Proceedings of the 2nd International Conference on Financial Technology and Business Analysis

DOI: 10.54254/2754-1169/52/20230758

Predicting the Direction of Stock Price Movement with

Machine Learning Algorithms
Minghan Yang1,a,*
1
Department of Mathematics, University College London, London, The United Kingdom
a. [email protected]
*corresponding author

Abstract: The application of machine learning algorithms in predicting stock price directional
movements has been a widely discussed problem. Due to the chaotic, uncertain, and dynamic
characteristics of stock markets, relatively accurate predictions could contribute to financial
benefits and risk reduction. This paper aims to assess the prediction performance of five well-
known machine learning models, which are Logistic Regression, Support Vector Machines,
Decision Trees, Random Forest, and Extreme Gradient Boosting. Data in this study covers
the period from 2020-01-02 to 2023-07-03, and assets studied in this paper are Nike, Amazon,
Microsoft, Tesco, and Airbnb. After feature selection, data pre-processing, cross-validation,
model selection, and evaluation, it is concluded that the Random Forest Classifier tends to
perform better in directional predictions, as it demonstrates higher accuracy and precision.
This research highlights the application of machine learning algorithms in financial area,
especially in the stock price prediction.

Keywords: stock price movement, prediction, machine learning algorithms

1. Introduction
Predicting the stock price movements has always been a widely discussed problem, as predictions
could offer insights for decision-making, risk management, and optimising investment returns.
However, the financial market represents a multifaceted, dynamic, chaotic, constantly evolving, and
nonlinear pattern [1]. Stock market movements are strongly linked to economic environments,
policies, and expectations from traders, as examples [2]. Hence it is argued that even a slight
improvement in prediction accuracy could lead to considerable benefits and reduce the risk in
investment [1,3].
Numerous studies on statistical time series forecasting methods have been discussed to analyze
and predict the stock price movements, such as the autoregressive integrated moving average
(ARIMA) and generalized autoregressive conditional heteroskedasticity (GARCH) models, however,
these models are constructed under the assumption of stationarity and lack the ability to react quickly
to the changes in temporary stock price movements, which may result in less satisfactory performance
[3]. Some studies also combine traditional statistical models such as ARIMA with Neural Networks,
which is said to improve prediction performance [4,5].
The applications of machine learning algorithms in stock price prediction have been studied using
a wide selection of methods and models, aiming to improve capacity, ability, and prediction
performance. Most of the machine learning algorithms considered in stock price direction forecasting
© 2023 The Authors. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0
(https://creativecommons.org/licenses/by/4.0/).

283
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

are Support Vector Machines, Logistic Regression, Random Forest, Boosting, and Neural Networks
[1]. Some popular tree-based ensemble machine learning models have also been studied. In one study,
it is suggested that AdaBoost outperforms other classifiers, such as Extra Trees, Base Classifier,
Random Forest, XGBoost, and Bagging [6]. Some other research suggested that Support Vector
Machines produce decent performance [7]. However, the best-performing models and algorithms for
the most recent stock price data have not been explored, and the selected assets in this paper were not
compared and combined in previous studies.
Hence, this paper evaluates prediction performances of five popular machine learning models
basing on the most recent stock market information, with a focus on stock price directional
movements prediction.
In the following parts of this paper, data description, feature selection, and preprocessing will be
discussed. After that, the five machine learning algorithms will be introduced in the methodology
section. Finally, the methods will be evaluated based on prediction performance, and conclusions will
be drawn.
2. Data
2.1. Data Description
This study collected stock price data from Yahoo Finance in five different industries. Most of the
five-time series cover the period from 01/01/2020 to 03/07/2023 and include Open-High-Low-Close-
Volume (OHLCV) information about the five assets. Descriptions of the five time series are presented
in Table 1.

Table 1: Descriptions of five stock price series.

Data set Industry Time Frame Number of Observation
NKE Manufacturing 2020-01-02 to 2023-07-03 881
AMZN Specialty Distribution 2020-01-02 to 2023-07-03 881
MSFT Technology 2020-01-02 to 2023-07-03 881
TSCO.L Retailer 2020-01-02 to 2023-07-03 881
ANBN Hotels/Resorts 2020-12-10 to 2023-07-03 643

In this paper, the target variable is the closing price directional movement indicator, calculated as
follows:

1, 𝑐𝑙𝑜𝑠𝑒 𝑝𝑟𝑖𝑐𝑒 𝑑𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑐𝑒 > 0, 𝑖. 𝑒. 𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒

