BANCA Y FINANZAS

Accounting information and stock returns in

Vietnam securities market: Machine learning
approach
Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao
Hanoi University of Industry, Vietnam
National Economics University, Vietnam
National Economics University, Vietnam

This paper studies the relationship between accounting information reflected in financial state-
ments and stock return in Vietnam Stock Market. The authors propose a research model to define
the relationship. Besides, machine learning algorithms are used in researching and forecasting,
with data obtained from observed firms during the period from 2009 to 2020. Research results
show that gradient boosting algorithm has the best self-reporting performance and finan-
cial ratios also have great impact on stock returns, including operating income growth, stock
earnings volatility, dividend yield, earnings before tax-to-equity ratio, cash holding ratio, and
accrual quality. Based on the research results, the authors make some recommendations for
investors, firms, and policy makers.

Keywords: stock return, machine learning, decision tree, gradient boosting, elastic net, accoun-
ting information, accrual quality

Información contable y rendimiento de acciones en el mercado de valores de

Vietnam: enfoque de aprendizaje automático

Este artículo estudia la relación entre la información contable reflejada en los estados financie-
ros y el rendimiento de las acciones en el mercado de valores de Vietnam. Se propone un modelo
de investigación para definir la relación. Además, los algoritmos de aprendizaje automático se
utilizan en la investigación y la previsión, con datos obtenidos de empresas observadas durante

[Link]
Contabilidad y Negocios (17) 33, 2022, pp. 94-118 / e-ISSN 2221-724X
 Accounting information and stock returns in Vietnam securities market 95

el período de 2009 a 2020. Los resultados de la investigación muestran que el algoritmo gradient
boosting tiene el mejor rendimiento de autoinforme y los índices financieros también tienen un
gran impacto en la rentabilidad de las acciones, al incluir el crecimiento de los ingresos operati-
vos, la volatilidad de las ganancias de las acciones, el rendimiento de los dividendos, la relación
entre las ganancias antes de impuestos y el capital, la relación de tenencia de efectivo, y la
calidad de la acumulación. Sobre la base de los resultados de la investigación, se hacen algunas
recomendaciones para inversionistas, empresas y formuladores de políticas.

Palabras clave: rentabilidad de acciones, aprendizaje automático decision tree, gradient boos-
ting, elastic net, información de cuenta, calidad devengada

Informações contábeis e retornos de ações no mercado de ações do Vietnã:

abordagem de aprendizado de máquina

Este artigo estuda a relação entre as informações contábeis refletidas nas demonstrações
financeiras e o retorno das ações no mercado de ações do Vietnã. Os autores propõem um
modelo de pesquisa para definir a relação. Além disso, algoritmos de aprendizado de máquina
são usados em pesquisas e previsões, com dados obtidos de empresas observadas durante o
período de 2009 a 2020. Os resultados da pesquisa mostram que o algoritmo gradient boosting
tem o melhor desempenho de autorrelato e os índices financeiros também têm grande impacto
nos retornos das ações, incluindo crescimento da receita operacional, volatilidade do lucro
das ações, rendimento de dividendos, lucro antes do índice de impostos sobre o patrimônio
líquido, índice de retenção de caixa e qualidade de acumulação. Com base nos resultados da
pesquisa, os autores fazem algumas recomendações para investidores, empresas e formula-
dores de políticas.

Palavras-chave: retorno de ações, aprendizado de máquina, decision tree, gradient boosting,

elastic net, informação contábil, qualidade da provisão

1. Introduction
Stock returns have been a topic of scientific research interest for many years (Ball
& Brown, 1968; Basu, 1983; Freeman, 1987; Collins & Kothari, 1989; Easton & Harris,
1991) have built yield models from income and return variables, by testing the rela-
tionship between current accounting earning divided by stock price at the beginning
of the period and stock return. Investors aim to maximize the value of their assets
by investing in firm shares as they offer the most significant potential for long-term

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

96 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

returns based on a trade-off of return and risk. Anwaar (2016) believes that a firm’s
financial and accounting information can be used effectively to evaluate investment
opportunities for investors. A firm’s information is divided into internal information
and external information, based on the source of the information (Emamgholipour et
al., 2013). As the name implies, external information is taken from the stock market,
while internal information appears on financial statements. Therefore, investors can
rely on accounting information on financial statements to evaluate stock returns over
time when making investment decisions.

Previous financial research literature contains debates about the effectiveness

of accounting information in forecasting stock returns. Previous studies have been
conducted over several decades, and there is still no consensus on which of the indica-
tors of accounting information seems to be most closely associated with stock returns.
For example, Barnes (1987) argues that the indicators commonly used to predict stock
returns are relevant indicators to a firm’s financial and operating performance and
can help predict future performance of the firm. According to Lewellen (2004), stock
returns are predictable. Similarly, a research by Delen et al. (2013) shows that the
most important financial ratio used to predict firm performance is net profit margin.
The authors use decision tree algorithms to evaluate the relationship between finan-
cial ratios and firm performance. In addition, the results from a research by Musallam
(2018) show that earnings per share, dividend yield and earnings yield are strongly
related to stock returns.

Recently, there is a promising research field which focuses more on artificial intel-
ligence techniques and can handle non-linearity and stochasticity. Improvements in
technology and computer science have spread implementations to a wider audience,
spurring the growth of research in machine learning. Financial market prediction
research is done by applying several machine learning algorithms, including gradient
boosting, support vector machine, and random forest to predict price, profitability,
direction and volatility of securities, stocks and commodities indicators (Henrique
et al., 2019).

