Implementations of Hybrid Prediction Models for Stock Price

Forecasting

Jiaji Zeng

Mathematical Institute, Southwestern University of Finance and Economics, Chengdu, China

Keywords: Prediction of Stock Price, Hybrid Prediction Model, Financial Markets.

Abstract: With the rapid advance of the world economy since the 20th century, many people are devoting their energy

to studying the stock market for profit. Therefore, various research methods for stock price prediction have

emerged, among which a single prediction model is an important part of these research methods. However,

with the progress of technology, the limitations of single model prediction are gradually amplified. It is

difficult to fully capture the dynamic changes and uncertainties of complex financial systems. So, aiming to

augment the accuracy and reliability of predictions, scientists began exploring the possibility of combining

multiple prediction models, namely hybrid prediction models. This research will start with stock price

prediction, based on variable analysis, representative configuration, model application results and

performance, as well as its limitations and prospects, to explore the implementation of hybrid prediction

models for the prediction of stock price. These results are of great significance in exploring the development

prospects, risk assessment, optimization and adjustment of financial markets.

1 INTRODUCTION

With the rapid progress of the times and economy,

plenty of people are participating in the stock market

with the expectation of getting rich. Therefore, the

prediction of stock price has become a crucial concept

in modern financial markets. There have been a lot of

methods for predicting stock price in the past, which

can be generally separated into several types:

technical analysis, basic analysis, macroeconomic

analysis, market sentiment analysis, and quantitative

analysis. In recent years, methods for predicting stock

prices have been constantly developing and

innovating. Stock price prediction is a rather

hazardous operation. A good analyst is therefore not

someone who is always right, someone who has a

higher efficiency than his colleagues (Op't Landt,

1997). In recent years, with the obvious progress of

new technologies, researchers have adopted various

cutting-edge hybrid models in the field of predicting

stock price to address the complexity and nonlinear

characteristics of the stock market. These hybrid

models combine several advantages of dissimilar

algorithms and techniques to promote the accuracy

and stability of these predictions (Ji et al., 2021).

The first new method is the combination of deep

learning and traditional statistical models, including

two types of models. Long Short Term Memory

Network (LSTM), as a variant of Recurrent Neural

Network (RNN) in deep learning, excels at handling

long-term dependency problems in sequential data.

The autoregressive integrated moving average model

(ARIMA) is a classic model in traditional time series

analysis. Some studies combine LSTM with ARIMA

to capture nonlinear trends and fluctuations, while

using ARIMA to process linear parts, thereby

improving prediction accuracy. Stock price

prediction is a challenging problem due to its random

movement. This hybrid model is a combination of

two well-known networks, LSTM and Gated

Recurrent Unit (GRU) (Hossain et al., 2018).

Transformer model has achieved significant results in

the area of natural language development with

powerful parallel computing capability and efficient

information extraction ability. In recent years,

researchers have begun to combine Transformer with

LSTM for stock prediction. This hybrid model can

simultaneously focus on global and local information,

further improving prediction performance (Jakkula,

2006).

The second method is the fusion of machine

learning algorithms. Support Vector Machine (SVM)

568

Zeng, J.

Implementations of Hybrid Prediction Models for Stock Price Forecasting.

DOI: 10.5220/0013270700004568

In Proceedings of the 1st International Conference on E-commerce and Artiﬁcial Intelligence (ECAI 2024), pages 568-573

ISBN: 978-989-758-726-9

is based on the principle of minimizing structural risk,

striving to minimize empirical risk while maximizing

the generalizability of the learning model

(Vishwanathan et al., 2002). However, a single SVM

model may not be capable of handling complex stock

markets. Therefore, researchers combined SVM with

self-organizing map neural networks (SOM), particle

swarm optimization algorithms (PSO), and other

methods to construct a hybrid model to cope with the

variability of the stock market (Biswas et al., 2021).

