comprehensively, such as the fundamentals of the operating enterprise, technical
indicators, etc., in order to achieve the investment goal of maximum profit or
avoidance of maximum risk.
2.3.1 Fully connected network
Fully-connected network (fully-connected network, or feedforward network) is a
kind of non- Linear model. It adds nonlinear functions (such as tanh, sigmod, ReLU,
etc.) after the linear transformation and realize the function of non-linear function.
Figure 1 shows a simplest fully connected network structure, including input
layer, hidden layer and output layer: the input x from the input layer to the hidden layer
undergoes a linear transformation and then undergoes a non-linear transformation.

Figure 1 Schematic diagram of a fully connected network

(1)

6
 o W1 W2 b1 b2
Among them, x Rn is input and is output. , , and are the parameter to

W1 b1 W2 b2
be learned. and are the parameter of the hidden layer and and are the

act
parameter of the output layer. is a non-linear activation function, such as tanh,

sigmod and ReLU, etc.

2.3.2 Convolutional Neural Networks
Convolutional neural networks (convolutional neural network, CNN) are widely
used in image processing related tasks (such as image classification, target detection,
object recognition, etc.), has also been applied to natural language speech processing
and speech processing tasks. The fully connected network requires corresponding
parameters for each dimension of data.
The step size indicates that the convolution kernel is slipping. The step length of
each sliding in the dynamic calculation process. Filling means adding specific
elements around the output image to control the size of the output.

Figure 2 Architecture of CNNs

A topological structure of BP training, and good performance has been obtained in

many experiments. A small part of the image called the local receptive area in CNNs is
used as the bottom input of the hierarchical structure. Information is transmitted
through different network levels, so the salient features of observation data that are
invariant to translation, zoom, and rotation can be obtained at each layer.
Deep learning has been widely used into large amount of pattern classification
tasks. Although this field is in the early stages of development, its development will
undoubtedly give an boost to machine learning and AI domain. At the same time, there
are still some specific tasks that are not suitable for processing.

7
2.3.3 Recurrent Neural Network
We treat RNN as a type of recursive neural network which takes sequential data
as input, recursively in the direction of sequence evolution, and all nodes are
connected in a chain, which aims to identify sequential features and use previous
patterns to predict the next possible situation , The structure is shown in Figure 3. In
order to solve the problems of RNN, LSTM is proposed, as shown in Figure 4. A
special storage unit is designed so that it can remember the input historical information
for a longer period of time. LSTM is composed of 3 gates, and the input gate You can
control whether new inputs are allowed, and the forget gate controls which
unimportant information is ignored, and finally the information is output through the
output gate. The network can learn the long-term dependence of the input data well
and remember the historical data information for a longer period of time. LSTM The
forward propagation algorithm of LSTM is similar to RNN. It takes a time series of
length T as input data. Each time the time step advances, the output result is updated.
The backward propagation algorithm of LSTM is also similar to that of RNN.
Beginning at the end, the gradient of each parameter is gradually calculated in the
reverse loop, and finally the network parameters are updated with the gradient of each
time step. Both RNN and LSTM can process time series data to learn time dependence,
and have been widely used in the field of time-space sequence prediction research.

