Innovative Applications of Artifici=
al
Neural Networks in Tax Forecasting
Aplicações inovadora=
s de
redes neurais artificiais na previsão fiscal
|
Bruno Couto de Abreu Rodolfo https://orcid.org/0009-0009-9=
071-3410 |
Master in Information Systems from =
Universidade Eduardo Mondlane (UEM), FC-Maputo,
Mozambique. cecybruna@gmail.com |
|
Bruno Miguel Ferreira Gonçalves h=
ttps://orcid.org/0000-0002-7541-3673 |
Resear=
ch
Centre in Basic Education (CIEB), Polytechnic Institute of Bragança, Port=
ugal |
ABSTRACT
T=
he
importance of forecasting tax revenues is vital for economic planning and
financial sustainability in Mozambique. This study addresses this topic by
exploring the potential of Artificial Neural Networks (ANNs) to improve such
forecasts. The central problem is the limitation of conventional methods in
capturing the complexity of tax data. The rationale for adopting ANNs lies =
in
their superior modeling and forecasting capacit=
y in
large and complex data environments. The results obtained demonstrate that =
ANNs
can forecast tax revenues with greater accuracy, outperforming traditional
models. The conclusion points to ANNs as a valuable tool for tax authoritie=
s,
increasing collection efficiency and contributing to the country's fiscal
stability.
RESUMO
A
importância da previsão das receitas fiscais é vital para o planeamento eco=
nômico
e sustentabilidade financeira em Moçambique. Este estudo aborda este tópico
explorando o potencial das Redes Neurais Artificiais (RNAs) para melhorar t=
ais
previsões. O problema central é a limitação dos métodos convencionais em ca=
ptar
a complexidade dos dados fiscais. A razão para a adoção de RNAs reside na s=
ua
superior capacidade de modelação e previsão em ambientes de dados grandes e
complexos. Os resultados obtidos demonstram que as RNAs podem prever as
receitas fiscais com maior precisão, superando os modelos tradicionais. A
conclusão aponta para a RNA como uma ferramenta valiosa para as autoridades
fiscais, aumentando a eficiência na cobrança e contribuindo para a estabili=
dade
fiscal do país.
P=
alavras-chave: redes neuronais; previsão fiscal;
sustentabilidade.
Recebido em 25/07/2024. Aprovado em 28/02/2025. Avaliado pelo s=
istema
double blind peer review. Publi=
cado
conforme normas da APA.
https://doi.org/10.22279/navus.v16.1953
1 INTRODU=
CT=
ION
Responsi=
ble
tax management forms the foundation for a nation's economic growth and
stability (Bartoluzzio & Anjos, 2020). In
Mozambique, where sustainable development is a priority, accurate forecasti=
ng
of tax revenues plays a critical role. Tax authorities face the continuing
challenge of estimating revenue effectively, a task complicated by economic
volatility and a constantly evolving fiscal environment. In this scenario,
technology emerges as a potential ally, and Artificial Neural Networks (ANN=
s)
represent an innovative frontier in predictive analysis.
The abil=
ity to
accurately predict tax revenues is an essential component in the administra=
tion
of any modern economy (da Silva et al., 2020). For Mozambique, a country
seeking to strengthen its fiscal infrastructure and promote economic
development, the adoption of advanced technologies such as ANNs can be a
divider. The complexity of today’s economic systems requires tools that can
learn and adapt to dynamic patterns, an area where ANNs excel. By integrati=
ng
historical data with contemporary economic variables, ANNs offer an opportu=
nity
to anticipate fiscal trends with unprecedented accuracy.
Furtherm=
ore,
the topic of ANNs in the fiscal forecast is relevant for Mozambique due to =
its
emerging economy and the need to optimize the collection of revenues (Lima
& Bezerra, 2022). Accurate tax forecasting is not only a matter of
administrative efficiency but also one of social justice (da Silva et al., =
2020).
Reliable revenue forecasts enable the government to allocate resources more
effectively, ensuring the provision of essential public services and
facilitating investment in critical areas (Bartoluzzio=
& Anjos, 2020). Therefore, the use of ANNs to improve fiscal forecasts =
represents
an initiative with the potential for a profound and lasting impact on the
country’s economic and social well-being.
The ques=
tion
that arises is: how can ANNs be applied to improve the accuracy of tax
collection forecasts in Mozambique? This issue arises in a context where
traditional forecasting methods often fail to capture the complexity and
dynamics of tax data. The objective of this study is to investigate the
applicability of ANNs in analysing historical data and economic factors, to
improve the accuracy of future revenue forecasts.
The
justification for exploring this topic is compelling. Accurate forecasts are
vital for tax planning and resource allocation, directly impacting the
government’s ability to finance public services and invest in infrastructur=
e. Furthermore,
the reliability of tax estimates is crucial for maintaining the confidence =
of
both investors and citizens.
The impo=
rtance
of this study for Mozambique cannot be overstated. By adopting ANNs in fisc=
al
forecasting, the country can make a significant leap in its economic manage=
ment
capabilities. This approach not only strengthens tax collection but also
promotes more transparent and efficient governance, which is essential for
socio-economic progress and the realization of a long-term vision for natio=
nal
development.
2 DISCUS=
SION
OF TOPICS
2.1 Neur=
al
Networks Modelling and Performance
ANN mode=
lling
is a key component in forecasting tax revenues, especially in complex econo=
mic
contexts such as Mozambique. Developing an ANN model involves carefully
selecting the network architecture, including the number of hidden layers a=
nd
neurons, which must be adequate to capture data complexity without leading =
to
"overfitting" (Almeida et al., 2020). The initial configuration of
weights and "biases", along with the selection of appropriate activation functions, are critical decisions that
impact the network's learning efficiency (Araújo 2022).