Besides, some other machine learning algorithms like

random forest, gradient boosting tree, etc. are also

used for stock prediction. Research work uses the

frequently used algorithms LSTM, Extreme Gradient

Boosting (XGBoost), Linear Regression, Moving

Average, and Last Value model, on more than twelve

months of past stock data to erect a prediction model

to forecast stock price. Ensemble learning method can

effectively improve the general prediction

performance by combining prediction outcomes of

many single models.

Thirdly, the combination of Big Data and

Artificial Intelligence (AI). Along with the

advancement of Big Data technology, researchers can

obtain more dimensions of stock market data. These

data not only include traditional indicators like past

stock prices and trade volumes, but also non-

traditional data like social media sentiment and

macroeconomic indicators. A hybrid model based on

big data could completely utilize this key information

to promote the accuracy and timeliness of predictions.

Some cutting-edge research has begun to apply

hybrid models to automated trading systems. By

monitoring market dynamics in real-time and

predicting stock price trends, automated trading

systems can automatically execute buy or sell

operations at the appropriate time, reducing the risk

of human intervention and improving trading

efficiency. In recent years, the number of academic

papers and patents on hybrid model prediction of

stock prices has been increasing. These research

results not only promote the development and

improvement of relevant theories, but also provide

strong support for practical applications.

Besides, some financial institutions and

technology companies have begun to apply hybrid

models to stock prediction and trading strategies. By

continuously optimizing models and algorithms,

these institutions are able to provide customers with

more accurate and personalized investment advice

and services. In summary, based on previous research,

this article will explore the implementation of a

hybrid prediction model for predicting of stock price

through specific examples or methods. The

framework would base on these topics: Descriptions

of stock price prediction, Configurations for hybrid

model, Implementation results and Limitations and

Prospects.

2 DESCRIPTIONS OF STOCK

PRICE PREDICTION

In stock price prediction, the dependent variable

usually refers to the target variable that one attempts

to predict or explain, namely the stock price itself or

its related indicators (Pinto & Asnani, 2011).

Specifically, the dependent variables for stock price

prediction can include these aspects. For prices:

 Opening price: First trading price of 1 stock at

the beginning of a market day.

 Closing price: Last trading price of 1 stock at the

end of a trading day.

 Highest and lowest prices: The highest and

lowest trading prices of a stock during a trading

day.

These prices directly reflect the market

performance of stocks and are one of the most

important indicators for investors to pay attention to.

For stock return rate

 Daily return: Percentage change in the price of

1 stock within a single market day.

 Weekly return, monthly return, and annual

return: represent percentage change of stock

prices over a week, month, and year,

respectively.

Stock return plays an important role in measuring

stock investment returns and a common dependent

variable in stock price prediction. Regarding to other

relevant indicators

 Trading volume: The number of trades in a

stock during a specific period of time, reflecting

the level of market activity and investor

participation (Cakra & Trisedya, 2015).

Some financial indicators, such as price to

earnings ratio and price to book ratio: These factors

are used to evaluate the investment value of stocks.

Although they are not direct price indicators, they are

often used as references in stock price forecasting

(Selvin et al., 2017).

In stock price prediction, the dependent variable

is usually the future price or price trend of the stock.

To predict this dependent variable, a series of

independent variables (also known as features, factors,

or input features) need to be used, as well as selecting

an appropriate prediction model (Ji et al., 2021). The

independent variables for stock price prediction can

Implementations of Hybrid Prediction Models for Stock Price Forecasting

569

include various types of data, which typically reflect

multiple aspects such as the historical performance of

the stock market, company fundamentals,

macroeconomic environment, and market sentiment

(Song & Lee, 2020). There are several common

independent variables:

 Historical price data: including past stock prices,

opening and closing prices, highest and lowest

prices, etc. These data can reflect the historical

fluctuations of stock prices.

 Trading volume data: The trading volume of

stocks is also an important predictor, which can

reflect the level of market activity and investor

participation.

 Technical indicators: like moving averages,

RSI, stochastic oscillators can help analyze the

overbought and oversold situation of stock

prices, trend strength, etc.

