Recent Methods in Stock Price Prediction: A Review

Yuxuan Li

College of Literature, Science and the Arts, University of Michigan, S State St, Ann Arbor, U.S.A.

Keywords: Stock Price Prediction, Neural Networks, Hybrid Models, ARIMA, BiCuDNNLSTM-1dCNN Model.

Abstract: This paper evaluates various methodologies for predicting stock prices, from traditional models such as the

Autoregressive Integrated Moving Average (ARIMA) to more advanced neural networks. It addresses the

inadequacies of the ARIMA model, particularly its limitations in predicting the future, which are

commonplace in dynamic financial markets. Then, it introduces a hybrid model, the Bidirectional Cuda Deep

Neural Network Long Short-Term Memory combined with a one-dimensional Convolutional Neural Network

(BiCuDNNLSTM-1dCNN). This model excels at capturing both the long-term trends and short-term

fluctuations essential for accurate financial forecasting. Through extensive preprocessing, the model ensures

the highest quality of input data, leading to more reliable predictions. Comparative results demonstrate that

the BiCuDNNLSTM-1dCNN model significantly surpasses both ARIMA and simpler neural networks in

accuracy and reliability. The paper concludes with a call for continued advancement of hybrid modeling

techniques to enhance the precision of forecasts and empower data-driven investment strategies in volatile

markets.

1 INTRODUCTION

The stock market functions via a network of

exchanges where shares of publicly traded

corporations are transacted. Investors depend

significantly on analytical insights to make prudent

financial decisions (Billah, Sultana, Bhuiyan, &

Kaosar 2024). Analyzing a company’s performance

through data is considered essential before making

any investment. This emphasizes the significant role

of data-driven strategies in helping investors evaluate

a company’s potential and make informed financial

commitments. The Autoregressive Integrated

Moving Average (ARIMA) model is a standard

method for forecasting stock values. This approach

integrates three components: autoregression (AR),

which forecasts future prices using historical data;

moving averages (MA), which smoothen fluctuations

by factoring in past errors; and differencing (I), which

stabilizes non-stationary data. While ARIMA

provides accurate short-term predictions due to its

simplicity, it struggles with non-linear patterns often

present in financial markets. As described, the

ARIMA model for stationary time series follows the

structure, where future values are estimated using a

combination of historical data points and previous

forecasting errors (Hossain et al. 2018). Although

effective for short-term predictions, ARIMA often

fails to account for the complexities of fast-changing

market trends. Although ARIMA offers reliable

results for short-term forecasts, it often fails to

capture the complexities of rapidly changing market

trends. In recent years, neural networks have become

prominent for their capacity to discern complex

patterns and relationships in financial data.

Researchers (Zhao, Hu, Liu, Lan, & Zhang, 2023)

show that analyzing historical stock prices through

recurrent neural networks (RNNs) or their variants

improves forecasting accuracy. These models

partition data into intervals, allowing them to uncover

meaningful patterns and predict future trends. Given

the growing reliance on neural networks for stock

price forecasting, this paper provides a detailed

review of various neural network models. It evaluates

their performance in predicting stock prices and

explores ways to enhance forecasting precision

through hybrid architectures, such as the

BiCuDNNLSTM-1dCNN model.

552

Li and Y.

Recent Methods in Stock Price Prediction: A Review.

DOI: 10.5220/0013528200004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 552-556

ISBN: 978-989-758-754-2

2 EVALUATING NEURAL

NETWORK ARCHITECTURES

FOR STOCK FORECASTING

2.1 CNN in Stock Price Forecasting

Convolutional Neural Networks (CNNs) are a

category of deep learning models designed to

examine structured inputs, such as photographs, by

emulating the activities of the visual cortex in

animals. These networks autonomously acquire and

identify spatial patterns, varying from simple to

intricate. A CNN has three fundamental types of

layers: convolutional, pooling, and fully connected

layers. The convolutional layers perform specialized

linear operations to identify features while pooling

layers help reduce dimensionality. Fully linked layers

translate the gathered features into outputs, including

class predictions (Yamashita, Nishio, Do, & Togashi,

2018).

When applied to stock price forecasting,

CNNs have demonstrated effectiveness in extracting

localized features from time-series data. In advanced

models like BiCuDNNLSTM-1dCNN, CNNs play a

vital role by identifying patterns within the data, and

facilitating accurate predictions (Selvin et al., 2017).