ANN trai=
ning
is an iterative process in which weights are adjusted to minimize prediction
error (dos Santos Neto, et al., 2020). This adjustment is achieved through
optimization algorithms, such as the descending gradient, which modify the
weights in response to the observed error between the network's forecasts a=
nd
actual values. During training, it is essential for the model not only to l=
earn
the patterns in the training data but also to generalize well to unseen dat=
a, a
quality verified through a validation set (Nascentes,
2020).
In the
literature, we can also cite the work of Simon Haykin<=
/span>
(1999), especially his book "Neural Networks: A Comprehensive Foundati=
on",
which thoroughly addresses the modelling and performance of artificial neur=
al
networks. Haykin explores the mathematical and
theoretical underpinnings of neural networks, providing essential insights =
into
how these networks can be modelled and optimized for improved performance. =
The
performance of ANNs is assessed based on their accuracy and reliability in
predictions (Announcement, 2023). Factors such as the quantity and quality =
of
training data, the complexity of the model, and the adequacy of optimization
techniques play significant roles in the effectiveness of predictions. In
Mozambique, where data can be scarce or noisy, the robustness of the ANN mo=
del
is important. Therefore, the modelling and performance of ANNs must be
approached with a deep understanding of the specifics of tax data and the <=
/span>country's economic
environment (Almeida et al., 2020).
The choi=
ce of
training data is another factor that significantly influences the performan=
ce
of ANNs. In Mozambique, selecting representative data is challenging due to
economic variability and the limited availability of tax data. The quality =
of
the data, including its completeness and accuracy, determines the ANN's abi=
lity
to learn relevant patterns and make reliable predictions (Araújo, 2022).
Therefore, data collection and pre-processing are fundamental steps that
precede network training.
Cross-va=
lidation
is a common technique used to evaluate the generalization of the ANN model.=
It
involves splitting the data set into several parts, training the model on s=
ome
of these parts, and validating it on the others. This method helps identify=
whether
the model is overfitted to the training data and whether it can accurately
predict previously unseen data (dos Santos Neto et al., 2020).
Furtherm=
ore,
the ability of an ANN to handle unstructured and noisy data is particularly
valuable in Mozambique, where data collection systems may be underdeveloped.
ANNs can extract useful information from imperfect data, which offers a
significant advantage over more traditional methods that require highly
structured and clean data (Almeida et al., 2020).
Finally, the interpretability of ANN models is a crucial factor, especially
when applied to tax decisions that impact the population and economy of
Mozambique. Highly complex ANN models can be difficult to interpret, raising
concerns about the transparency and accountability of predictions. Therefor=
e,
it is important to consider not only the model's performance in terms of
accuracy but also its ability to be explained and justified to stakeholders
(Araújo, 2022).
2.2
Applications of ANNs in Financial Forecasts
ANNs have
emerged as a powerful tool for predicting financial indicators, including s=
tock
market trends (de Oliveira & dos Santos, 2020). In Mozambique, where the
economy faces complex challenges and financial data is highly volatile, ANNs
offer an innovative approach to improving the accuracy of fiscal forecasts.=
Dornelles et al. (2022) highlight several application=
s,
including:
Stock Pr=
ice
Forecast: ANNs
have been successfully applied to predict stock prices, assess risks, and
support investment decisions. A classic example is the study by Kimoto et a=
l.
(1990), in which Artificial Neural Networks were used to predict stock pric=
es
on the Tokyo Stock Exchange. These networks analyze
historical price time series and incorporate information such as trading
volume, past trends, and technical indicators to generate future forecasts.=
Financial Time Series Models: ANNs can be adapted to model financial time seri=
es,
including exchange rates, interest rates, inflation, and commodity prices. =
The
work of Zhang et al. (1998) demonstrated the effectiveness of ANNs in modeling financial time series, such as predicting ex=
change
rates. ANNs learn from past patterns and capture nonlinear relationships
between variables. In Mozambique, where economic data may be scarce and noi=
sy,
ANNs offer an advantage in handling the complexity of these time series.
Overcoming Limitations of Traditional Models: ANNs overcome the limitatio=
ns
of traditional models, such as linear regression, which often fail to captu=
re
the nonlinear characteristics of financial data. Hutchinson et al. (1994)
compared ANNs with traditional regression models in predicting option prices
and demonstrated the superiority of ANNs. In the context of Mozambique’s
growing economy and volatile markets, ANNs represent a valuable alternative=
for
predicting fiscal trends with greater accuracy.
Challenges and
Opportunities
A renowned study that can be cited here is
Rosenblatt's foundational work (1958). Although the author focused on the
development of the Perceptron, a precursor to modern neural networks, his
research opened the door to diverse applications, including financial
forecasting. The evolution of neural networks for complex tasks such as fin=
ancial
forecasting stems from Rosenblatt’s original concept that machines could le=
arn
and make predictions based on data (Rosenblatt, 1958).
De Oliveira and dos Santos (2020) report that the
application of ANNs in financial forecasting is a field that is gaining
increasing prominence, especially in emerging markets such as Mozambique. T=
he
ability of ANNs to process and learn from large volumes of data makes them
suitable for analysing financial markets, which are characterized by their
complexity and uncertainty.
Forecasting economic indicators, such as gross
domestic product (GDP) and inflation, is another area where ANNs can be
extremely useful. By incorporating a wide range of economic and social
variables, ANNs can help identify trends and patterns that may not be
immediately apparent to human analysts or through traditional statistical
methods (Rodella, 2023).
In addition, ANNs can be applied to forecast tax
revenues, a crucial aspect of economic governance. In Mozambique, where tax
planning and resource allocation present significant challenges, ANNs can
provide more accurate and reliable forecasts, enabling policymakers to make
more informed decisions (Dornellas et al., 2022).