𝑇𝑎𝑟𝑔𝑒𝑡 = { (1)
0, 𝑐𝑙𝑜𝑠𝑒 𝑝𝑟𝑖𝑐𝑒 𝑑𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑐𝑒 < 0, 𝑖. 𝑒. 𝑑𝑒𝑐𝑟𝑒𝑎𝑠𝑒

2.2. Feature Extraction and Selection

Feature extraction for machine learning methods hold great importance as the movements of stock
prices are affected by various factors, for example historic price movements, periodic patterns, and
historic market behaviors [8]. Apart from the basic features of OHLCV stock price features, some
widely used technical features include Relative Strength Index (RSI), On Balance Volume (OBV),
Moving Average Convergence Divergence (MACD), among others [9]. Based on prior research and
common features to measure stock price movement, features including SMA, RSI, MACD, OBV,
STDEV, H-L, C-O are constructed by using OHLCV data obtained, as shown in Table 2 [10,11]. The

284
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

features are normalized to assist modelling and reduce the impact of inconsistent scales of the feature
values.

Table 2: Features extracted.

H-L Difference between daily highest price and lowest price
C-O Difference between daily close price and open price
RSI A momentum oscillator to evaluate the velocity of stock price movement
SMA Average of the closing price in the past n days
MACD A momentum indicator that compares
STDEV Rolling standard deviation
On balance volume. A cumulative indicator that estimate the change of price by using
OBV
the change of volume.

In previous studies, multiple methods for feature selection and dimension reduction were
introduced. Correlation is one of the most common and straightforward criteria for feature selection
as it measures the relationship between features and the target variable [8,12]. Based on the
correlation matrix computed for each of the five stocks, the selected features are shown in Table 3.

Table 3: Features selected for each stock price time series.

Data set Feature selected based on correlation matrix
NKE Volume, H-L, C-O, RSI_14, MACDh_12_26_9, STDEV_5, OBV, SMA_50
AMZN Volume, H-L, C-O, RSI_14, MACDh_12_26_9, STDEV_5, OBV, SMA_100
MSFT Volume, H-L, C-O, RSI_14, MACDh_12_26_9, STDEV_5, OBV
Adj Close, Volume, H-L, C-O, RSI_14, MACDh_12_26_9, STDEV_5, OBV,
TSCO.L
SMA_100
ANBN Volume, H-L, C-O, RSI_14, MACDh_12_26_9, STDEV_5, SMA_50

The training sets for each time series in this paper are set to cover the period from 2020-01-02 to
2022-12-31, and the test sets include data from 2023-01-01 to 2023-07-03. A five-fold time series
cross-validation is conducted to each of the five training sets to select the best models. After that, the
models will be evaluated based on accuracy, precision, and recall, which are obtained by comparing
the predictions with the test sets.
3. Methodology
In this paper, five machine learning algorithms are considered: Logistic Regression, Support Vector
Machine, Decision Tree, Random Forest, and Extreme Gradient Boosting.
3.1. Logistic Regression Classifier (LGC)
LRC is a classification algorithm utilized in the field of machine learning and statistics, particularly
advantageous for binary classification problems. Rooted in the concept of probability, it applies a
logistic function to a linear combination of features to estimate the probability of an event. In essence,
it models the relationship between a set of independent variables and a binary dependent variable. Its
simplicity, interpretability, and computation efficiency make it a widely used approach in many
predictive modeling scenarios (See Figure 1).

285
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

Figure 1: Logistic Regression (Picture credit: Original).

3.2. Support Vector Classifier (SVC)

SVC is another powerful supervised learning model used for classification tasks. It constructs an N-
dimensional hyperplane in a transformed problem space to create the optimal decision boundary
between different classes, with the goal of maximize the distance between the nearest data points
(support vectors) of each class to the decision boundary. Its inherent abilities to manage data of higher
dimensions, generating non-linear boundaries using kernel functions, and preventing overfitting make
it an invaluable tool in complex classification problems, ranging from image recognition to biological
and medical applications. The interpretability and versatility of Support Vector Classifier make it a
well-regarded tool in machine learning (See Figure 2).

Figure 2: Support Vector Machines (Picture credit: Original).

3.3. Decision Tree Classifier (DTC)

Decision Tree Classifier operates by dividing the feature space into a set of rules, which can be
visually represented in tree-like structure of decisions. Every node in the tree stands for a decision
based on the value of an input feature. The final decision or prediction is determined by the majority
vote or average of the leaf nodes reached by the input data. With the capability to handle both
categorical and numerical data, decision trees excel in exploratory knowledge discovery. The

286
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

Decision Tree Classifier, with its interpretability and straightforward implementation, remains a
steadfast tool in predictive analytics and machine learning (See Figure 3).