Nonetheless, despite their efficacy in capturing short-

term dependencies, CNNs are limited in modeling

long-term trends. The challenge has resulted in the

development of hybrid approaches, such as the CNN-

LSTM model, which integrates the feature extraction

capabilities of CNNs with the temporal dependency

modeling of Long Short-Term Memory (LSTM)

networks to proficiently capture both short- and long-

term dependencies (Hiransha et al., 2018).

2.2 RNN and Its Role in Sequential

Data Analysis

RNNs are designed for the processing of sequential

input through a loop structure that retains information

from previous phases. These networks perform

repetitive computations on each element within a

sequence, where the outcome at each step depends on

both the current input and the results from previous

steps (Shiri, Perumal, Mustapha, & Mohamed, 2023).

These abilities make RNNs useful in forecasting

stock prices, as they can identify temporal patterns

and correlations in financial data, helping the model

analyze how historical prices influence future trends.

However, basic RNNs face challenges related

to limited memory retention, which reduces their

effectiveness when processing long sequences. This

shortcoming, known as short-term memory, arises

from disappearing or exploding gradient issues

during training, especially with extensive datasets. As

a result, while RNNs are well-suited for short-term

predictions, they struggle with capturing complex

dependencies over longer periods—an essential

requirement for accurate stock forecasting. To

alleviate these limitations, sophisticated models like

LSTM networks and Gated Recurrent Units (GRUs)

were created. These models incorporate memory

mechanisms that allow the network to save and utilize

essential information over prolonged sequences,

hence enhancing their effectiveness in making

financial predictions.

2.3 Advancements with LSTM

Networks

LSTM networks are widely used in deep learning

models, notably acknowledged for their capability to

capture long-term dependencies in serial data, such as

stock prices. LSTM models have been prevalent for

financial prediction because of their ability to capture

temporal correlations among data points (Chung &

Shin, 2018). Their ability to remember pertinent

information facilitates jobs that necessitate the

examination of historical trends across time. The

Grey Wolf Optimization-Elman Neural Network

(GWO-ENN) model references LSTM as one of the

benchmark models in its evaluation of various stock

forecasting techniques. While LSTM models

generally achieve competitive results based on

metrics like mean square error (MSE), the GWO-

ENN model, in this study, exhibited superior

performance for one-day-ahead forecasts (Chung &

Shin, 2018).

Compared to simpler models like the Elman

Neural Network (ENN), which performs well for

short-term memory retention (Zheng, 2015), LSTMs

offer a more advanced approach by incorporating cell

states and gating mechanisms. These features enable

the model to make decisions on whether to keep or

discard information, ensuring that relevant data is

preserved for future use. Hybrid models, such as GA-

LSTM— which combines genetic algorithms with

LSTM networks—further improve forecasting

accuracy, outperforming traditional approaches in

stock market predictions (Chung & Shin, 2018).

LSTMs have proven particularly effective in

predicting stock prices over extended periods,

ranging from multiple days to even longer

timeframes. Although ENN models are suitable for

short-term tasks, such as forecasting the next day’s

price, LSTM networks excel when it comes to

Recent Methods in Stock Price Prediction: A Review

553

capturing long-term trends, thanks to their

sophisticated memory mechanisms

2.4 Incorporating GRU for Enhanced

Efficiency

The Structure of the GRU Neural Network According

to CEEMDAN-Wavelet is an advanced technique

aimed at improving the precision of time series

forecasting, (Qi, Ren, & Su 2023). Especially in the

financial market, including stock index prediction.

The model combines the CEEMDAN (Complete

Ensemble Empirical Mode Decomposition with

Adaptive Noise) signal processing technique with the

Wavelet Denoising method with the GRU neural

network to make a financial forecast.

Initially, the original input noised data is

deconstructed utilizing the CEEMDAN approach.

CEEMDAN decomposes the signal into many

frequency components known as Intrinsic Mode

Functions (IMFs). The high-frequency intrinsic mode

functions frequently encompass greater noise and

demonstrate a diminished signal-to-noise ratio

(SNR). Nevertheless, these high-frequency

components may include significant short-term

information that can affect the precision of forecasts.

Wavelet denoising is utilized instead of outright

discarding them. The wavelet thresholding method

distinguishes valuable signal information from noise

by examining the wavelet coefficients, which denote

the intensity of various frequency components within

the signal. By establishing a suitable threshold, the

framework eliminates noise while preserving critical

features, so ensuring that essential information is not

compromised during the preprocessing phase.

After denoising the high-frequency

components, the subsequent step is to amalgamate

them with the intermediate and low-frequency IMFs,

which generally exhibit reduced noise and

encapsulate the long-term patterns of the data. This

component combination process yields a

reconstructed, denoised signal that embodies both the

short-term and long-term attributes of the original

data. Therefore, the Wavelet denoising method

retains critical information across all frequencies,

facilitating precise forecasting of the data.