However, it is important to note that while ANNs o=
ffer
many advantages, they also present challenges. The quality of input data is
crucial to the success of forecasts. In countries such as Mozambique, where
there may be restrictions on data collection and processing, it is essentia=
l to
ensure that data is of high quality and representative of the economic real=
ity
(Rodella, 2023).
2.3 ANNs=
in
Decision Support Systems
ANNs are transforming decision-making processes wi=
thin
organizations. By providing data-based predictions, ANNs enable managers and
administrators to make more informed and strategic choices (Schuch, 2021). =
The
accuracy of these predictions is crucial, as decisions based on inaccurate
information can lead to unwanted results.
In decision support systems, ANNs analyse large
volumes of data to identify patterns and trends that may not be obvious to
humans. This is useful in complex and dynamic environments, where the amoun=
t of
data can be overwhelming. ANNs help filter the noise and focus on the most =
relevant
information (de Souza, et al., 2022).
The integration of ANNs into decision support syst=
ems
is also beneficial in terms of efficiency (Bastos et al., 2019). By automat=
ing
data analysis, ANNs reduce the need for manual analysis, which can be
time-consuming and error-prone. This enables decision-makers to focus on
interpreting results and planning actions rather than on data processing (F=
igueiredo,
2022).
A relevant study to cite here is the work of McCul=
loch
and Pitts (1943), whose formal neuron model marked a milestone in understan=
ding
how the brain could be mathematically represented and simulated by machines.
This concept is fundamental to neural network-based decision support system=
s,
where networks process complex information to aid in decision-making (McCul=
loch
& Pitts, 1943).
However, implementing ANNs in decision support sys=
tems
presents certain challenges. One of the keys is to ensure that ANN models a=
re
transparent and explainable (Schuch, 2021). Decision-makers need to underst=
and
how predictions are made to trust them. Therefore, the interpretation of ANN
models remains an active and critical area of research.
Another challenge is the need for high-quality dat=
a,
as ANNs are only as effective as the data on which they are trained (de Sou=
za,
et al., 2022). In scenarios where the data is incomplete, <=
/span>inaccurate, or
difficult, the predictions generated by ANNs can be questionable. Thus, data
collection and pre-processing are vital steps in developing effective decis=
ion
support systems (Figueiredo, 2022).
Furthermore, ANNs must be adapted to the specific
context in which they are applied. This means that the models need to be
customized to reflect the "nuances" of the decision-making
environment. For example, in Mozambique, ANNs used in tax decision support
systems must account for local economic factors and tax collection patterns=
(de
Souza et al., 2022).
Collaboration between data experts and decision-ma=
kers
is essential for the successful implementation of ANNs in decision support
systems. Data experts can build and adjust ANN models, while decision-makers
provide data about the needs and objectives of the organization. This
collaboration ensures that ANNs are used effectively and aligned with
organizational strategies (Schuch, 2021).
In this way, Fernandes (2020), describes steps that
demonstrate the application of ANNs in decision support systems, such as:
· =
Data collection:
Gathering data relevant to the decision problem;
· =
Data Processing:
Cleaning and preparing of data for analysis;
· =
Definition of the=
ANN
Model: Choosing the neural network architecture appropriate to the problem;=
· =
ANN Training: Usi=
ng
the data to train the neural network model;
· =
ANN Validation: T=
esting
the model with a separate dataset to ensure that it generalizes well for new
data;
· =
Results
Interpretation: Analysing the ANN outputs to understand the predictions or
classifications made;
· =
Integration with =
the Decision
System: Implementing the ANN as part of the decision support system;
· = Decision making: using the information provided by the ANN to make informed decisions, and;<= o:p>
·&nb=
sp;
Evaluation and Adjustment: Monitoring the
performance of the decisions taken and adjusting the ANN
In conclusion, ANNs play a significant role in
decision support systems. They hold the potential for more accurate predict=
ions
and more efficient decision-making processes. However, to integrate them ef=
fectively,
challenges such as model interpretability, data quality, and context-specif=
ic
customization must be addressed. Once these challenges are overcome, ANNs c=
an
become an invaluable tool for decision-makers in Mozambique and beyond.
2.4 Succ=
ess
cases with ANNs in Tax Forecast
The use of ANNs in fiscal forecasting has proven t=
o be
a promising approach in numerous case studies worldwide. For example, a stu=
dy
conducted in the state of Rio de Janeiro, Brazil, employed ANNs to predict =
the
collection of the Tax on Circulation of Goods and Provision of Services (IC=
MS),
one of the country’s main taxes. The ANN model chosen was the Long Short-Te=
rm
Memory (LSTM), which is suitable for time series due to its ability to reme=
mber
information over extended periods (Figueiredo, 2022).
Again, the work of Rosenblatt (1958) is relevant h=
ere,
as his Perceptron was one of the first models to demonstrate the ability of
machines to perform predictive tasks. Applications of neural networks in
specific areas, such as fiscal forecasting, are offshoots of the supervised
learning concept he introduced.
Another case study in Rio Grande do Sul, Brazil,
developed a short-term univariate model using LSTM ANNs to predict monthly =
ICMS
revenue. The model showed a cumulative forecasting error of -2.33% in six
one-step forecasts, demonstrating significant gains compared to other
predictive methods previously used by the State Treasury Secretariat (Dornelles et al., 2022).
For the evaluation model, Dor=
nelles
et al. (2022) used the Mean Squared Error (MSE), as explained be
In the studies by Oliveira and dos Santos (2020)
entitled “Strategies to Combat Tax Evasion: A Model for Artificial Neural
Networks-Based ICMS”, the authors discuss the development of an ANN model a=
imed
at improving the prediction and detection of tax evasion related to the ICM=
S.