Figure 3 Structure form of standard RNN

8
 Figure 4 Structure of LSTM

3. Methodology

Of course, stock market prediction technology has great economic value for stock
market investors and investment institutions, helping investors and investment
institutions to make profits and avoid investment risks. But the value of stock market
forecasting technology is far more than that. From a social perspective, stock market
forecasting technology can prevent systemic risks in the financial market, help
rationally allocate social funds, and contribute to the harmonious and stable
development of the economy. Stock data has its own characteristics, and the existing
forecasting technology methods are not fully used, so its research brings new
challenges to the technology. In particular, the multi-scale and multi-source
heterogeneous prediction technology can not only be used for stock market prediction,
but also has broad application prospects in many fields such as personal health state
prediction, energy demand prediction and website traffic prediction. This research not
only has important socio-economic value, but also has important academic research
value.
The multi-scale property of stock market data refers to the existence of data at
different time intervals, and the data at different scales will reflect the stock movement
state of different time periods. Large-scale stock market data can reflect the long-term
movement state of the stock market, and small-scale stock market data can reflect the
short-term movement state of the stock market. Data of different scales have associated
information and their own unique information. In order to describe the current market
state more accurately, it is necessary to comprehensively consider stock market data of
multiple scales. However, most of the existing researches only focus on the single-
scale data of the stock market. This can lead to less-than-expected forecast
performance due to an inaccurate description of the state of the stock market. How to
effectively use multi-scale data is the key to accurately describe and predict the market
state.

9
3.1 Data source
This article selects S&P500 index through yahoo finance, and the transaction data
for each trading day from September 26th 2001 to September 24th 2021. The data
includes 5000 observations. The selected data are divided into two parts. First part
occupied 70% of the selected data to train the model , and the remaining observations
are considered to test and validation.

3.2 Methodology
The volatility of stock prices is controlled by the trend of the stock, but is also
sensitive to many other factors. Due to the relative stability and predictability of the
intrinsic value of stocks, the factors that have impacts on the stock market price mainly
include the following aspects: 1. Macro factors; 2. Industrial and regional factors; 3.
Company factors; 4. market factors. This article predicts the closing index of the
S&P500 rather than specific company stock price forecasts, so aside from the more
microscopic industry and company factors, it mainly focuses on the influence of
macroeconomic factors and market factors. Macroeconomic factors refer to the impact
of macroeconomic environment and its changes on stock prices, including regular
factors such as cyclical fluctuations in macroeconomic operations and policy factors
such as monetary policy implemented by the government. This article predicts the
daily data of the S&P500 closing index, mainly focusing on the impact of monetary
policy and other policy factors on stock prices.
There are two types of stock price forecasting methods: qualitative analysis and
quantitative analysis. The qualitative analysis method is the fundamental analysis
method, which is a subjective analysis method relying on the experience of financial
practitioners. This thesis is a numerical prediction of the daily closing index of the
S&P500 rather than a trend judgment of price fluctuations, so this thesis mainly
focuses on the literature review of quantitative analysis methods.
Numerical data-based stock market forecasting research uses numerical data on a
certain time scale in the stock market, such as sky-level index prices and stock price
volume data, to predict specific stocks or other investments in the stock market on the

10
same scale. Predict the future price of the underlying. According to the focus of the
research, these studies can be divided into research on the characteristics of numerical
data stock market forecasting and research on the numerical data stock market
forecasting model.
In order to build our model, in addition to the traditional ARIMA model, this
article will also use the LSTM model. The model in this article uses 70% of the data
for training, and the remaining 30% of the data is used for testing. For training, we use
Root Mean Square Error and Adam algorithm to optimize the model. This Article will
use Stata12 to calculate the ARIMA and GARCH model and use Matlab for the
training.

(1). ARIMA model

As the stock data is noisy, we must first perform stationarity test on the stock
sequence. The test method is to observe the sequence diagram, autocorrelation
diagram, and partial autocorrelation diagram of the sequence first, and then do a unit
root (ADF) test to test its P If the sequence is non-stationary, then we choose the
difference for smoothing. After determining the order of the difference, confirm that it
is a stationary sequence, which can be used to determine the order of the model, that is,
p, q. This article chooses to use the BIC value to determine Order. After the
determination is completed, the model is tested, mainly the LB test, to confirm whether
the residual is white noise. If it is, then the model passes the test and we can make
predictions.

(2). Single Feature LSTM neural network

This model chooses the s&p500 return as the only input feature. First, it is
necessary to test the stationarity of the closing price series: generally choose to draw a
time series diagram first, and check whether the image has an obvious trend; then, we
also need to draw a correlation diagram, and through observing whether the acf image
is quickly reduced to 0 to judge the stationarity; then perform the ADF test; finally, if
the data does not have stationarity, then the difference method is needed to smooth the
data.