 Fundamental data: including the company's

financial statement data (such as revenue, profit,

balance sheet, etc.), price to earnings ratio or

book ratio, dividend yield, etc. Those data can

reflect the company's profitability, the financial

condition, and the market valuation.

 Macroeconomic data: Rate of GDP growth, rate

of inflation, interest rate level, can reflect the

impulse of overall economic environment on

stock market.

 Market sentiment data: including news

sentiment analysis, social media sentiment

index, etc., can reflect investors' overall views

and emotional changes towards the stock

market.

In stock price prediction, there are multiple

models to choose from, each with its own

characteristics and scope of application. Here are

some common prediction models:

 Linear regression model: Predicting stock prices

by fitting a linear relation between the

independent variables and dependent variables.

Although simple, it may be effective in certain

situations.

 The time series analysis models, like ARIMA

models, seasonal decomposition models are

particularly suitable for processing data with

temporal dependencies.

 Machine Learning (ML) models: including

decision trees, random forests, SVM, neural

networks, etc. These models are capable of

handling complex nonlinear relationships and

perform well on large amounts of data.

 Deep learning models: RNN, LSTM,

Transformers, etc. These models perform well

in processing time series data and capturing

long-term dependencies, making them

particularly suitable for stock price prediction.

3 CONFIGURATIONS FOR

HYBRID MODEL

Regarding the LSTM-SDE configuration of hybrid

models, there is no widely recognized hybrid model

directly named LSTM-SDE in standard deep learning

architectures. However, the LSTM-SDE is an attempt

to combine LSTM with a method based on Stochastic

Differential Equations (SDE) to integrate the

advantages of both in processing time series data and

dynamic system modeling (Melo et al., 2022). Base

on both advantages of LSTM model and advantages

of SDE, here are some potential characteristics of

LSTM-SDE configuration. Firstly, LSTM is a special

type of recurrent neural network (RNN) that excels at

processing sequential data and capturing long-term

dependencies. Through structures such as forget gates,

input gates, and output gates, LSTM can effectively

control the flow of information and reduce the

problem of gradient vanishing or exploding. Secondly,

SDE (stochastic differential equation) is a

mathematical model that describes the dynamic

changes of stochastic processes and is widely used in

fields such as finance, physics, and biology. SDE can

capture the uncertainty and randomness of systems,

and has unique advantages for modeling complex

dynamic systems (Araújo et al., 2015). As a result,

there is a combination method. In the LSTM-SDE

configuration, LSTM may be used for feature

extraction and temporal modeling of time series data,

while SDE is used to describe the random changes in

system state. The two can be combined in some way,

such as using the output of LSTM as the input of SDE,

or associating certain parameters of SDE with the

hidden state of LSTM. Because of the high

nonlinearity and randomness Financial market data

always has, LSTM can capture the long-term trends

of the market, while SDE can describe the random

fluctuations of the market. By combining LSTM and

SDE, more accurate financial market prediction

models can be constructed as well.

Before studying the CNN-LSTM-AM model, it is

necessary to first introduce the CNN-LSTM model.

The CNN-LSTM model has combined both the

advantages of CNN and LSTM. CNN excels at

processing data with grid structures, such as images

and videos, and automatically extracts features from

input data through convolutional layers. LSTM is

good at processing sequential data and can capture

ECAI 2024 - International Conference on E-commerce and Artiﬁcial Intelligence

570

long-term dependencies. In the CNN-LSTM model,

CNN is first used to extract spatial features of input

data, and then these feature sequences are used as

inputs to LSTM to learn the temporal dependencies

between features (Sun et al., 2024). In that case, the

CNN-LSTM-AM model introduces an attention

mechanism based on CNN-LSTM, allowing the

model to focus more on important features or time

points. The attention mechanism allows the model to

dynamically adjust the importance weights of

different time steps or features when processing

sequential data. Therefore, the application of CNN-

LSTM-AM (Convolutional Neural Network Long

Short Term Memory Network Attention Mechanism)

in stock price prediction is a comprehensive method

that combines multiple deep learning techniques,

aiming to improve the accuracy and efficiency of

prediction.