Research on evaluating stock returns of Vietnamese firms using machine learning

algorithms, together with accounting information, has not been studied yet. Besides,
the literature is not consistent in previous studies in the world about whether financial
ratios are closely related to stock returns. Forecasting a firm’s stock return is certainly
a difficult task, and even more so when just simply applying a series of financial indi-
cators take from accounting information. However, it is a demanding task for deci-
sion makers, including stocks analysts, creditors, portfolio managers, and investors,

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 97

to determine which indicators are suitable for making predictions. Therefore, there
is a need to study the impact of accounting information on stock returns on Vietnam
Stock Market.

The study aims to give an overview of the relationship between accounting infor-
mation, through relevant financial ratios, and stock returns prediction by answering
the research question and the following questions: “How is the accounting information
on financial statements reflected in financial ratios associated with the stock returns
on Vietnam Stock Market?”. Furthermore, the following detailed questions are formu-
lated to answer whether certain financial ratios have a stronger association with stock
returns than others. To determine the association, we use and compare machine learn-
ing models to each other to answer the questions: Which financial ratios have the
strongest association with stock returns? Which applied machine learning model has
the highest predicting accuracy?

Literature review
Using accounting information is perhaps the most common way to evaluate firm
performance and provides essential inside information about firm value and market
capitalization. Accounting information is widely used for many years all over the world,
especially to predict the future performance of firms. Additionally, it can be used as
a tool to predict stock returns in the stock market (Lewellen, 2004). It is also used to
determine a firm’s creditworthiness.

Financial theory suggests that financial ratios can provide an insight into a firm’s
operating and financial performance by rearranging accounting information on finan-
cial statements such as returns and firm value, its ability to allocate cash flows to
investments, its liquidity and its debt size. Financial ratios are also used for future
performance prediction. For example, they are used as explanatory variables in predic-
tive modeling to predict financial distress, failure, and bankruptcy (Sun et al., 2011;
Zięba et al., 2016; Kim & Upneja, 2014). Similarly, financial ratios are used in modeling
the relationship between them and stock returns, by using multiple or sets of indi-
vidual financial variables or using statistical methods or machine learning techniques
(Emamgholipour et al., 2013; Lewellen, 2004). Although most of the previous studies are
successful in modeling the relationship between financial ratios and firm returns, they
lack determination about the importance features of these ratios used to evaluate firm
performance.

In Vietnam, there have been a number of studies by Hai et al. (2015), Ha et al. (2018),
Hung et al. (2018), and Hung and Van (2020) showing the impact of accounting infor-

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

98 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

mation on stock returns. However, studies which are based on machine learning algo-
rithms and the importance of financial indicators on stock returns have not been yet
considered. Results from previous studies show evidence of relationship between
accounting information and returns. Yet, those studies are based on various theories
and models. Each model has its own advantages and disadvantages, and the data
collection methods are different. This study will determine which algorithm is most
suitable in determining the impact of accounting information on stock returns, and
which financial ratios have the strongest impact.

3. Research methodology
3.1. Research model

From the studies presented in the literature review such as Lewellen (2004), Delen et
al. (2013), Emamgholipour et al. (2013), Musallam (2018), Hai et al. (2015), Ha et al. (2018),
Hung et al. (2018), Hung and Van (2020), and Metsomäki (2020), the authors establish
the model as follows:

STOCK_RETURNsit = α + βInformation Accoutingit + εit

Stock return is a ratio that measures the percentage of a stock’s net income per
dollar of invested capital. The formula for the stock returns is the appreciation in the
price plus any dividends paid, divided by price at the beginning of the period. This
calculation method is frequently mentioned in studies on stock returns and in the
original model of Easton and Harris (1991), which is:

(Pit -Pit-1 ) +Dit

STOCK_RETURNsit =
Pit-1

Pit-1 , Pit : price of share at the end of year t and t-1, respectively
Dit : dividend per share for the year t.

This study examines 11 financial indicators of accounting information on financial

statements, including Debt-to-equity ratio, Earnings before tax-to-equity ratio, Total
assets turnover, Fixed assets turnover, Operating income growth, Stock earnings vola-
tility, Cash holding ratio, Dividend yield, Earnings management, Accrual quality, Earn-
ings per share. The measurement of these financial indicators are described briefly in
appendix 1. To conduct this study, we used Python 3.8 to analyze, forecast and evaluate
the models.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 99

3.2. Research algorithms

3.2.1. Linear models

Linear regression: As the first linear model, linear regression with the ordinary least
square method is usually implemented. its aim is to minimize the sum of squares
between the true and estimated values by fitting the linear model with coefficients
(Pedregosa et al., 2011).

Ridge regression and Lasso regression are two regression models that apply regu-
larization techniques to avoid overfitting. Overfitting is a phenomenon where the
model only fits well on training dataset but does not predict well on test data. This is
often the case when training machine learning models. This phenomenon has a nega-
tive effect and leads to inapplicable models because the predictions turn wrong when
applied in practice. There are many causes of overfitting. One of the common causes is
that the training dataset and the test dataset have different distribution, making the
rules learned in the training dataset no longer valid in the test dataset, or it is because
when a model has too many parameters, its data is not representative. Regularization
is a technique to avoid overfiting by adding a regularization term to the loss function.
Usually, this term is in the form of L1 norm or L2 norm of the coefficients. They are
called ridge regression and Lasso regression respectively.