11
 After smoothing the data, you can construct a single-feature LSTM neural
network. This article chooses a three-layer LSTM network, that is, there is only one
hidden layer, and the input layer has 20 neurons, so that it can process 20 days of stock
prices , Because we are calculating the closing of the next day, so the output layer has
only 1 neuron, which is used to output the stock price of the twenty-first day. 20 is
chosen because after n-fold cross-checking, 20 is found to be the optimal parameter.

(3) GARCH model

In the 1980s, Engel proposed ARCH (auto regressive conditional heteroskedastic
process) model, which is an autoregressive conditional heteroskedastic process model,
can be used to make such predictions. The ARCH model defined by Engel.
The GARCH model hold an idea that the variance for the change of return can be
predictable, not only the latest information, but also the previous conditional variance
will have an affection on the conditional variance. In order to simplify the calculation,
the risk metrics proposed by the JP Morgan Group's risk management company uses a
simple and practical GARCH(1,1).

(4) Mixed model construction

The mixed model is constructed by ensembling three models including ARIMA
model, Garch Model and LSTM model. The innovation of the article emphasizes the
long-term dependence of LSTM on performance to improve accuracy, and ensemble
can improve the robustness of the model.

(5) Estimator parameters

Since the ultimate goal of stock market forecasting is profit, how to correctly
evaluate the model and select the model with the best profitability is very important in
stock market forecasting. The current stock market forecast research generally adopts a
two-stage model evaluation method: first, the performance of the model is evaluated,
and then the model with the best performance is selected to evaluate the profitability of
the model. The performance evaluation of the stock market forecasting model usually
adopts the classification evaluation indicators, such as the accuracy rate and F1 value,
and the profitability of the stock market forecasting model is estimated by various
simulated trading algorithms. There may be a lack of consistency between the above
12
two evaluation methods, that is, the profitability of the model with the best
classification evaluation performance is not necessarily the best. This inconsistency
can lead to the improvement of stock market forecasting models without valuable
guidance. How to reduce this inconsistency and improve the validity of model
evaluation is a difficult point in stock market research.

4. Results and discussion

4.1 Arima process

The ARIMA time series model is a differential processing of the autoregressive
moving average model. Its main methods are modeling, evaluation, verification and
control, which are expressed as ARIMA(p, d, q). The main idea of the model is to
regard the known data as a random sequence when it is formed in the order of time
development, and then describe the random sequence by mathematical modeling. Time
series values predict future values.
This project chooses the closing price sequence as the time sequence, and the
sequence diagram of the closing price sequence is shown in Figure 5:

Figure 5 Timing chart of return series

As observed in the above figure, it can be seen that the stock price fluctuated to a
certain extent during the period, and it was not stable.

13
 Observing the sequence diagram of the sequence, we can consider the data to be
non-stationary. In order to confirm our conjecture, we then draw the autocorrelation
graph and the partial autocorrelation graph, as shown in Figure 6:

Figure 6 The results for Arima Prediction

As can be seen from figure above, the ARIMA model is not very effective in
predicting stock data, the error fluctuates between +7.2 and -2.6, and the upward or
downward trend is basically the same as the change trend of the original data.
4.2 GARCH process