This type of model could be useful for tax authorities, enabling them to
identify evasion patterns and optimize audit and tax collection strategies.=
The
application of ANNs in this context suggests an innovative and data-driven
approach to addressing a significant tax problem, potentially leading to
greater efficiency and fairness within the tax system.
Bastos et. al (2019), in their study entitled “Fin=
ancial
Validation of Neural Network Training Algorithms for Financial Series Trend
Prediction”, explore the application of ANNs for predicting trends in finan=
cial
data. The authors investigated the effectiveness of different ANN training
algorithms in financial time series, focusing on the validation and accurac=
y of
predictions. Although the exact content of the study is not directly
accessible, the research appears to be relevant for enhancing financial for=
ecasting
techniques and optimizing decision-making in economic contexts.
The implementation of ANNs in tax projections also
requires a careful analysis of available data and a clear understanding of =
tax
objectives. The quality of input data, the selection of relevant variables,=
and
the definition of appropriate parameters are critical factors for the model=
's
success (Rodella, 2023).
In addition, case studies highlight the importance=
of
validating and testing ANN models. Using techniques such as cross-validation
and comparison with "benchmarks", it is possible to assess the
robustness and reliability of predictions generated by ANNs.
In short, case studies with =
ANNs in
tax prediction provide valuable data on how these models can be adapted and
optimized for different tax contexts. They also highlight the potential of =
ANNs
to improve the accuracy of tax predictions, which is essential for effective
financial planning and management.
Artificial Neural Networks (ANNs) have been
increasingly used in tax prediction, demonstrating their ability to model
complex nonlinear relationships that traditional methods may not capture
efficiently. Linear regression, for example, is limited by its assumption of
linearity between variables (Bartoluzzio & =
Anjos,
2020). While time series and econometric models are useful, they may not ad=
apt
well to volatile or unstable data patterns (Peixoto et al., 2016).
In contrast, ANNs, with their flexible structure a=
nd
learning ability, can identify hidden patterns in data, which is particular=
ly
valuable in tax prediction, where anomalies and unforeseen events are commo=
n.
Studies such as Souza’s (2011), which compared ANNs to traditional methods =
for
predicting the BOVESPA index, demonstrate that ANNs can surpass traditional
techniques in terms of accuracy.
In addition, ANNs can process an abundance of input
variables without the need for pre-selection or transformation, unlike
econometric models that often require variables to be carefully chosen and
transformed (Araújo, 2020). This allows ANNs to capture complex interactions
between variables that can be neglected in traditional methods.
Haykin (1998) addressed=
the
comparison between neural networks and traditional methods of analysis and
prediction in his work. He discussed the advantages of ANNs, particularly t=
heir
flexibility and ability to handle nonlinear data, comparing them to more tr=
aditional
techniques such as linear statistical models.
However, it is important to note that ANNs also ha=
ve
disadvantages. They can be opaque, making it difficult to interpret the res=
ults
and understand how inputs influence forecasts. Additionally, ANNs require l=
arge
datasets for training and may be prone to overfitting, particularly if not
properly regularized (Almeida et al., 2020).
Empirical evidence suggests that ANNs offer
significant improvements in terms of accuracy and reliability in tax
predictions. For example, a dissertation from the Federal University of Rio
Grande do Norte found that ANNs provided more accurate predictions of the
BOVESPA index compared to time series methods (Souza, 2011). Another study =
by
the Federal University of Itajubá highlighted the applicability of ANNs in =
forecasting
economic indicators, surpassing traditional statistical models (Freiman, 20=
04).
However, the choice between ANNs and traditional
methods should not be made in isolation. The decision must consider the
specific context of the forecast, the availability of data, the need for
interpretability, and the modeler’s experience. In some cases, a combinatio=
n of
methods may provide the most robust approach, taking advantage of the stren=
gths
of each method.
In short, while ANNs present clear advantages in t=
erms
of flexibility and modeling ability, they also
require careful application to avoid pitfalls such as overfitting. The
literature indicates that, when properly applied, ANNs can indeed offer
substantial improvements over traditional methods. However, a thorou=
gh
assessment of each specific situation is essential.
2.6 The Connection Between Neural Networks a=
nd
Artificial Intelligence: The Role of Deep Learning
Artificial Neural Networks (ANNs) are a central
technique within the field of Artificial Intelligence (AI), inspired by the
structure and functioning of the human brain. Used to solve complex problems
involving large volumes of data and nonlinear patterns, ANNs have been
essential in the advancement of AI (IBM, n.d.).
The emergence of Deep Learning, a specific subfiel=
d of
AI, has broadened the capabilities of ANNs by introducing deep networks with
multiple hidden layers. These networks can extract high-level features from=
the
data, leading to significant advances in areas such as image recognition,
natural language processing, and financial forecasting (Goodfellow et al., =
2016).
ANNs began to gain prominence in the 1980s with the
introduction of the backpropagation algorithm, which enabled the efficient
training of multilayer networks (Rumelhart et al.,1986). However, it was on=
ly
with the advent of deep learning, driven by advancements in computational p=
ower
and the availability of large volumes of data, that ANNs truly flourished
(LeCun et al., 2015).
Deep learning differs from traditional machine
learning approaches by utilizing deep neural networks, which consist of
multiple layers of artificial neurons. These additional layers enable the
network to learn more complex representations of the data, which is crucial=
for
tasks such as speech recognition and computer vision (=
Krizhevsky
et al., 2012).
One of the most notable examples of the impact of =
deep
learning is in image recognition. Convolutional neural networks (CNNs), a
specific type of ANN, have proven to be highly effective at identifying obj=
ects
in images with great accuracy. This has applications in a variety of areas,
from medical diagnostics to autonomous vehicles (He et al., 2016).