For these regressions, we need to fit coefficient α to find a coefficient that is best
for each dataset. If data is severely overfitting, it is necessary to reduce overfitting by
increasing the effect of the regularization term by increasing coefficient α. If the model
is not overfitting, α can be close to 0. If α = 0, the regression equation is equivalent to
multivariable linear regression.

Elastic net: To add more explanatory power to the family of linear models, elastic
net (Zou & Hastie, 2005) was applied. Firstly, it is a continuation of linear regression
models trained with Lasso’s L1 penalty and Ridge’s L2 penalty. Combining the penalties
of both methods in one model produces a regular, competitive model in which the
weight of parameters is non-zero (Pedregosa et al., 2011).

3.2.2. Decision tree

Decision tree: Decision tree is a classification model introduced by Belson (1959),

widely used in various fields. After the introduction of the machine learning method
system, decision tree was further developed with C4.5 algorithm by J. Ross Quinlan
(1996) and ID3 algorithm by J. Ross Quinlan (1986). Decision tree is a structured classifi-
cation tree that classifies objects based on sequences of rules. Independent variables

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

100 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

and ­attributes can be of different data types such as binary, nominal, ordinal, quanti-
tative data. To determine which variable is classified first, which variable is classified
later, the weight of evidence (Entropy) for each variable is calculated, the higher the
information value, the more classification information the variable carries.

Random forest: Random forest is an attribute classification method developed by

Leo Breiman at the University of California, Berkeley. Breiman is also the co-author
of CART (classification and regression trees) method, which is one of ten data mining
methods. In Random Forest, a significant improvement in classification accuracy is the
result from the growth of a set of trees, each of which will then “vote” for the most
popular class. Normally, to develop these sets of trees, random vectors are gener-
ated, which will govern the growth of each tree in the aforementioned sets. For the k th
tree, a random vector Vk is generated, which is independent from previously generated
vectors V1, V2, ..., Vk-1, although the distribution of these vectors is similar. A tree is
grown based on the training set and vector Vk, resulting in a subclass h(x, Vk) where x
is the input vector. After a large number of trees are created, these trees will “vote” for
the most popular class.

AdaBoost: Boosting is a technique that uses a combination of machine learning

algorithms on a sample space sequentially, then combines separate classification
results to get an effective classifier. An efficient algorithm in boosting is AdaBoost
(adaptive boosting), which uses classification error weights assigned to each sample.
Firstly, the algorithm classifies equivalent weights on each training sample. In each
iteration, the algorithm: (i) trains the sample by a weak classifier; (ii) checks whether
the classification results on the training sample are correct; (iii) recalculates classifica-
tion error weights on the sample by: increasing error weights on misclassified samples
and decreasing error weights on properly classified samples. After finishing the loop,
the algorithm will combine the sub-classifiers to construct a composite classifier.

Gradient boosting: Gradient boosting is also an algorithm that uses a combination

of boosting methods to develop an advanced prediction tool. In many ways, gradient
boosting is similar to AdaBoost, but with a few key differences: unlike AdaBoost which
builds decision trees, gradient boosting builds trees that typically have 8–32 leaves.
Gradient boosting views boosting problem as an optimization problem, where it uses a
loss function and tries to minimize errors. This is why it is called gradient boost, as it is
inspired by gradient descent. Finally, trees are used to predict the residuals of samples
(prediction minus reality). Gradient boosting starts by building a tree to try to fit the data,
and subsequent trees are built for the purpose of reducing residuals (errors). It does this
to areas where existing learners are underperforming, which is similar to AdaBoost.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 101

3.2.3. SVM and KNN models

The other two applied models are grouped together, although they are not as similar
as previous models, they use the same method to evaluate the relationship between
financial ratios and stock returns.

Support vector machine (SVM): SVM is a binary classification algorithm, SVM takes input
data and classifies them into two different classes. Given a set of training examples of two
given categories, SVM algorithm builds a SVM model to classify other examples into those
two categories. SVM builds a hyperplane to classify a dataset into two separate classes.
To do this, first SVM will construct a hyperplane or a set of hyperplanes in a multi-dimen-
sional or infinite-dimensional space, which can also be used for classification, regres-
sion, or other tasks. For the best classification, it is necessary that optimal hyperplane is
located as far away from the data points of all classes (margins) as possible, because the
larger the margins, the smaller the generalization errors of the classification algorithm.

K-nearest neighbors (K-NN): K-nearest neighbors algorithm (K-NN) is commonly

used in data mining. K-NN is a method to classify objects based on the closest distance
between query points and all objects in training data. An object is classified based on
its K neighbors. K is a positive integer determined before the execution of the algo-
rithm. Euclidean distance is often used to calculate the distance between objects. This
simple algorithm is capable of solving the regression problems proposed by Altman
(1992). It assumes that similar observations exist in close proximity. K-NN predicts
output by using number k of training data.

3.3. Feature importance

The technique of assigning scores to input features in a predictive model is called

feature importance. Feature importance scores are an essential part of a predictive
model because they can be used to enhance a model’s performance and provide insi-
ght into the model and dataset. Provided relative importance scores can be used
to determine which features are most relevant for a study. There are many types of
feature importance scores in these techniques, those that are simple to calculate are
statistical correlation coefficients such as Pearson’s correlation and Spearman’s rank,
for linear and nonlinear correlation respectively.

There are three types of importance scores for more advanced features that are
also implemented from model coefficients as part of the linear model, from decision
model and permutation importance, which are described in detail by Pedregosa et al.
(2011). We describe three important features as follows:

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

102 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

Coefficients as feature importance: After fitting a linear machine learning model on

dataset, the coefficient of each input variable can be retrieved and stated as a feature
importance score. Comparison is possible because the dataset is normalized and the
variables have the same scale. This approach is applied to linear regression and elastic
net model for importance scores retrieval.