4 IMPLEMENTATION RESULTS

In prediction of stock price, hybrid prediction models

have been widely applied and researched due to their

ability to combine multiple methods and techniques

to promote prediction accuracy and stability. There

are some specific results and performance of

implementing hybrid prediction model for predicting

stock prices. This model is constructed by combining

self attention mechanism, LSTM, and GRU.

Compared with models such as LSTM, GRU, RNN-

LSTM, and RNN-GRU, the ATLG model has higher

accuracy. Introducing self attention mechanism

makes the model to better focus on stock feature

information at important time points. Through

backtesting using the MACD (Moving Average

Convergence and Divergence) indicator, a return of

53% was obtained, which is higher than the return of

the Shanghai and Shenzhen 300 during the same

period, proving the effectiveness and practicality of

the model (Xie et al., 2021). Some researchers use

SVR. This model has combined SVR (Support Vector

Regression) with various methods such as wavelet

transform, sliding window, etc. The original data is

transformed using the difference method to highlight

the trend of stock price fluctuations. Wavelet

transform is introduced to separate the main and

detailed components of stock prices, and multi-scale

sliding windows are used to smooth stock prices,

especially those with strong noise (Liu et al., 2022).

Compared with the SVR model based on a single

feature, this hybrid model has improved prediction

accuracy. However, due to the lack of detailed

experimental data and results in the reference article,

specific accuracy improvement values cannot be

provided.

This model is constructed by combining

Variational Mode Decomposition (VMD) with two

different neural networks (such as GRU and Echo

State Network ESN) for mixed prediction. VMD

decomposition effect: VMD decomposition

effectively reduces the complexity of raw data and

helps extract different frequency features of stock

prices. By separately predicting the sub sequences

decomposed by VMD and combining them with

ARIMA model for error correction, the overall

prediction accuracy and stability of the model have

been significantly improved. The experimental

results have showed that this model has a advance

accuracy than the traditional single model and other

hybrid models when predicting the historical data of

stocks such as Shanghai Shenzhen 300 and the US

Standard&Poor's 500.

To sum up, these hybrid forecasting models show

high accuracy and stability in stock price forecasting,

and effectively improve the forecasting ability of a

single model by combining different technologies and

methods. However, the specific prediction

performance will be affected by various factors such

as model design, data quality, and market

environment, so adjustments and optimizations need

to be made according to specific situations in

practical applications.

5 LIMITATIONS AND

PROSPECTS

When using mixed forecasting models for stock price

prediction, although these models combine the

superiority of various techniques and algorithms to

promote the accuracy and stability of predictions,

there are still some limitations. Firstly, hybrid

prediction models typically combine multiple

algorithms and techniques, making the model

structure more complex. This complexity may

increase the training difficulty and computational cost

of the model, while also potentially reducing its

interpretability. Secondly, due to the combination of

multiple algorithms in the hybrid model, its

prediction results may be difficult to explain directly.

This may be a problem for investors as they need to

understand the logic and basis behind their

predictions. Hybrid prediction models are highly

depending on quality and completeness of input data.

Whether there are deviations, missing or outliers in

the data, it will significantly affect the training

Implementations of Hybrid Prediction Models for Stock Price Forecasting

571

effectiveness and prediction accuracy of the model.