In the field of natural language processing (NLP),
Deep Learning has also demonstrated impressive results. Models such as GPT-=
3,
developed by OpenAI, are capable of generating human-like text in a coherent
and contextually relevant manner (Brown et al., 2020). These advances have
significant implications for virtual assistants, machine translation, and s=
entiment
analysis (Devlin et al., 2018).
In the financial sector, ANNs and Deep Learning ar=
e employed
to predict market movements and detect fraud. The ability of these networks=
to analyze large volumes of historical data and identify
subtle patterns makes them valuable tools.
Despite the advances, deep learning faces signific=
ant
challenges. Training deep networks requires large amounts of data and
computational power, which can be an obstacle for many organizations. In
addition, the interpretability of deep learning models remains an active ar=
ea
of research, as understanding how these networks make decisions is crucial =
for
their application in sensitive areas (Doshi-Velez & Kim, 2017).
The future of ANNs and deep learning appears promi=
sing,
with continued advances in hardware, algorithms, and training techniques. T=
he
integration of these technologies with other emerging areas, such as quantum
computing and explainable AI, may open new frontiers in the field of AI (
Therefore, it can be stated by the group of authors
mentioned above that Artificial Neural Networks and Deep Learning have play=
ed a
fundamental role in the advancement of Artificial Intelligence. With
applications ranging from image recognition to financial forecasting, these
technologies are transforming various industries. However, challenges such =
as
the need for large volumes of data and the interpretability of models still
need to be overcome for their full potential to be realized.
2.7 Example and Visualization
Examining the application of ANNs in tax forecasti=
ng
is essential to understand how these models can be implemented in practice.=
A
practical example is the use of ANNs to predict the collection of ICMS in R=
io
Grande do Sul. In this case, a short-term univariate model utilizing Long
Short-Term Memory (LSTM) networks was employed, resulting in a cumulative
forecasting error of −2.33% in six forecasts (Do=
rnelles
et al., 2022).
The exemplification and visualization of neural
networks can be related to the work of McCulloch and Pitts (1943) in the
formalization of brain processes. They created mathematical representations
that facilitated the visualization and understanding of how neural networks
work, which is fundamental to the modern visualization of these networks (M=
cCulloch
& Pitts, 1943).
Another example is the application of ANNs in the
forecast of the quotation of beef straw, demonstrating the versatility of A=
NNs
in different tax contexts (Freiman, 2004). In addition, LSTM ANN models have
been employed for tax predictions, showing superior performance compared to
traditional methods (Figueiredo, 2022).
To illustrate the process, we can consider the
following simplified pseudocode of an ANN for tax prediction:
Python
# Simplified
Pseudocode of an ANN for Tax Forecast
Import necessary libraries
Define network parameters (number of layers, neurons, etc.)
Load historical tax data.
Prepare the data (normalization, division into training and test sets)
Create the ANN architecture.
Train ANN with training data.
Evaluate ANN performance with test data.
Use trained ANN to make future predictions.
This pseudocode
represents a basic structure that can be adapted and expanded based on spec=
ific
tax forecasting needs. The results can be visualized using graphs that show=
ANN
predictions compared to actual data, allowing a visual analysis of the mode=
l's
performance.
It is important to emphasize that, in practice, the
implementation of ANNs requires careful data analysis, selection of appropr=
iate
hyperparameters, and rigorous validation to ensure reliable and accurate
predictions. Furthermore, the interpretation of ANN models can present
challenges, often requiring the application of additional techniques to
understand how inputs influence the outputs of the model.
In general, regardless of the framework applied, <=
span
class=3DSpellE>Dornelles (2022) outlines practical applications of A=
NNs in
fiscal forecasting, including the following:
Python
#
Pseudocode for predicting ICMS collection using ANN
# Import required libraries
import numpy as np
import pandas as pd
from keras.models import Sequential
from keras.layers import Dense, LSTM Ffrom sklearn.preprocessing
import MinMaxScaler
# Load historical data of revenue of ICMS
data_icms =3D pd.read_csv<=
/span>('recover_icMS.csv')
# Data Preprocessing
scaler =3D MinMaxScaler(fe=
ature_range=3D(0,
1))
standardized_data =3D scal=
er.fit_transform(data_icms)
# Split the data into training and test sets
training_size =3D int(len<=
/span>(standard_data) * 0.67)
training, test =3D data_normalized[0:size_train=
ing,:],
data_normalized[size-training:le=
n(data_normalized),:]
# Convert arrays to matrices that RNA can interpret
def create_dataset(dataset, look_back=3D1):
dataX, dataY =3D []=
, []
for i in range(len(=
dataset)-look_back-1):
a =3D dataset[i:(i+look_ba=
ck),
0]
dataX.append(a)
dataY.append(dataset[i + look_back, 0])
Return np. array(dataX), n=
p.array(dataY)
look_back =3D 1
trainingX, trainingY =3D create_dataset(training, look_ba=
ck)
testX, testY =3D
# Reshape for [samples, time steps, features]
trinoX =3D np.reshape(trinoX, (treinoX.shape[0]=
, 1, treinoX.shape[1]))
testX =3D np.reshape(testX, (testX.shape[0], 1=
, testX. shape[1]))
# Create and train
ANN
model =3D Sequential()
model.add(LSTM(4, input_sh=
ape=3D(1,
look_back)))
model.add(Dense(1))
model.compile(loss=3D'mean=
_squared_error',
optimizer=3D'adam')
model.fit(treinoX, =
treinoY, epochs=3D100, batch_siz=
e=3D1,
verbose=3D2)
# Make predictions
forecast_training =3D mode=
l.predict(treinoX)
forecast_test =3D model.pr=
edict(testeX)
# Invert predictions to the original scale
forecast_training =3D scal=
er.inverse_transform(forecast-training)
forecast_test =3D scaler.i=
nverse_transform(forecasts_test)
trainingY =3D scaler.inver=
se_transform([treinoY])
testY =3D scaler.inverse_t=
ransform([testeY])
# Calculate forecast error
error_train =3D np.sqrt(np.mean((previsions_training - trainingY) ** 2))
error_test =3D np.sqrt(np.mean((previsions_test =
- testY) ** 2))
print(f'Training error: {training_error:.2f}') =
print(f'Error in test: {error_test:.2f}')
This pseudocode is a simplified representation and
should not be directly applied in a real production environment. It is inte=
nded
to illustrate the process of creating an ANN model for tax prediction, from=
the
data loading to the model evaluation. In practice, fine adjustment of
parameters, cross-validation, and other techniques would be required to ens=
ure
the robustness and accuracy of the model.