Decision tree feature importance: Decision tree algorithms, like the CART algorithm
used in this study, are provided in scikit-learning implementations with feature impor-
tance scores, based on criterion reduction used to select split points. This approach
is applied to decision tree model and all the tree-based combination methods such as
random forest, gradient boosting, and AdaBoost.

Permutation feature importance: This technique computes relative importance

scores independently to the model used. After fitting a model on a dataset, a predic-
tion is made, then this process is repeated five times for each feature in dataset, result-
ing in an average importance score for each input feature. This technique is suitable
for models that do not provide original feature importance scores, such as k-Nearest
Neighbors and SVM in this study. Important characteristics are identified as follows:

Fn(fused) = w1 F1n + w 2 F2n + ... + wm Fmn

1
RMSEi
Wi = , for i = 1,...n
1
∑ j RMSE
n

F: absolute value of the relative importance score of the feature generated by

the model
W: normalized weighted value based on the model’s predictability
m: number of models
n: number of financial ratios (n = 11 in this study)

3.4. Model evaluation

Evaluating a regression model is perhaps the most important part of building a super-
vised machine learning model. This determines whether a model is ready for deploy-
ment. Indicators only show numbers, but if examined carefully, they can tell us about
feature selection and feature engineering. In this study, we will look at the regression
metrics and discover why we use these following regression metrics: mean absolute
error, mean squared error and root mean squared error.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 103

Mean absolute error (MAE) is a measure of the mean absolute error between
predicted and actual values.

1 n
MAE = ∑ yi - yiµ
n i=1

Basically, MAE is a L1 norm. The smaller MAE, the smaller the distance between
predicted and actual values, and the better the model. However, MAE value does not
cover unit difference.

The mean squared error (MSE) of an estimator measures the average squares of
errors – that is, the average squared difference between predicted values and actual
values. MSE is a risk function, corresponding to the expected value of squared error
loss or quadratic loss. MSE is the second moment (about the origin) of errors.

1 n
( )
2
µ
MAE =
n
∑ i=1
yi - yi

Root mean squared error (RMSE) is the standard deviation of residuals (prediction
errors). Residuals are a measure of how far from the regression line data points are. RMSE
is a measure of how spread-out residuals are. In other words, it tells you how concentrated
data is around the line of best fit. RMSE is commonly used in forecasting and regression
analysis to verify experimental results and is also a measure of how effective models are,
by measuring the difference between predicted and actual values. The smaller RMSE, the
smaller the errors, the higher the level of reliability estimation of models.

1 n
( )
2
µ
RMSE =
n
∑ i=1
yi - yi

3.5. Research data

This study uses data collected from Vietnam Stock Exchange over the period of
2009-2020. Data is collected from audited financial statements of listed firms after
excluding firms in banking, securities and insurance sectors. After determining the
indicators, there are 3397 observations to be analyzed and forecasted, presented in
table 1 by year and by industry.

Based on figure 1, the stock return of firms witnessed huge fluctuations over the
period of 2009-2010, especially in 2010, stock return decreased to -9%, while during
2011-2013, it increased to 81% in 2013. Then from 2014 to 2017, the return was stable
at around 20%. In 2019, the stock return dropped to -15%, but by 2020, it increased
dramatically to 66%.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

104 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

Table 1. Descriptive statistics among the variables in the model

Table A: Data by year Table B: Data by industry
Number of Number of
Year Percentage Industry Percentage
observations observations
2009 139 4.1% Real estate & 1,194 35.1%
construction
2010 213 6.3% Technology 102 3.0%
2011 301 8.9% Industrial 400 11.8%
2012 314 9.2% Service 371 10.9%
2013 334 9.8% Consumer goods 309 9.1%
2014 323 9.5% Energy 221 6.5%
2015 341 10.0% Agriculture 293 8.6%
2016 349 10.3% Materials 378 11.1%
2017 350 10.3% Medical 129 3.8%
2018 330 9.7%
2019 274 8.1%
2020 129 3.8%
Total 3397 100.0% Total 3397 100.0%

Figure 1. Stock returns volatility by year

140%

120% 120%

100%

80% 81%
66%
60%

40%
32%
28% 27%
20% 22% 21%

0% 4% 2%
2009 -9%2011 2013 2015 2017 2019-15% 2021
-20%

-40%

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 105

Figure 2 shows observed stock returns by industry in the period of 2009-2020, in which
the industry with the highest stock return is the health industry, followed by the consumer
goods industry. The industries with low stock return are technology and industrial.

Figure 2. Stock returns by industry

40.0% 35.4%
34.3%
35.0% 32.0%
28.5% 28.2%
30.0%
24.5% 24.0% 24.3% 24.7%
25.0%
20.0%
15.0%
10.0%
5.0%
0.0%

Statistical data in table 2 shows that the average stock return is 27.0%, the standard
deviation is 54.4%, the lowest return is -59.9% and the highest is 244.9%. The account-
ing information is reflected through 11 financial ratios in terms of mean, standard devi-
ation, minimum value and maximum value.