Figure 7 The results for Garch Prediction

14
 As can be seen from Figure 7, the Garch model is better than the single ARIMA
model in predicting stock data. For the Garch model, the error fluctuates between
+2.1~-0.6, but the error is basically concentrated between +1~-1, the rising or falling
trend is almost the same as the change trend of the original data , and even have
overlapping intervals. It can be seen that the Garch model is stronger than ARIMA
model's prediction of stocks in both the accuracy of stock prediction and the trend of
stock changes.
4.3 LSTM process
In traditional neural networks, neurons in the same hidden layer are not connected
to each other, and this structural defect directly leads to their poor performance in
dealing with certain problems. This shortcoming becomes especially acute when
dealing with time series and speech recognition problems where information is
contextualized. The emergence of the recurrent neural network solves this problem
very well. The neurons in the same hidden layer are connected to each other, which
can effectively obtain the contextual information of the data. The output of the
recurrent neural network is determined according to the input and the previous related
information, so it can play its short-term memory when dealing with time series
problems.
Although the effect of recurrent neural network in dealing with time series
problems is very good, there are still some problems. The more serious one is that
gradient disappears or explodes easily in the processing of long-term span problems.
Causes the phenomenon of small memory value. After the cyclic neural network is
expanded, it can be regarded as a multi-layer feedforward neural network with each
layer sharing the same weight parameters. Although it keeps trying to learn the
longterm dependencies of sequences, actual research finds this to be a difficult task
indeed. Long-term reliance on signals tends to become very weak and highly
susceptible to short-term signal fluctuations. There is a multiplier of the derivative of
the activation function in time-based backpropagation, and the continuous
accumulation will cause uncontrollable problems. Although it can be solved

15
theoretically by adjusting the parameters, it is found that this problem is difficult to
solve in practice, so it still needs to be optimized from the structure. This leads to its
improved structure - the LSTM neural network.
Next, we can start the construction of the LSTM neural network. The first is the
determination of several parameters. After n-fold cross-validation, we choose the
hidden layer to have 10 neurons. The number of iterations is selected 50 times, and
each 72 sample data is formed into a batch for training, that is, batchsize = 72, Adam
algorithm is used as the optimizer of the model, the learning rate is 0.001, and the
training set data is randomly scrambled. Use the MSE indicator as the loss function of
the model for training. Figure 8 is the training diagram of the neural network. It can be
seen from this diagram that after iteration, the loss function of the model decreases
quickly and tends to converge. It can be seen that the prediction model is more
reasonable.

Figure 8 The results for LSTM Prediction

The code of LSTM can be seen in Apppendix.

4.4 Mixed model process

16
 Figure 9 The results for Ensemble Prediction
This chapter chooses the traditional time series model and the LSTM neural
network model to construct the stock price model and make predictions. First, the
ARIMA model is used, and only the closing price sequence is stabilized, the model is
determined, and the model is checked. Finally, the stock price forecast was made; then
a new forecast of the stock price was made using the GARCH model; at this point, the
application of the traditional time series model ended. Next, using the same data, the
LSTM neural network with single-feature input and multi-feature input was
constructed, the number of layers was selected and the neurons were determined, and
when the model parameters were trained to meet the standards, predictions of closing
prices were given. Finally, the mean square error (MSE) of each model was calculated
separately to compare the models.
We are going to leverage the model of LSTM into traditional financial time series
forecasting, and a stock forecasting model based on long short-term memory neural
network (LSTM) is established. The absolute error and the coefficient of determination
were evaluated, and a better prediction effect was obtained. It proves the feasibility of
deep learning in financial time series forecasting, which can guide the investment
behavior of institutions and individuals to a certain extent, and provides new ideas for
stock forecasting research.
Firstly, an ARIMA model was established based on the closing index sequence.
The input feature was the closing index of the previous day. The MSE of the initial test
was 2.185, and the average error rate was 4.48%. After hyperparameter tuning, the
prediction effect was significantly improved, and the MSE dropped to 1.213, and the
average error rate is reduced to 3.19%. During the test period, the average yield
fluctuation of the closing index was 0.62%, which was far lower than the optimal
forecasting effect of the model. It can be considered that the forecasting effect of the
ARIMA model on the S&P500 is extremely poor and has no practical significance. At
the same time, to a certain extent, it shows that there are many factors affecting the