For real examples of the implementation of ANNs in
fiscal forecasting, one can reference studies such as “Applied Neural Netwo=
rks
in ICMS Revenue Forecast in Rio Grande do Sul”, which details the applicati=
on
of LSTM ANNs to predict the monthly ICMS revenue (Dorn=
elles,
2022), or “Models for Tax Forecasting Using LTSM Neural networks” which
compares multivariate and univariate approaches of LTSRs in tax forecasts (=
Figueiredo,
2022). These studies provide valuable insights into the practical applicati=
on
of ANNs for tax prediction.
3 METHOD=
OLOGY
In this section, we will describe the methodology
adopted to investigate how ANNs can improve the accuracy of tax collection
forecasts in Mozambique.
This study addresses this topic by exploring the
potential of Artificial Neural Networks (ANNs) to improve such forecasts. T=
he
central problem is the limitation of conventional methods in capturing the
complexity of tax data.
Data Source and Selection Criteria:
· =
Historical Tax Data: We colle=
cted
historical tax collection data in Mozambique; including tax revenues, fees,=
and
contributions;
· =
Economic indicators: We
incorporate relevant economic indicators, such as GDP growth, inflation,
unemployment and investment;
· =
Source Selection: We used
official sources, such as government reports, economic databases, and acade=
mic
publications, over up to 5 years, and;
· =
Selection criteria: We
prioritized reliable, up-to-date, and representative data, within academic
bases of relevance, as well as data linked to the Mozambique Revenue Author=
ity.
Data
Analysis Techniques:
· =
Artificial Neural Networks: A=
NN
models, such as feedforward networks or recurrent networks, will be exempli=
fied
to predict future tax revenues;
· =
Training and Validation: We
divided data into training and validation sets, hyperparameters were adjust=
ed,
and evaluate the performance of the model;
· =
Result Analysis: We will eval=
uate
the accuracy of predictions by comparing them with traditional methods, and=
;
·&nb=
sp;
Qualitative Analysis: It will explore examples of
application in the Tax Revenue Authority, highlighting practical models and
emphasizing how these models can assist the country’s tax units.
Legal
and Ethical Considerations:
· =
Data Privacy: We will ensure =
the
anonymization of tax data and compliance with privacy regulations;
· =
Informed Consent: We will obt=
ain
consent for the data from the authorities of the tax authority of Mozambiqu=
e;
· =
Transparency and Integrity: We
will report all methodological steps with transparency and potential biases
will be avoided,
· =
Social Responsibility: We will
consider the social and economic impact of our findings.
This methodology aims to provide a solid foundation
for our research and contribute to advancing the forecast of tax collection=
in
Mozambique.
4 RESULT=
S AND
DISCUSSION
For this section, we can highlight the following
items:
Main Discoveries
· =
ANNs have demonstrated a supe=
rior
ability to model nonlinear complexities in fiscal data when compared with
traditional methods such as linear regression and time series;
· =
The application of ANNs to tax
data has resulted in more accurate and reliable predictions, as evidenced by
empirical studies and practical examples.
Theoretical and Practical Implications
· =
Theoretically, the results
reinforce the importance of exploring advanced computational models in fiel=
ds
traditionally dominated by statistical methods;
· =
In practice, the
implementation of ANNs can help tax authorities improve their forecasts,
resulting in better planning and resource allocation.
Limitations
of the Study
· =
A significant limitation was =
the
inability to access data from the Mozambique Revenue Authority, which could
have enriched the analysis with local information.
· =
ANNs require large volumes of
quality historical data and can be complex to set up and train properly.
Suggestions for Future Research
· =
Future research could explore=
the
integration of ANNs with other "machine learning" models to create
hybrid tax prediction systems,
· =
It would be valuable to carry=
out
studies that overcome barriers to access to data in different geographical
contexts, including Mozambique.
Artificial Neural Networks (ANNs) have emerged as a
promising alternative to improving the accuracy of tax predictions. Their ability to model complex nonlinear
relationships within tax data is well-documented. Empirical studies, such as
Souza's (2011) comparison of ANNs with traditional methods for predicting t=
he
BOVESPA index, have consistently demonstrated that ANNs can outperform line=
ar
and econometric models. However, this advantage is not universal and depend=
s on
the specific context and quality of the data.
The comparison between ANNs and traditional methods
reveals important distinctions. While ANNs offer significant advantages, su=
ch
as the ability to handle nonlinear data and flexibility for modeling
complex relationships, traditional methods should not be dismissed. Linear
regression, for instance, remains valuable in scenarios where interpretabil=
ity
is crucial. Furthermore, econometric models have a strong foundation in
economic theory and can provide relevant insights.
A central limitation is the need for large volumes=
of
data to adequately train ANNs. In addition, ANNs can be opaque, making it
challenging to interpret their results. The lack of access to data from the
Mozambique Tax Unit also posed a significant limitation in this study. The
absence of local data may impact the applicability of ANNs in specific cont=
exts.