Table 2. Descriptive statistics among the variables in the model

Std.
Variable Obs. Mean Min Max
Dev.
SR 3,397 0.270 0.544 -0.599 2.449
A1-Debt-to-equity ratio 3,397 1.485 1.287 0.074 7.022
A2-Earnings before tax-to-equity ratio 3,397 0.123 0.084 -0.108 0.398
A3-Total assets turnover 3,397 1.195 0.774 0.087 4.617
A4-Fixed assets turnover 3,397 17.222 29.542 0.303 273.554
A5-Operating income growth 3,397 1.120 0.908 -2.520 8.892
A6-Stock earnings volatility 3,397 0.029 0.893 -3.316 7.920
A7-Cash holding ratio 3,397 0.094 0.090 0.001 0.467
A8-Dividend yield 3,397 0.095 0.107 0.000 0.589
A9-Earnings management 3,397 0.577 0.815 0.004 5.843
A10-Accrual quality 3,397 -0.005 0.126 -0.236 0.657
A11-Earnings per share 3,397 2.275 1.941 -1.408 9.850

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

106 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

4. Result and discussion

Figure 3 shows the result of the correlation coefficient between variables. The purpose
of testing the correlation between independent and dependent variables is to elimi-
nate factors that can lead to multicollinearity, before running the regression model.
There are no two variables that have correlation coefficient greater than 0.6, except for
that of A2 - Earnings before tax-to-equity ratio and A11 - Earnings per share.

Figure 3. Autocorrelation matrix between variables in research model

To evaluate the effectiveness of algorithms in predicting the impact of accounting

information on financial statements on stock returns, we use three following metrics:
mean absolute error (MAE), mean squared error (MSE) and root mean squared error
(RMSE), presented in table 3. The algorithm with the smallest value has the highest
measurement performance.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 107

Table 3. Algorithm’s evaluation result

Mean absolute Mean squared Root mean
error error squared error
Linear regression 0.372 0.236 0.485
Lasso 0.375 0.238 0.488
Ridge 0.372 0.236 0.485
Elastic net 0.373 0.237 0.487
Random forest 0.370 0.230 0.480
Decision tree 0.487 0.425 0.652
AdaBoost 0.498 0.336 0.580
Gradient boosting 0.350 0.208 0.456
K-neighbor 0.368 0.239 0.489
SVM 0.339 0.219 0.468

With MAE measurement, AdaBoost algorithm has the highest value of 0.498, which
proves that this algorithm has the lowest measurement performance, while SVM algo-
rithm has a value of 0.339, which proves that it has the best measuremance perfor-
mance among these ten algorithms in applied dataset. Meanwhile, with MSE and RMSE
measurement, the algorithm with the lowest value, 0.208 and 0.456 respectively, is
gradient boosting algorithm. Thus, gradient boosting algorithm has the best predic-
tion performance among the algorithms used in predicting the impact of accounting
information from financial indicators on stock returns.

When applying regression algorithms, we choose to use linear regression, Lasso,

ridge, elastic net algorithms to determine importance coefficient of each financial
ratio, presented in table 4. Based on RMSE, we determine F (Fused) for the importance
regression coefficient for each ratio, shown in table 4 and figure 4.

Among 11 financial indicators of accounting information on financial statements,

earnings before tax-to-equity ratio has the highest importance score of 0.093, the next
one is operating income growth with a score of 0.063 and the third one which scores
0.053 is dividend yield. in contrast, the three indicators with the lowest importance
scores are debt-to-equity ratio (0.003), accrual quality (0.010) and total assets turn-
over (0.011).

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

108 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

Table 4. Coefficients as feature importance

Linear regression Lasso Ridge Elastic net F (Fused)
Debt-to-equity ratio 0.0063 0.0000 0.0063 0.0000 0.003
Earnings before tax-to- 0.0921 0.0949 0.0921 0.0921 0.093
equity ratio
Total assets turnover 0.0154 0.0037 0.0154 0.0097 0.011
Fixed assets turnover 0.0222 0.0027 0.0222 0.0121 0.015
Operating income growth 0.0651 0.0599 0.0651 0.0617 0.063
Stock earnings volatility 0.0235 0.0188 0.0235 0.0218 0.022
Cash holding ratio 0.0432 0.0349 0.0432 0.0382 0.040
Dividend yield 0.0554 0.0499 0.0554 0.0529 0.053
Earnings management 0.0194 0.0099 0.0194 0.0139 0.016
Accrual quality 0.0162 0.0000 0.0162 0.0082 0.010
Earnings per share 0.0150 0.0061 0.0151 0.0115 0.012
Root mean squared error 0.485 0.488 0.485 0.487 1.945
Weight 0.250 0.249 0.250 0.250

Figure 4. Coefficients as feature importance

A2 0.093
A5 0.063
A8 0.053
A7 0.040
A6 0.022
A9 0.016
A4 0.015
A11 0.012
A3 0.011
A10 0.010
A1 0.003

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 109

Table 5. Feature importance

Random Decision Gradient
AdaBoost F (Fused)
forest tree boosting
Debt-to-equity ratio 0.075 0.068 0.042 0.033 0.054
Earnings before tax-to- 0.100 0.093 0.066 0.113 0.095
equity ratio
Total assets turnover 0.075 0.067 0.065 0.040 0.061
Fixed assets turnover 0.064 0.064 0.042 0.015 0.045
Operating income growth 0.160 0.171 0.209 0.244 0.198
Stock earnings volatility 0.102 0.083 0.108 0.096 0.098
Cash holding ratio 0.074 0.077 0.053 0.054 0.064
Dividend yield 0.094 0.084 0.035 0.105 0.081
Earnings management 0.076 0.070 0.022 0.025 0.048
Accrual quality 0.113 0.139 0.305 0.246 0.201
Earnings per share 0.066 0.084 0.054 0.031 0.056
Root mean squared error 0.480 0.652 0.580 0.456 2.168
Weight 0.277 0.204 0.229 0.291

Next, we used decision tree algorithm group including random forest, decision tree,
AdaBoost, gradient boosting. Feature importance values of financial ratios are shown
in table 5 and their permutation feature importance values are presented in figure 5.