17
fluctuation of stock prices, and the historical price of the S&P500 closing index cannot
fully reflect the relevant information of the stock market.
Secondly, the optimal prediction effect of the GARCH model after
hyperparameter tuning is MSE of 1.923 and an average error rate of about 1.65% in
the test period, which are both better than the prediction evaluation indicators of the
ARIMA model, indicating that the comprehensive index of historical transaction
information (SH index) ), Investor Sentiment Composite Index (IS Index) and
Monetary Policy Composite Index (MP Index), compared with a single sequence of
closing indices, they express more information on the S&P500, and the three-factor
forecast model has good practical significance. However, the average error rate at this
time is still higher than the average yield volatility of the closing index of 0.62%, and
the applicability of the GARCH method to the forecasting of the Shanghai Composite
Index still needs further investigation.
Finally, examining the predictive ability of the LSTM model established based on
13 original indicators after tuning the hyperparameters, it can be found that the MSE
between the predicted value of the model and the actual value during the test period
was 0.876, the average error rate was further reduced to 0.40%, and the predictive
ability was significantly better. The MSE of the Mixed Model is 0.412 and the average
error rate is 0.27%.
Table1 Accuracy Comparison For regression models
Model MSE Average Error Rate

ARIMA 1.213 3.19%

GARCH 1.923 1.65%

LSTM 0.876 0.40%

Mixed Model 0.412 0.27%

Base on Table1, the MSE and Average Error Rate of each model was calculated
separately for comparison between the models. It can be seen that the MSE value and
Average Error Rate of the Mixed Model is the lowest, so it can be clearly seen that the
Mixed Model is better than the traditional time series model.

18
5. Conclusion and Implication

5.1 Conclusion
The importance of the stock market to a country's economy will make the types of
stock price forecasting methods continue to develop and grow, and will continue to be
derived from the development of other disciplines. In the development process of the
follow-up forecasting method, it is necessary to continuously explore and deeply study
the characteristics of the stock market, so as to make the model closer to reality,
expand the applicability of the method, and obtain better forecasting accuracy.
Because stock data is affected by economic factors, political factors or
environmental factors, the law of its change is elusive, and the cycle of the law of
change is difficult to determine. Therefore, the model still needs a lot of historical data
and selection of appropriate variables for analysis to obtain the desired results. In the
traditional ARIMA model, when analyzing complex stock markets, its prediction
results are not particularly ideal, and there are still certain errors in price prediction. As
a technology in the field of deep learning, neural network can solve non-linear
problems well. LSTM neural network is optimized on traditional neural network and
introduces the concept of "gate", which enhances the long-term memory ability of the
model , Which enhances its generalization ability. Therefore, the application of LSTM
neural network in analyzing financial-related time series data is promising.
Based on the understanding of traditional time series analysis and RNN and
LSTM neural network, this paper constructs a stock price prediction model based on
LSTM neural network. For better comparison, we also established a traditional
ARIMA model for comparison. As the neural network has a good predictive effect on
nonlinear problems, this article chooses the optimized neural network-LSTM model,
and also chooses the use of single-feature and multi-feature input models to seek better
prediction results. The traditional time series model focuses on the role of time in stock
forecasting. However, certain errors will occur when the model deals with complex
19
nonlinear stock data, and the model does not consider other factors, such as economics
and politics, so the prediction error of the ARIMA model will be large. Next, this thesis
considers the ARIMA model, the GARCH model, and the singlefeature input LSTM
model that can handle nonlinear data, but they all have unconsidered problems, and
their prediction results will also appear to be certain. The error. The multi-feature input
LSTM model not only takes into account the influence of external factors, but also can
process non-linear data, and its prediction performance is better. Through the result of
the prediction, we can see that the prediction result of the mixed model is the best.
For the work of this article, the following points can be summarized:
(1) Carry out the steps of smoothing, model ordering, and model checking on our
stock data, and finally establish the ARIMA model and predict the stock price; (2)
Carry out the steps of smoothing, model ordering, and model checking on our stock
data, and finally establish the GARCH model and predict the stock price; (3) Construct
an LSTM neural network, determine the number of neural network layers and neurons,
and train the parameters; (4) Construct a mixed model to predict stock prices. (5)
Compare the prediction results of the above models.