In practice, implementing ANNs requires technical
expertise and substantial computational resources. Tax authorities should
consider investing in team training and necessary infrastructure to support=
the
adoption of these advanced models. Collaboration between machine learning
experts and economists is essential to maximize the benefits of ANNs.
Thus, this study reinforces the relevance of ANNs =
in
fiscal forecasting while highlighting the need for hybrid approaches. An intelligent combination of ANNs with
traditional methods can be a key to obtaining more robust and reliable
predictions. Moreover, searching f=
or
local data and validation in different geographical contexts are promising
areas for future research.
This critical and constructive analysis aims to
provide a comprehensive view of the implications and challenges associated =
with
the application of ANNs in fiscal forecasting.
It is essential to acknowledge both the advantages and limitations of
these models, promoting a balanced and informed approach.
3 CONCLU=
SION
A study on tax revenue forecasting using Artificial
Neural Networks (ANNs) and a comparison with traditional methods revealed
valuable insights and challenges inherent in tax forecasting. Tax revenue
forecasting is a fundamental pillar of a country’s economic planning, and A=
NNs
emerge as a promising tool, potentially outperforming traditional methods. =
The
study evaluated the effectiveness of ANNs in improving the accuracy of tax
forecasts. The results demonstrated that ANNs perform remarkably well, adap=
ting
to the complexities and volatility of tax data. Their flexible and
comprehensive approach allows them to capture non-linear relationships that
traditional methods often fail to model. Compared to linear regression and
econometric models, ANNs showed significant improvements in accuracy and
reliability. However, traditional methods are still relevant, especially in
scenarios that require interpretability and simplicity.
A limitation of the study was the inaccessibility =
of
data from the Mozambique Tax Unit, which could have enriched the analysis.
Furthermore, ANNs require large volumes of high-quality data for training a=
nd
can be complex in terms of setup and interpretation. From a practical
perspective, the adoption of ANNs in tax forecasting can optimize resource
allocation and support more informed tax policies, but it requires that tax
authorities are prepared to implement and manage these advanced models. The
study reinforces the potential of ANNs in tax forecasting while highlighting
the importance of hybrid approaches that combine their advantages with thos=
e of
traditional methods. Collaboration between machine learning experts and
economists is essential to boost research in this area.
For future research, it is recommended to explore
hybrid models and overcome barriers to data access. Additional studies could
focus on diverse geographic contexts, including those with data constraints,
such as Mozambique, to validate and expand the current findings. This paper
reflects on the advances and challenges faced in the research, adopting a
critical perspective on the methodologies and results, and suggesting
directions for future studies in tax forecasting with ANNs.
REFERENCES
Almeida, A., Amaris, M., Merlin, B., & Veras, A. (2020). Modelage=
m e
prediçao temporal de parâmetros de qualidade de água usando redes neurais
profundas. Anais do XI Workshop de
Computação Aplicada à Gestão do Meio Ambiente e Recursos Naturais (pp.
121-130). SBC. Available in: https://sol.sbc.org.br/index=
.php/wcama/article/view/11026 Accessed at: 24 April =
2024.
Anunciação, G. A. (2023). Aná= lise comparativa de estruturas de redes neurais artificiais para modelagem basea= da em dados do bombeio centrífugo submerso. [Trabalho de Conclusão de Curs= o, Universidade Federal da Bahia]. Available in: https://repositorio.ufba.br/= handle/ri/38957 Accessed at: 24 April 2024.<= o:p>
Araújo, J. L. G. D. (2022). Modelagem de reforma catalítica seca de
metano a gás de síntese, utilizando machine learning e redes neurais
(Bachelor's thesis, Federal University of Pernambuco). Available in: https://repositorio.ufpe.br/=
handle/123456789/47706 Accessed at: 24 de
Bartoluzzio, A. I. S. D. S., & Anjos, L. C. M. D. (2020). Ciclos políticos e gestão fiscal nos municípios brasileiros. Revista de Administração Contemporânea, 24, 167-180. Available = in: <= span lang=3DPT style=3D'font-family:"Myriad Pro",sans-serif;mso-bidi-font-family= :Arial; color:black;mso-themecolor:text1;mso-ansi-language:PT;mso-bidi-font-weight: bold;text-decoration:none;text-underline:none'>https://www.scielo.br/j/rac/= a/h5QvmkQ9JD8hNRm5mrPmpLr/?lang=3Dpt Accessed at: 24 April 2024.<= o:p>
Bastos, M. V., Carrano, E. G., Batista, L. S., & Minas Gerais, M.= G. (2019) Validação financeira de algoritmos de treinamento de redes neurais p= ara predição de tendência em séries financeiras. Anais Simpósio Brasileiro de Automação Inteligente. Available i= n: https://scholar.archive.org/= work/p5m5vmxuubgetnfcvp3wmk6gti/access/wayback/https://proceedings.science/= proceedings/100113/_papers/111135/download/fulltext_file2 Accessed at: 24 April 2024.<= o:p>
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal,=
P.,
… & Amodei, D. (2020). Language models are few-shot learners. Advances in neural information process=
ing
systems, 33, 1877-1901Available in: https://splab.sdu.edu.cn/GPT3.pdf Accessed at: 26 August=
2024.
da Silva, M. C., de Souza, F. J. V., Martins, J. D. M., & de Barr=
os
Câmara, R. P. (2020). Fatores explicativos da gestão fiscal em municípios
brasileiros. Revista Contemporânea =
de
Contabilidade, 17(42), 26-37. Available in: https://periodicos.ufsc.br/i=
ndex.php/contabilidade/article/view/56116 Accessed at: 24 April =
2024.