Figure 5. Feature importance

A10 0.201
A5 0.198
A6 0.098
A2- 0.095
A8 0.081
A7 0.064
A3 0.061
A11 0.056
A1 0.054
A9 0.048
A4- 0.045

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

110 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

As shown in figure 5, the three financial indicators that have the highest feature
importance score are accrual quality, operating income growth and stock earnings
volatility, with a value of 0.201, 0.198 and 0.098, respectively. The three financial ratios
with the lowest importance score are fixed assets turnover (0.045), earnings manage-
ment (0.0048) and debt-to-equity ratio (0.054).

The feature importance of support vector machine model and K-nearest neighbors
model are calculated as permutation feature importance. The initial values provided by
this technique are shown in table 6. The result is the average weighted importance score
for each financial indicator as determined in column F (Fused). The value of permutation
feature importance is much smaller than feature importance and coefficients as feature
importance presented in the previous sections, therefore they cannot be directly compared
with each other. However, importance feature rankings from highest to lowest can still
be presented as the average weighted importance score for each financial indicator.

Table 6. Permutation feature importance

K-Neighbor SVM F (Fused)
A1-Debt-to-equity ratio 0.0063 0.0039 0.0051
A2-Earnings before tax-to-equity ratio 0.0061 0.0072 0.0066
A3-Total assets turnover 0.0065 0.0053 0.0059
A4-Fixed assets turnover 0.0035 0.0035 0.0035
A5-Operating income growth 0.0051 0.0080 0.0065
A6-Stock earnings volatility 0.0056 0.0061 0.0059
A7-Cash holding ratio 0.0089 0.0063 0.0076
A8-Dividend yield 0.0057 0.0075 0.0067
A9-Earnings management 0.0040 0.0043 0.0042
A10-Accrual quality 0.0043 0.0048 0.0045
A11-Earnings per share 0.0046 0.0056 0.0051
Root mean squared error 0.4889 0.4682 0.9571
Weight 0.4892 0.5108

According to figure 6, the three most important financial indicators of accounting

information on the financial statements according to the permutation feature impor-
tance, are cash holding (0.008), dividend yield (0.007) and earnings before tax-to-eq-
uity ratio (0.007). In contrast, the financial indicators of the lowest importance are
fixed asset turnover (0.004), earnings management (0.004) and accrual quality (0.005).

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 111

Figure 6. Permutation feature importance

0.008 0.008

0.007 0.007 0.007 0.007

0.006 0.006
0.006
0.005 0.005
0.005 0.005
0.004
0.004 0.004

0.003

0.002

0.001

0.000
A4 A9 A10 A11 A1 A6 A3 A5 A2 A8 A7

To evaluate the importance of financial indicators of accounting information

on financial statements in a comprehensive way, we rank the results from three
approaches, from 1 to 11, to examine the overall importance of all financial ratios.
The results are based on the importance of financial ratios in the analysis above and
are visualized as a bar graph below. The highest importance score takes the value
11 and the lowest one takes the value 1, which means the higher the bar, the higher
the importance of the financial ratios. Each financial ratio is grouped with all three
approaches: “Coefficients as feature importance”, “Feature importance” and “Permu-
tation feature importance” for a better approach-based comparison. Figure 7 shows
five financial indicators of accounting information on financial statements that can be
considered important: Operating income growth, dividend yield, cash holding ratio,
earnings before tax-to-equity ratio, and stock earnings volatility. One indicator that
can be listed in this group is accrual quality.

Earnings before tax-to-equity ratio is an important indicator of stock returns.

Besides, earnings before tax-to-equity ratio coefficient is positive in all four linear
models, which is consistent with the study of (Anwaar, 2016). Previous research has
found out that net profit margin is an important indicator for both firm performance
and stock return (Delen et al., 2013; Anwaar, 2016). The importance of operating income
growth is consistent with previous research in both linear and decision tree models.
Based on positive coefficient values, the results show that stock returns increase as
Earnings before tax-to-equity ratio increases, which is consistent with the research
results of Öztürk & Karabulut (2018). They found that net profit margin has a positive

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

112 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

effect on stock returns, and higher net profit margin generates higher profits for the
next period. Previous research has also found that dividend yield can predict stock
returns (Fama & French, 2021; Lewellen, 2004). Similarly, the research results of Musal-
lam (2018) show that dividend yield is a highly important financial ratio, with positive
coefficient values, they show a positive association with stock returns.

Figure 7. Feature importance ranking

Coefficients as Feature Importance Feature importance Permuta�on Feature Importance

12

10

0
A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11

Research results of Musallam (2018), and Emamgholipour et al. (2013) indicate that
earnings per share has a positive and significant association with stock returns. This
result is consistent with these following research: Easton and Harris (1991), Ball and
Brown (1968), Basu (1983), and Hung et al. (2018). Stock earnings volatility also has
an important impact on stock returns, this result is consistent with that of Freeman
(1987), Collins and Kothari (1989), and Easton and Harris (1991). Besides, consistent with
a research by Hung and Van (2020), accrual quality also has a significant impact on
stock returns.