During the second phase of the framework, the

purified and denoised data is input into a GRU neural

network for forecasting. GRUs are especially adept at

processing time series data as they effectively learn

long-term dependencies, managing sequential

information across time while avoiding

complications such as the vanishing gradient problem

that can impact conventional RNNs. The GRU

network manages the denoised data by regulating the

retention of prior information via its update gate and

the elimination of information through its reset gate.

This enables the GRU to constantly equilibrate the

retention of pertinent historical knowledge with the

assimilation of fresh input to produce precise

predictions regarding future patterns, such as stock

prices.

The application of CEEMDAN and wavelet

denoising guarantees that the GRU model is provided

with superior input data. The framework enhances the

forecasting accuracy of the GRU by efficiently

distinguishing noise from the signal. This is

especially beneficial in financial forecasting, where

erratic data may result in imprecise projections.

Furthermore, the GRU's more straightforward

architecture relative to LSTMs enhances its

computational efficiency, while yet preserving the

capacity to catch both short-term variations and long-

term trends in the data. The integration of

sophisticated denoising methods and a robust

predictive model renders the system exceptionally

useful for time series forecasting, particularly in

contexts such as stock market prediction, where data

is frequently noisy yet holds significant information.

The CEEMDAN–wavelet model underwent

thorough testing through ten trials on two prominent

financial indices, the CSI300 and the S&P 500, to

assess its predictive accuracy (Qi et al., 2023). The

model's performance in these tests was evaluated

using Mean Squared Error (MSE) and Mean Absolute

Error (MAE), which are conventional metrics for

measuring prediction accuracy. The model integrates

CEEMDAN (Complete Ensemble Empirical Mode

Decomposition with Adaptive Noise) and wavelet

denoising methods to preprocess the data, efficiently

distinguishing valuable signal information from

noise. The denoised data serves as input for a GRU

neural network, which learns and forecasts future

trends. Comparative evaluations were performed

against other prominent models, including LSTM,

GRU, CNN-BiLSTM, and ANN. Statistical research,

including t-tests, demonstrated that the CEEMDAN–

wavelet model consistently attained the lowest MSE

and MAE values, exhibiting significant differences

from the other models. The CEEMDAN–wavelet

model successfully manages noisy data and captures

intricate patterns, rendering it more accurate and

dependable for time series forecasting compared to

the classical and neural network models evaluated.

DAML 2024 - International Conference on Data Analysis and Machine Learning

554

2.5 A Hybrid Model

The BiCuDNNLSTM-1dCNN model, presented by

(Kanwal, Lau, Ng, Sim, & Chandrasekaran, 2022)

commences with a crucial data preparation phase.

Given the inconsistent nature of financial datasets,

particular attention is given to handling missing

values (NaNs). To maintain the integrity and

continuity of the data stream, NaN values are replaced

with the average of adjacent values (Kanwal et al.,

2022). Furthermore, normalization is executed via the

Min-Max scaler, which adjusts the feature values to a

range of 0 to 1, thus enhancing the model's

performance and computational efficiency. Once

preprocessing is complete, the data is divided into

90% for model training and 10% for testing the

predicted accuracy.

The configuration of hyperparameters is the

initial phase in the complex training process of

BiCuDNNLSTM-1dCNN. A Dropout layer is

utilized to eliminate nodes at a rate of 20% during

training to mitigate overfitting, alongside a

bidirectional CuDNNLSTM layer that proficiently

collects features from sequential input, followed by a

one-dimensional CNN layer that identifies sudden

market swings, crucial for capturing quick changes in

the stock market. (Kanwal et al., 2022).

Hyperparameter tuning is performed with a random

search strategy, which navigates the parameter space

more effectively than conventional grid search

methods, resulting in decreased computational time

and enhanced model performance (Bergstra &

Bengio, 2012).

The practical execution of the model

commences with the preparation of stock market data,

encompassing open, maximum, minimum, and

closing prices, along with trading volume, to ensure

uniform inputs for training. The model uses a

lookback window of 50-time steps to examine

historical patterns, enabling the BiCuDNNLSTM

component to leverage its bidirectional functionality

effectively.