Figure 3: Decision Tree (Picture credit: Original).

3.4. Random Forest Classifier (RFC)

RFC harnesses the power of several decision trees, which are based on randomly chosen subsets of
the data and features, and aggregates their outputs, typically through a majority voting system. This
method helps mitigate the risk of overfitting and boosts model generalization by introducing
randomness into the model creation process and combining the predictions of individually weak, but
collectively strong, decision trees. Consequently, the Random Forest Classifier demonstrates
exceptional performance across diverse domains, by effectively handling large datasets with high
dimensionality. With its inherent feature importance evaluation capability and its robustness to
outliers and non-linear data, it constitutes a pivotal component of machine learning (See Figure 4).

Figure 4: Random Forest (Picture credit: Original).

3.5. Extreme Gradient Boosting Classifier (XGBC)

XGBC is based on the gradient boosting framework. It capitalizes on the principle of building a
collection of predictive models sequentially, which are typically decision trees. Each subsequent
model in the sequence learns from the errors of its predecessor, aiming to minimize the overall
prediction error. XGBoost includes regularization parameters to prevent overfitting, making it a
standout in terms of both accuracy and efficiency (See Figure 5).

287
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

Figure 5: Extreme Gradient Boosting (Picture credit: Original).

4. Results
The model performance is evaluated by three factors: accuracy, precision, and recall. For Random
Forest and XGBoost, each of the accuracy, precision and recall values were calculated by taking
averages of resulting values from running models 10 times.
4.1. Accuracy
Accuracy measures the ability to correctly predict the decreases and increases, and is calculated by
the following formula:
𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (2)
𝑎𝑙𝑙 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

The accuracy scores of the five algorithms for each of the five stocks are presented in Table 4
below.

Table 4: Accuracy values.

Accuracy Nike Amazon Microsoft Tesco Airbnb
LRC 0.512 0.592 0.528 0.603 0.544
SVC 0.504 0.584 0.600 0.571 0.520
DTC 0.552 0.544 0.528 0.595 0.520
RFC 0.554 0.616 0.592 0.611 0.524
XGBC 0.546 0.552 0.544 0.588 0.496

In terms of accuracy, Random Forest Classifier appears to produce a generally better performance
than other algorithms, as it achieves higher accuracy values in predicting 3 out of 5 stocks.
4.2. Precision
Precision is evaluated by the following formula:
𝑡𝑝
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (3)
𝑡𝑝+𝑓𝑝

288
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

where 𝑡𝑝 is the true positive, and 𝑓𝑝 is the false positive. The precision values measure the proportion
of true increases among all the predicted increases. Hence higher precision could represent a lower
likelihood of false investment choices based on the incorrect predictions of price increases, thereby
reducing the risk of losses due to erroneous investments.
From Table 5, it is evident that the Random Forest Classifier generally demonstrates decent
performance, achieving the highest precision values in predicting the stock price movements of
Amazon, Tesco, and Airbnb. In contrast, the Decision Tree Classifier seems to yield the poorest
prediction performance. The highest precision value, 0.71, is observed when forecasting the direction
of Tesco's stock prices using the Random Forest Classifier.

Table 5: Precision values.

Precision Nike Amazon Microsoft Tesco Airbnb
LRC 0.49 0.59 0.52 0.63 0.59
SVC 0.48 0.57 0.60 0.64 0.56
DTC 0.50 0.57 0.53 0.60 0.58
RFC 0.50 0.65 0.57 0.71 0.59
XGBC 0.53 0.58 0.54 0.67 0.55

4.3. Recall
Recall values are calculated by using the following formula:
𝑡𝑝
𝑟𝑒𝑐𝑎𝑙𝑙 = (4)
𝑡𝑝+𝑓𝑛

where 𝑓𝑛 is the false negative. It calculates the proportion of correctly forecasted increases out of the
total number of actual increases. Higher recall values indicate a higher ability of capturing the increase
in stock prices, and a lower likelihood of missing investment opportunities.
From table 6, it is noticeable that Support Vector Classifier seems to generate an overall better
performance, by achieving highest recall values when predicting Nike, Amazon and Airbnb stock
price movements. The highest precision value takes place when predicting the Amazon stock price
movement direction by using Support Vector Classifier.

Table 6: Recall values.