5.2 Implication
The prediction model studied in this article is based on the LSTM neural network,
and preliminary results have been obtained. However, due to objective factors such as
research time and data sources, there is still a lot of room for research in this article.
There is still much to be done for the model constructed in this article. Update and
improvement, the follow-up work mainly includes the following aspects:
(1)Handling of abnormal values

Because stock market has a certain degree of speculation and is also susceptible to
policy influences, there are often skyrocketing and plummeting situations. This leads
to outliers in the stock data we obtain. There are many reasons for the occurrence of
outlier points in stock data, which cannot be obtained by quantitative analysis. This
makes the problem unable to simply use the LSTM neural network constructed in this

20
article. Therefore, some methods to deal with outliers can be used to perform data
processing. Noise reduction, such as wavelet transform, Fourier transform, etc.
(2)About feature selection

The number of features of the data set obtained in this article is not very large,
and for real stock data, in addition to the features in the stock market, can you use
other features, such as corporate financial reports; also, because the stock market is
affected by policy Larger, using the application of LSTM neural network in text
learning and text sentiment analysis, can we obtain some features from news and
financial reports, so as to enable the model to make corresponding judgments on stocks
in an economic sense, thereby improving the accuracy of our predictions rate.
(3)About model optimization
When constructing the neural network model, whether the number of hidden
layers is small, and whether more hidden layers will have better prediction results, this
is also the lack of research in this article.

21
Reference
Addison, P. S. (2002). The illustrated wavelet transform handbook. Napier University.
Avramov, D. (2002). Stock returns predictability and model uncertainty. Journal of
Financial Economics, 64, 423–458.

Brock, W., Lakonishok, J., & LeBaron, B. (1992). Simple technical trading rules and
the stochastic properties of stock returns. The Journal of Finance, 47, 1731–1764.

Campbell, J. Y., & Thompson, S. B. (2008). Predicting excess stock returns out of
sample: Can anything beat the historical average? Review of Financial Studies,
21,1509–1531.

Campbell, J. Y., & Vuolteenaho, T. (2004). Bad beta, good beta. The American
Economic Review, 94, 1249–1275.

Chen, J., Jiang, F., & Tong, G. (2017). Economic policy uncertainty in China and stock
market expected returns. Accounting and Finance, 57, 1265–1286.

Clark, T. E., & West, K. D. (2007). Approximately normal tests for equal predictive
accuracy in nested models. Journal of Econometrics, 138, 291–311.
Cochrane, J. H. (2007). The dog that did not bark: A defense of returns predictability.
Review of Financial Studies, 21, 1533–1575.

Conrad, J., & Kaul, G. (1998). An anatomy of trading strategies. Review of Financial
Studies, 11, 489–515.

Cowles, A., 3rd (1933). Can stock market forecasters forecast? Econometrica. Journal
of the Econometric Society, 309–324.

Dai, Z., Zhou, H., Wen, F., & He, S. (2020a). Efficient predictability of stock return
volatility: The role of stock market implied volatility. The North American Journal of
Economics and Finance, 52, 101174.

Dai, Z., & Zhu, H. (2020). Stock returns predictability from a mixed model
perspective. Pacific-Basin Finance Journal, 60, 101267.

Dai, Z. F., Dong, X. D., Kang, J., & Hong, L. (2020b). Forecasting stock market
returns: New Technical indicators and two-step economic constraint method. The
North American Journal of Economics and Finance, 53, 101216.

22
Dangl, T., & Halling, M. (2012). Predictive regressions with time-varying coefficients.
Journal of Financial Economics, 106, 157–181.