5. Conclusion and recommendation

To get experimental research results for the evaluation and measurement of the
importance of accounting information on financial statements as financial indica-
tors, on stock returns, we use panel data with 3397 observations from listed firms on
Vietnam Stock Market from 2009-2020 and machine learning algorithms. The research
results show that operating income growth, stock earnings volatility, dividend yield,
earnings before tax-to-equity ratio ratio, cash holding ratio and accrual quality have
an important and positive impact on stock returns. Based on the results, we propose
some recommendations as follows:

• When deciding to invest in stocks, investors need to pay attention to accoun-

ting information because this type of information affects stock prices, such as
operating income growth and stock earnings volatility on the income statement.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 113

However, there are other factors that can affect stock returns and announced
accounting information has not yet strongly affected stock return volatility
on Vietnam Stock Market. Before making investment decisions, investors can
consider earnings before tax-to-equity ratio, dividend yield because these
indicators have an important and positive impact on stock returns. Investors
should also refer to other additional information to make the best decisions.
Empirical results of other indicators like cash holding ratio, on the relationship
between earnings and stock returns, have suggested the difference between
firms with the same earnings rate.

• Firms need to disclose complete and timely financial statements. According

to the recommendations for investors above, investors can rely on accoun-
ting information to make investment decisions, so a firm itself needs to ensure
the quality of disclosed financial statements to create and maintain investors’
confidence in the firm. This not only creates a premise for investors’ confidence
in accounting and financial information that firms announce, but also helps
firms gain a position and improve their firm value to attract potential inves-
tors. complete and timely disclosure of financial statements, audit reports, and
board of directors’ reports will create belief in investors about the transparency
in information disclosure, which is a good signal to attract more investors.

Contribución de autores

Hung DN: Conceptualización, Metodología, Validación, Análisis formal, Investigación,

Recursos, Redacción borrador original, Validación, Administración del proyecto.
Van VT: Conceptualización, Metodología, Software, Validación, Análisis formal, Investi-
gación, Curación de datos, Redacción borrador original, Supervisión, Validación. Le DT:
Conceptualización, Metodología, Software, Análisis formal, Investigación, Curación
de datos.

Declaración de conflicto de intereses

El (los) autor(es) declara(n) que, durante el proceso de investigación, no ha existido

ningún tipo de interés personal, profesional o económico que haya podido influen-
ciar el juicio y/o accionar de los investigadores al momento de elaborar y publicar el
presente artículo.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

114 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

References
Altman, N. S. (1992). An introduction to kernel and nearest-neighbor nonparametric
regression. The American Statistician, 46(3), 175-185. [Link]
031305.1992.10475879
Anwaar, M. (2016). Impact of firms performance on stock returns (Evidence from listed
companies of ftse-100 index London, UK). Global Journal of Management and
Business Research, 16(1), 1-10.
Ball, R., & Brown, P. (1968). An empirical evaluation of accounting income numbers.
Journal of Accounting Research, 6(2), 159-178. [Link]
Barnes, P. (1987). The analysis and use of financial ratios. Journal of Business Finance
dan Accounting, 14(4), 449-461. [Link]
Basu, S. (1983). The relationship between earnings’ yield, market value and return for
NYSE common stocks: Further evidence. Journal of financial economics, 12(1),
129-156. [Link]
Belson, W. A. (1959). Matching and prediction on the principle of biological classifica-
tion. Journal of the Royal Statistical Society: Series C (Applied Statistics), 8(2),
65-75. [Link]
Collins, D. W., & Kothari, S. (1989). An analysis of intertemporal and cross-sectional
determinants of earnings response coefficients. Journal of Accounting and
Economics, 11(2-3), 143-181. [Link]
Dechow, P. M., & Dichev, I. D. (2002). The quality of accruals and earnings: The role
of accrual estimation errors. The Accounting Review, 77, 35-59. [Link]
org/10.2308/accr.2002.77.s-1.35
Delen, D., Kuzey, C., & Uyar, A. (2013). Measuring firm performance using financial ratios:
A decision tree approach. Expert Systems with Applications, 40(10), 3970-3983.
[Link]
Easton, P. D., & Harris, T. S. (1991). Earnings as an explanatory variable for returns. Jour-
nal of Accounting Research, 29(1), 19-36. [Link]
Emamgholipour, M., Pouraghajan, A., Tabari, N. A. Y., Haghparast, M., & Shirsavar, A.
A. A. (2013). The effects of performance evaluation market ratios on the stock
return: Evidence from the Tehran stock exchange. International Research Journal
of Applied and Basic Sciences, 4(3), 696-703. [Link]
Fama, E. F., & French, K. R. (2021). Dividend yields and expected stock returns. Univer-
sity of Chicago Press.