The BiCuDNNLSTM-1dCNN model has

consistently demonstrated superior performance

across various datasets, including individual stocks

like Crude Oil and DAX ETF, as well as major

indices such as GDAXI and HSI. It outperforms other

models, including CNN, RNN, and standard LSTM,

with lower RMSE and MAE values, indicating

enhanced accuracy (Kanwal et al., 2022). This model

functions as a dependable instrument for financial

analysts and traders, demonstrating the efficacy of

sophisticated machine-learning methodologies in

intricate financial forecasting endeavors. By

integrating LSTM and CNN architectures, the model

effectively combines temporal sequence analysis

with pattern recognition, establishing

BiCuDNNLSTM-1dCNN as a robust solution for

stock market prediction in volatile environments.

3 CONCLUSIONS

Forecasting stock prices has always presented a

difficulty due to the inherent complexity and

volatility of trade markets. Although conventional

methods such as ARIMA are proficient for short-term

forecasts, they frequently fail to account for the

nonlinear dynamics inherent in stock data.

Conversely, neural networks—such as CNNs, RNNs,

LSTMs, and GRUs—have arisen as formidable

instruments for stock price prediction, delivering

superior performance through modeling intricate

temporal dependencies and nonlinearities. The

creation of hybrid architectures, shown by the

BiLSTM-CNN model, enhances prediction accuracy

and stability by integrating the advantages of various

neural network models.

CNNs excel at recognizing limited, transient

patterns in time series data, rendering them effective

for forecasting rapid market fluctuations.

Nonetheless, CNNs independently may encounter

difficulties in capturing long-range relationships,

hence constraining their utility in contexts

necessitating extensive trend analysis. RNNs,

especially LSTMs, and BiLSTMs, excel in handling

sequential dependencies by retaining crucial

information over extended periods, making them

popular choices for financial forecasting. GRUs

provide a more efficient solution by streamlining the

architecture, diminishing computing complexity, and

accelerating training without compromising

accuracy, rendering them suitable for situations

necessitating rapid, dependable predictions.

In conclusion, neural network models,

particularly hybrid systems that amalgamate various

designs, have continuously surpassed traditional

forecasting methods. Their capacity to comprehend

intricate patterns, retain vital information over time,

and adeptly handle sequential dependencies enables

them to deliver enhanced accuracy and reliability,

which is crucial for investors making data-driven

judgments in the dynamic and unpredictable realm of

finance. As financial markets progress, the demand

for advanced forecasting techniques will increase.

Consequently, continuous research and development

in machine learning, especially in the creation and

enhancement of hybrid neural network models, will

Recent Methods in Stock Price Prediction: A Review

555

be crucial. These models can provide profound

insights, adjust to fluctuating market conditions, and

deliver more accurate predictions, allowing investors

to make informed decisions that enhance their

returns. Utilizing these improvements, the future of

stock price forecasting will probably experience

enhanced incorporation of artificial intelligence,

promoting data-driven investment techniques that are

resilient and adaptable to market complexity.

REFERENCES

Billah, M. M., Sultana, A., Bhuiyan, F., & Kaosar, M. G.

2024. Paper templates. Stock price prediction:

comparison of different moving average techniques

using deep learning model. NEURAL COMPUTING

AND APPLICATIONS, 36(11), 5861-5871.

Kanwal, A., Lau, M. F., Ng, S. P., Sim, K. Y., &

Chandrasekaran, S. (2022). BiCuDNNLSTM-1dCNN—

A hybrid deep learning-based predictive model for

stock price prediction. EXPERT SYSTEMS WITH

APPLICATIONS, 202, 117123.

Kumar Chandar, S. (2021). Grey Wolf optimization-Elman

neural network model for stock price prediction. SOFT

COMPUTING, 25, 649-658.

Ma, Q. (2020). Comparison of ARIMA, ANN and LSTM for

stock price prediction. IN E3S WEB OF

CONFERENCES (VOL. 218, P. 01026). EDP

SCIENCES.

Qi, C., Ren, J., & Su, J. (2023). GRU neural network based

on CEEMDAN–wavelet for stock price

prediction. APPLIED SCIENCES, 13(12), 7104.

Shiri, F. M., Perumal, T., Mustapha, N., & Mohamed, R.

(2023). A comprehensive overview and comparative

analysis on deep learning models: CNN, RNN, LSTM,

GRU. ARXIV PREPRINT ARXIV:2305.17473.

Yamashita, R., Nishio, M., Do, R. K. G., & Togashi, K.

(2018). Convolutional neural networks: an overview

and application in radiology. INSIGHTS INTO

IMAGING, 9, 611-629.

Zhao, C., Hu, P., Liu, X., Lan, X., & Zhang, H. (2023).

Stock market analysis using time series relational

models for stock price

prediction. MATHEMATICS, 11(5), 1130.

DAML 2024 - International Conference on Data Analysis and Machine Learning

556