Daubechies, I. (1992). Ten lectures on wavelets. Philadelphia, PA: SIAM (Society for
Industrial and Applied Mathematics).

DeMiguel, V., Garlappi, L., Nogales, F. J., & Uppal, R. (2009). A generalized approach
to portfolio optimization: Improving performance by constraining portfolio norms.
Management Science, 55, 798–812.

Fama, E. F., & Blume, M. F. (1966). Filter rules and stock market trading. Journal of
Business, 39, 226–241.

Fama, E. F., & French, K. R. (1988). Dividend yields and expected stock returns.
Journal of Financial Economics, 22, 3–25.
Faria, G., & Verona, F. (2018). Forecasting stock market returns by summing the
frequency-decomposed parts. Journal of Empirical Finance, 45, 228–242.

Ferreira, M. I., & Santa-Clara, P. (2011). Forecasting stock market returns: The sum of
the parts is more than the whole. Journal of Financial Economics, 100, 514–537.

Gençay, R., Selçuk, F., & Whitcher, B. (2002). An introduction to wavelets and other
filtering methods in finance and economics. Academic Press.
Goyal, A., & Welch, I. (2003). Predicting the equity premium with dividend ratios.
Management Science, 49, 639–654.

Goyal, A., & Welch, I. (2008). A comprehensive look at the empirical performance of
equity premium prediction. Review of Financial Studies, 21, 1455–1508.

Guo, H. (2006). Time-varying risk premia and the cross section of stock returns.
Journal of Banking & Finance, 30, 2087–2107.

Haven, E., Liu, X., & Shen, L. (2012). De-noising option prices with the wavelet
method. European Journal of Operational Research, 222(1), 104–112.

Inoue, A., & Kilian, L. (2004). In-sample or out-of-sample tests of predictability:

Which one should we use? Economic Review, 23, 371–402.

Jaffard, S., Meyer, Y., & Ryan, R. D. (2001). Wavelets from a historical perspective.
Philadelphia, PA: SIAM (Society for Industrial and Applied Mathematics).

Jiang, F., Lee, J. A., Martin, X., & Zhou, G. (2019). Manager sentiment and stock
returns. Journal of Financial Economics, 132, 126–149.

23
Appendix
clear all; close all;clc
rng(20211004)
mode = 1; %% mode=2 LSTM training， mode = 1 Loading

if mode == 2 %%
Data Process
filename = "sp500.xls";
sheet = "Sheet1";
[num,txt,raw] = xlsread(filename,sheet);

data = num; date =

datenum(raw(3:end,1));

figure plot(date,data,'linewidth',1,'color',[0 0 1]);

xticks([datenum('2001/1/1'),datenum('2002/1/1'),datenum('2003/1/1'),datenum('2004/1/1'),daten
u
m('2005/1/1'), datenum('2006/1/1'),
datenum('2007/1/1') ,datenum('2008/1/1') ,datenum('2009/1/1') ,datenum('2010/1/1')...
,datenum('2011/1/1'),datenum('2012/1/1'),datenum('2013/1/1'),datenum('2014/1/1'),da
tenu m('2015/1/1'),datenum('2016/1/1')...
,datenum('2017/1/1'),datenum('2018/1/1'),datenum('2019/1/1'),datenum('2020/1/1'),da
tenu m('2021/1/1'),datenum('2022/1/1')])
xticklabels(['2001','2002','2003','2004','2005','2006','2007','2008','2009','2010'... ,'2011','2012','
2013','2014','2015','2016'...
,'2017','2018','2019','2020','2021','2022'])
datetick('x','yyyy','keepticks');
xlabel('Year','FontSize',12); ylabel('Stock
Return','FontSize',12)
set(gcf,'unit','centimeters','position',[10 5 30 15]);
axis normal;

numTimeStepsTrain = floor(0.7*numel(data));
dataTrain = data(1:numTimeStepsTrain+1);
dataTest = data(numTimeStepsTrain+1:end);

mu = mean(data); sig = std(data);

dataTrainStandardized = (dataTrain - mu) /
sig; dataTestStandardized = (dataTest - mu) /
sig; stay_to_predict = 10;