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 115

Freeman, R. N. (1987). The association between accounting earnings and security

returns for large and small firms. Journal of Accounting and Economics, 9(2),
195-228. [Link]
Ha, H. T. V., Hung, D. N., & Dung, T. M. (2018). Impact of accounting data on stock prices:
the case of Vietnam. International Journal of Accounting and Financial Repor-
ting, 8(1), 140-154. [Link]
Hai, T. T. T., Diem, N. N., & Binh, H. Q. (2015). The relationship between accounting infor-
mation reported in financial statements and stock returns-empirical evidence
from Vietnam. International Journal of Accounting and Financial Reporting, 5(1),
229-238. [Link]
Henrique, B. M., Sobreiro, V. A., & Kimura, H. (2019). Literature review: Machine learning
techniques applied to financial market prediction. Expert Systems with Applica-
tions, 124, 226-251. [Link]
Hung, D. N., Ha, H. T. V., & Binh, Đ. T. (2018). Impact of accounting information on finan-
cial statements to the stock price of the energy enterprises listed on Vietnam’s
Stock Market. International Journal of Energy Economics and Policy, 8(2), 1-6.
Hung, D. N., & Van, V. T. T. (2020). Studying the impacts of earnings quality on stock
return: Experiments in Vietnam. International Journal of Advanced and Applied
Sciences, 7, 45-53. [Link]
Kim, S. Y., & Upneja, A. (2014). Predicting restaurant financial distress using decision
tree and AdaBoosted decision tree models. Economic Modelling, 36, 354-362.
[Link]
Lewellen, J. (2004). Predicting returns with financial ratios. Journal of Financial Econo-
mics, 74(2), 209-235. [Link]
Metsomäki, J. (2020). Relation between financial ratios and stock returns: Machine lear-
ning approach [Master thesis, Lappeenranta-Lahti University of Technology],
[Link]
Musallam, S. R. (2018). Exploring the relationship between financial ratios and market
stock returns. Eurasian Journal of Business and Economics, 11(21), 101-116.
[Link]
Öztürk, H., & Karabulut, T. A. (2018). The relationship between earnings-to-price,
current ratio, profit margin and return: an empirical analysis on Istanbul
stock exchange. Accounting and Finance Research, 7(1), 109-115. [Link]
org/10.5430/afr.v7n1p109

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

116 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Dubourg, V.
(2011). Scikit-learn: Machine learning in Python. The Journal of Machine Learning
Research, 12, 2825-2830.
Quinlan, J. R. (1986). Induction of decision trees. Machine Learning, 1(1), 81-106. https://
[Link]/10.1007/BF00116251
Quinlan, J. R. (1996). Bagging, boosting, and C4. 5. En American Association for Artificial
Intelligence (Ed.), AAAI’96: Proceedings of the thirteenth national conference on
artificial intelligence. Vol 1 (pp. 725-730). AAAI Press.
Sun, J., Jia, M.-y., & Li, H. (2011). AdaBoost ensemble for financial distress predic-
tion: An empirical comparison with data from Chinese listed companies.
Expert Systems with Applications, 38(8), 9305-9312. [Link]
eswa.2011.01.042
Zięba, M., Tomczak, S. K., & Tomczak, J. M. (2016). Ensemble boosted trees with synthe-
tic features generation in application to bankruptcy prediction. Expert Systems
with Applications, 58, 93-101. [Link]
Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2),
301-320. [Link]

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

 Accounting information and stock returns in Vietnam securities market 117

Appendix 1. Attributes in research model

Attribute Notation Measurement
Debt-to-equity ratio A1 Total debts / Equity
Earnings before tax- A2 Earnings before tax / Equity
to-equity ratio
Total assets turnover A3 Net income / Total assets
Fixed assets turnover A4 Net income / Total fixed assets
Operating income A5 Operating incomet / Operating incomet-1
growth
Stock earnings vola- A6 Earnings before taxt - Earnings before taxt-1 / Ave-
tility rage number of outstanding shares
Cash holding ratio A7 Cash / Total assets
Dividend yield A8 Annual dividends per share / Price per share
Earnings management A9 There are many ways to measure/ manage ear-
nings, the authors use the model of Jones (1991).
Variable earnings management (EM) is measured
through a proxy, which is the residual of equation
(1). Earnings quality (EQ) is inversely proportional
to the residual of the following equation:
ACCit = β0 + β1 (REVit − ARit) + β2PPEit + εit (1)
In which:
DREVit is the difference between revenue of firm i
in year t and year t-1
PPEit is the cost of fixed assets of firm i in year t
ARit – 1 is total assets of year t-1
α1, α2, α3, are the parameters of each firm
Earnings management will be taken as a residual
of εit because it is an earnings-adjusting behavior,
so whether the adjustment is increased or decrea-
sed (equivalent to the adjusted accruals variable
with negative values or positive) are all beha-
viors. Thus, εit is the measurement of EM variable
(Earnings quality), the higher the deviation εit, the
lower earnings quality. Earnings quality is measu-
red by Earnings management: EQ_EM=EM*(-1)

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X

118 Ngoc Hung Dang, Thi Thuy Van Vu, Thi Nhat Le Dao

Attribute Notation Measurement

Accrual quality A10 To estimate the quality of accruals, the authors
used a model developed by Dechow and Dichev
(2002), in which Accrual quality is presented as
current Working capital accruals and it is calcu-
lated by Operating cash flows of the previous
period, this period, and the following period, all
divided by Total assets at the beginning of the
period.
To measure EQ, which is an AQ (Accrual quality)
variable and is considered as the negative stan-
dard deviation of residual εi,t of equation (2) after
regression.
CACCit = β0 + β1CFOit-1 + β2CFOit + β3CFOit+1 + εit (2)
In which:
CFOit-1, CFOit, CFOit+1 are respectively the cash flows
of year t-1, year t and year t+1, all divided by total
assets at the beginning of the period (Ait-1).
A higher value of AQ indicates poorer Accrual qua-
lity because less variation in current accruals is
explained by cash flows performance. Since ear-
nings are the sum of accruals and cash flows, and
the cash flow component is generally considered
to be objective and unmanipulated, the quality of
earnings depends on the quality of accruals.
Earnings per share A11 Net income / Average number of outstanding
shares

Fecha de recepción: 21/09/2021

Fecha de aceptación: 21/02/2022
Correspondencia: dangngochung@[Link]

Contabilidad y Negocios 17 (33), 2022 / e-ISSN 2221-724X