Besides, stock market is influenced by various factors,

including macroeconomics, industry policies, market

sentiment, etc. However, data on these factors may be

incomplete or difficult to obtain, resulting in the

model being unable to fully consider all relevant

factors when making predictions. Hybrid prediction

models are often built based on a series of

assumptions. These assumptions may not fully

conform to actual situation of the stock market,

thereby affecting the accuracy of the model's

predictions. Also, the parameter settings in the hybrid

model have a obvious impulse on the prediction

results. Whereas, determining the optimal parameter

combination is a challenging problem. Inappropriate

parameter settings may lead to overfitting or

underfitting of the model, thereby reducing predictive

performance. On one hand, current technologies and

algorithms still have certain limitations when dealing

with complex data and relationships. These

limitations may limit the effectiveness of hybrid

prediction models in predicting stock prices. On the

other hand, different hybrid prediction models may

use different combinations of algorithms. However,

not all algorithms are suitable for stock price

prediction. Improper selection of algorithms may lead

to a decrease in model performance. Under the

influence of numerous factors, the stock market has a

great degree of uncertainty and is influenced by

various unpredictable factors. These factors may

include policy changes, natural disasters,

international situations, etc. Although hybrid

prediction models can handle complex data and

relationships, their predictive ability may be limited

when faced with these sudden unpredictable events.

What’s more, the volatility of the stock market is also

a crucial factor affecting the accuracy of predictions.

In that case of significant market volatility, the

predicted results of the hybrid model may have

significant deviations from the actual trend.

Although there are still many limitations to

prediction, one can still optimize it through various

methods. For instance, one can improve data quality

and integrity to make sure enough accuracy and

reliability of input data, and simplify the model

structure and improve its interpretability. One needs

to be cautious in setting model parameters to avoid

overfitting or underfitting. Besides, both paying

attention to market dynamics and changes and

adjusting models in a timely manner are significant

for adapting to market changes. At last, one should

continuously explore and try new technologies and

algorithms to improve the performance and

application effectiveness of hybrid prediction models.

All in all, the development prospects of using

mixed forecasting models for stock price prediction

are quite broad, mainly due to several factors. With

the continuous advancement of deep learning

technology, models such as LSTM and CNN have

demonstrated forceful capabilities in processing time

series and spatial feature data. These technologies

provide more accurate and efficient algorithmic

support for hybrid prediction models. Additionally,

hybrid prediction models can more comprehensively

capture the complexity and nonlinear characteristics

of stock price fluctuations by combining the

advantages of different models. For example, LSTM

models are good at processing time series data, while

CNN is good at extracting spatial features. The

combination of the two can significantly improve the

accuracy and stability of predictions. In the future,

with the continuous development of model fusion

technology, hybrid prediction models will become

more intelligent and efficient. The widespread use of

big data technology has led to a sharp increasing in

the amount of data in stock market, including

historical trading data, market sentiment data,

macroeconomic data, etc. These data provide a rich

source of information for hybrid forecasting models,

helping them to more accurately capture market

trends and changes. At the same time, with the

continuous improvement of data cleaning and

preprocessing techniques, hybrid prediction models

can more effectively handle noisy data and outliers,

improving the accuracy and reliability of predictions.

Along with the continuous advancement of financial

markets and the multiplicative demand for risk

control among investors, stock price forecasting, as

an important investment decision-making tool, has

seen a continuous increase in market demand. The

hybrid prediction model, with its high accuracy and

stability, is expected to become the mainstream tool

in the field of stock price prediction in the future. In

that case, spontaneously, governments and regulatory

agencies of various countries have attached great

importance to and supported the development of

financial technology. The promotion of policies will

facilitate the application and dissemination of hybrid

forecasting models in the financial sector, providing

a broader space for their development. In addition to

deep learning technology, many emerging

technologies such as natural language processing

(NLP), knowledge graphs, etc. are also constantly

emerging and developing. These innovative

technologies will provide richer information sources

and more efficient algorithm support for hybrid

prediction models, promoting their application and

development in the area of stock price prediction.

ECAI 2024 - International Conference on E-commerce and Artiﬁcial Intelligence

572

To sum up, the development prospects of using

mixed forecasting models for stock price prediction

are very broad. With the optimization of technology,

the improvement of data quality, market demands and

policy supports, and the continuous promotion of

interdisciplinary integration and innovation, hybrid

forecasting models will play a more crucial role in

financial field, constantly providing the investors

with the most accurate and reliable decision support.