Train_shift = []; for i = 1 : 1 :

stay_to_predict + 1 data_temp =
dataTrainStandardized; data_temp =
circshift(data_temp,-(i-1));

24
 Train_shift = [Train_shift data_temp];
end
Train_shift(end - stay_to_predict:end,:) = [];
XTrain = Train_shift(:,1:stay_to_predict);
YTrain = Train_shift(:,stay_to_predict+1);

Test_shift = []; for i = 1 : 1 : stay_to_predict

+ 1 data_temp = dataTestStandardized;
data_temp = circshift(data_temp,-(i-1));
Test_shift = [Test_shift data_temp];
end
Test_shift(end - stay_to_predict:end,:) = [];
XTest = Test_shift(:,1:stay_to_predict);
YTest = Test_shift(:,stay_to_predict+1);

%% LSTM

numFeatures = stay_to_predict;
numResponses = 1; numHiddenUnits = 300;

layers = [ ...
sequenceInputLayer(numFeatures)
lstmLayer(numHiddenUnits)
fullyConnectedLayer(numResponses)
regressionLayer];
options = trainingOptions('adam', ...
'MaxEpochs',500, ...
'GradientThreshold',1, ...
'InitialLearnRate',0.01, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropPeriod',125, ...
'LearnRateDropFactor',0.2, ...
'Verbose',0, ...
'Plots','training-progress'); net = trainNetwork([XTrain;XTest]',
[YTrain;YTest]',layers,options);
[net,YPred] = predictAndUpdateState(net,[XTrain;XTest]');
YPred = sig*YPred(end - 1488:end) + mu;
YTest = sig*YTest + mu; rmse =
sqrt(mean((YPred-YTest').^2)) figure
plot(YTest','r')
hold on
plot(YPred,'.-b') hold off
legend(["Observed"
"Predicted"])

25
save('YPred.mat','YPred')
save('YTest.mat','YTest')
end %% Results and
Plots if mode == 1
filename = "sp500.xls";
sheet = "Sheet1";
[num,txt,raw] = xlsread(filename,sheet);

data = num; date =

datenum(raw(3:end,1));
load('YPred.mat')
load('YTest.mat')
rmse = sqrt(mean((YPred-YTest').^2))
%%
figure
subplot(2,1,1) plot(date(end -
1488:end),YPred,'linewidth',1,'color',[0 0 1]);
xticks([datenum('2015/1/1'),datenum('2016/1/1')...
,datenum('2017/1/1'),datenum('2018/1/1'),datenum('2019/1/1'),datenum('2020/1/1'),da
tenu m('2021/1/1'),datenum('2022/1/1')]) xticklabels(['2015','2016'...
,'2017','2018','2019','2020','2021','2022'])
datetick('x','yyyy','keepticks');
xlabel('Year','FontSize',12); ylabel('Stock
Return','FontSize',12) title('LSTM Prediction')
set(gcf,'unit','centimeters','position',[10 5 30 15]);
axis normal;

subplot(2,1,2) plot(date(end - 1488:end),YPred -

YTest','linewidth',1,'color',[1 0 0]);
xticks([datenum('2015/1/1'),datenum('2016/1/1')...
,datenum('2017/1/1'),datenum('2018/1/1'),datenum('2019/1/1'),datenum('2020/1/1'),da
tenu m('2021/1/1'),datenum('2022/1/1')]) xticklabels(['2015','2016'...
,'2017','2018','2019','2020','2021','2022'])
datetick('x','yyyy','keepticks');
xlabel('Year','FontSize',12);
ylabel('Error','FontSize',12) title('LSTM
Prediction Error')
set(gcf,'unit','centimeters','position',[10 5 30 15]);
axis normal; end

26