Recall Nike Amazon Microsoft Tesco Airbnb
LRC 0.87 0.87 0.83 0.63 0.61
SVC 0.96 0.97 0.67 0.48 0.71
DTC 0.93 0.72 0.73 0.72 0.64
RFC 0.69 0.67 0.80 0.47 0.59
XGBC 0.60 0.65 0.73 0.48 0.59

It is highlighted that a high recall value and a high precision value cannot be obtained at the same
time in any case. In this paper, precision performance is prioritized over recall performance, as
investors often seek to reduce risk rather than missing potential profit opportunities. In light of this,
the Random Forest Classifier is likely to outperform other algorithms, which agrees with the findings
from previous studies [1,13].

289
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

5. Discussion
In order to produce a more inclusive and general evaluation of the models’ performances, it is
recommended to consider a broader range of assets in the study. In a study evaluating multiple
classifiers for stock price movements, stock price data of 5767 companies were considered in the
study, covering 19 industries [1]. Including a larger and more diverse set of assets would allow for a
more generalized evaluation and comparison of model performance, even though the assets
considered in the study do not encompass the entire selection of all stocks in the financial market.
Additionally, it could be advantageous to incorporate other financial indicators such as liquidity,
solvency, profitability indicators, and general features such as public debt, GDP, unemployment, and
trade balance, into the models [1].
One limitation of this study is that it considered only five most common machine learning
algorithms. Hence, in future work, more advanced algorithms such as Neural Networks, AdaBoosting,
and other deep learning models could be included in comparison and evaluation.
6. Conclusion
This project evaluates the performances of five machine learning algorithms in predicting stock price
directional movements. Five assets, NKE, AMZN, MSFT, TSCO.L, and ABNB are included in this
study. Each time series, obtained from yahoo finance, contains OHLCV data within period
2020/01/02 – 2023/07/03. Technical features such as RSI and MACD are constructed and selected
based on correlation matrix. After normalization, each time series is split into training and test sets.
Then five-fold cross validation is applied to the training set to evaluate the models. Following that,
the models are evaluated using accuracy, precision, and recall on the test set. The results indicate that
the Random Forest Classifier provides superior prediction performance on the test set compared to
the other models.
References
[1] Ballings, M., Van den Poel, D., Hespeels, N., and Gryp, R. (2015) Evaluating multiple classifiers for stock price
direction prediction. Expert systems with Applications, 42(20), 7046-7056.
[2] Kumar, M., and Thenmozhi, M. (2006) Forecasting stock index movement: A comparison of support vector
machines and random forest. In Indian institute of capital markets 9th capital markets conference paper.
[3] Timmermann, A., and Granger, C. W. (2004) Efficient market hypothesis and forecasting. International Journal of
forecasting, 20(1), 15-27..
[4] Khandelwal, I., Adhikari, R., and Verma, G. (2015) Time series forecasting using hybrid ARIMA and ANN models
based on DWT decomposition. Procedia Computer Science, 48, 173-179.
[5] Zhang, G. P. (2003) Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing,
50, 159-175
[6] Ampomah, E. K., Qin, Z., and Nyame, G. (2020) Evaluation of tree-based ensemble machine learning models in
predicting stock price direction of movement. Information, 11(6), 332.
[7] Lee, M. C. (2009) Using support vector machine with a hybrid feature selection method to the stock trend prediction.
Expert Systems with Applications, 36(8), 10896-10904.
[8] Htun, H. H., Biehl, M., & Petkov, N. (2023). Survey of feature selection and extraction techniques for stock market
prediction. Financial Innovation, 9(1), 26.
[9] Alsubaie, Y., El Hindi, K., and Alsalman, H. (2019) Cost-sensitive prediction of stock price direction: Selection of
technical indicators. IEEE Access, 7, 146876-146892.
[10] Khalid, S., Khalil, T., and Nasreen, S. (2014) A survey of feature selection and feature extraction techniques in
machine learning. In 2014 science and information conference, 372-378.
[11] Nabipour, M., Nayyeri, P., Jabani, H., Shahab, S., and Mosavi, A. (2020) Predicting stock market trends using
machine learning and deep learning algorithms via continuous and binary data; a comparative analysis. IEEE
Access, 8, 150199-150212.
[12] Nabi, R. M., Saeed, S. A. M., Harron, H. B., and Fujita, H. (2019) Ultimate prediction of stock market price
movement. Journal of Computer Science, 15, 1795.

290
 Proceedings of the 2nd International Conference on Financial Technology and Business Analysis
DOI: 10.54254/2754-1169/52/20230758

[13] Patel, J., Shah, S., Thakkar, P., and Kotecha, K. (2015) Predicting stock and stock price index movement using
trend deterministic data preparation and machine learning techniques. Expert systems with applications, 42(1),
259-268.

291