6 CONCLUSIONS

To sum up, the hybrid forecasting model is not only

playing a significant role in the field of stock prices,

but also directly or indirectly affecting the entire

financial market and other related fields. In summary,

the research on using hybrid prediction models to

predict stock prices is of great significance and has

broad prospects. In the future, string along with both

constantly innovation of technology and continuous

development of the market, hybrid forecasting

models will play a big role in the financial market.

REFERENCES

Araújo, R. D. A., Oliveira, A. L., Meira, S., 2015. A hybrid

model for high-frequency stock market forecasting.

Expert Systems with Applications, 42(8), 4081-4096.

Biswas, M., Shome, A., Islam, M. A., Nova, A. J., Ahmed,

S., 2021. Predicting stock market price: A logical

strategy using deep learning. 2021 IEEE 11th IEEE

symposium on computer applications & industrial

electronics (ISCAIE), 218-223.

Cakra, Y. E., Trisedya, B. D., 2015. Stock price prediction

using linear regression based on sentiment analysis.

2015 international conference on advanced computer

science and information systems (ICACSIS), 147-154.

Hossain, M. A., Karim, R., Thulasiram, R., Bruce, N. D.,

Wang, Y., 2018. Hybrid deep learning model for stock

price prediction. 2018 IEEE symposium series on

computational intelligence (ssci), 1837-1844.

Jakkula, V., 2006. Tutorial on support vector machine svm.

School of EECS, Washington State University, 37(2.5),

Ji, X., Wang, J., Yan, Z., 2021. A stock price prediction

method based on deep learning technology.

International Journal of Crowd Science, 5(1), 55-72.

Liu, T., Ma, X., Li, S., Li, X., Zhang, C., 2022. A stock price

prediction method based on meta-learning and

variational mode decomposition. Knowledge-Based

Systems, 252, 109324.

Ma, C., Ning, Y., Jin, H., Wu, J., 2019. The hybrid dynamic

stock forecasting model based on ANN and SVR. 2019

International Conference on Intelligent Computing,

Automation and Systems (ICICAS), 715-718.

Melo, F. M., Magnani, F. S., Carvalho, M., 2022. A

decision-making method to choose optimal systems

considering financial and environmental aspects:

Application in hybrid CCHP systems. Energy, 250,

123816.

Op't Landt, F. W., 1997. Stock price prediction using neural

networks. Maters Thesis, Leiden University.

Pinto, M. V., Asnani, K., 2011. Stock price prediction using

quotes and financial news. International Journal of Soft

Computing and Engineering (IJSCE), 1(5), 266-269.

Selvin, S., Vinayakumar, R., Gopalakrishnan, E. A., Menon,

V. K., Soman, K. P., 2017. Stock price prediction using

LSTM, RNN and CNN-sliding window model. 2017

international conference on advances in computing,

communications and informatics (icacci), 1643-1647.

Song, Y., Lee, J., 2020. Importance of event binary features

in stock price prediction. Applied Sciences, 10(5), 1597.

Sun, Y., Zhou, Q., Sun, L., Sun, L., Kang, J., Li, H., 2024.

CNN–LSTM–AM: A power prediction model for

offshore wind turbines. Ocean Engineering, 301,

117598.

Vishwanathan, S. V. M., Murty, M. N., 2002. SSVM: a

simple SVM algorithm. Proceedings of the 2002

International Joint Conference on Neural Networks.

IJCNN'02 (Cat. No. 02CH37290), 3, 2393-2398.

Xie, Y., Gui, F. X., Wang, W. J., Chien, C. F., 2021. A two-

stage multi-population genetic algorithm with

heuristics for workflow scheduling in heterogeneous

distributed computing environments. IEEE

Transactions on Cloud Computing, 11(2), 1446-1460.

Implementations of Hybrid Prediction Models for Stock Price Forecasting

573