de Oliveira, F. N., & dos Santos, L. P. G. (2020). Estratégias pa=
ra
combater a sonegação fiscal: Um modelo para o icms baseado em redes neurais
artificiais. Revista de Gestão, Fin=
anças
e Contabilidade, 10(1), 42-64. Available in: https://www.researchgate.net=
/profile/Luis-Santos-8/publication/353889808_ESTRATEGIAS_PARA_COMBATER_A_SO=
NEGACAO_FISCAL_UM_MODELO_PARA_O_ICMS_BASEADO_EM_REDES_NEURAIS_ARTIFICIAIS/l=
inks/6116cd0a1ca20f6f861e55b0/ESTRATEGIAS-PARA-COMBATER-A-SONEGACAO-FISCAL-=
UM-MODELO-PARA-O-ICMS-BASEADO-EM-REDES-NEURAIS-ARTIFICIAIS.pdf
de Souza, C. C., Juniior, J. B. A. C., Cristaldo, M. F., Castelão, R.=
A.,
Frainer, D. M., da Gama Viganó, H. H., ... & de Souza Vieira, J. M. C.
(2022). Comparação dos modelos ARIMA, RNA e híbrido ARIMA-RNA para a previs=
ão
dos custos de internações hospitalares pelo Sistema Único de Saúde (SUS) na
região Centro-Oeste do Brasil. Research, Society and
Development, 11(16). Available in: https://rsdjournal.org/index=
.php/rsd/article/view/37547 Accessed at: 24 April =
2024.
Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of de=
ep
bidirectional transformers for language understanding. arXiv
preprint arXiv:1810.04805. Available in: https://arxiv.org/abs/1810.0=
4805?amp=3D1 Accessed at: 26 Au=
gust
2024.
Dornelles, G. Z., Schwartzer, F. R., & Braatz, J. (2022). =
Redes Neurais Aplicadas na
Previsão de Receita de ICMS no Rio Grande do Sul. Secretaria da Fazenda do
Estado do Rio Grande do Sul, Tesouro do Estado, Divisão de Estudos Econômic=
os e
Fiscais e Qualidade do Gasto. Available in: https://tesouro.fazenda.rs.g=
ov.br/upload/1643376127_Artigo%20Modelo%20Redes%20Neurais_Texto%20de%20Disc=
ussao.pdf Accessed at: 23 April =
2024.
Dos Santos Neto, L. A., Maniesi, V., Querino, C. A. S., da Silva, M. =
J.
G., & Brown, V. R. (2020). Modelagem hidroclimatologica utilizando redes
neurais multi layer perceptron em bacia hidrográfica no sudoeste da amazôni=
a. Revista Brasileira de Climatologia=
, 26. Available
in:
Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable machine learning. ar=
Xiv
preprint arXiv:1702.08608. Available in: https://arxiv.org/abs/1702.08608 Accessed at: 26 August 2024.
Fernandes, B. D. L. (2020). Business intelligence no suporte
à decisão estratégica [Doctoral
dissertation, Higher Institute of Engineering of Coimbra]. Available in: https://comum.rcaap.pt/handl=
e/10400.26/40355 Accessed at: 24 April =
2024.
Figueiredo, K. (2022). Modelos para Previsão Tributária Utilizan=
do
Redes Neurais LSTM. Anais do XIX En=
contro
Nacional de Inteligência Artificial e Computacional. Disponível em: https://www.academia.edu/98052542/Modelos_para_Previs%C3%A3o_Tribut%C=
3%A1ria_Utilizando_Redes_Neurais_LSTM Ac Accessed at: 23 de abri=
l de
2024.
Fischer, T., & Krauss, C. (2018). Deep learning with long short-t=
erm
memory networks for financial market predictions. European Journal of Operational Research, 270(2), 654-679. Available in: https://www.sciencedirect.co=
m/science/article/pii/S0377221717310652 Accessed at: 26 August 2024.
Freiman, J. P. (2004). Utilização de Redes Neurais Artificiais na
previsão de indicadores financeiros para avaliação econômica de negócios em
situação de risco. [Master's thesis, Federal University of Itajubá]. Available in: https://repositorio.unifei.e=
du.br/jspui/handle/123456789/3748 Accessed at: 23 April =
2024.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. <=
span
lang=3DEN-GB style=3D'font-family:"Myriad Pro",sans-serif;mso-bidi-font-fam=
ily:
Arial;color:black;mso-themecolor:text1;mso-ansi-language:EN-GB;mso-bidi-fon=
t-weight:
bold'>
Haykin, S. (1998). Neural networks<=
/i>: a
comprehensive foundation. Prentice Hall PTR.
Hornik, K., Stinchcombe, M., & White, H. (1989). Multilayer
feedforward networks are universal approximators. Neural Networks, 2(5), 359-366. Available in: https://www.sciencedirect.co=
m/science/article/pii/0893608089900208 Accessed at: 27 August 2024.
Hutchinson, J. M., Lo, A. W., & Poggio, T. (1994). A Nonparametric Approach to Pricing and Hedging Derivative Securities=
via
Learning Networks. The Journal of F=
inance,
49(3), 851-889. Available in: https://onlinelibrary.wiley.=
com/doi/abs/10.1111/j.1540-6261.1994.tb00081.x Accessed at: 27 August 2024.
IBM. (n.d.). O que é uma rede
neural? Available in: https://www.ibm.com/br-pt/to=
pics/neural-networks Accessed at: 26 August 2024.
Kimoto, T., Asakawa, K., Yoda, M., & Takeoka=
,
M. (1990, June). Stock market prediction system with modular neural network=
s. IJCNN international joint conference on
neural networks (pp. 1-6). IEEE. Available in:
https://ieeexplore.ieee.org/abstract/document/5726498/ Accessed at: 26 Augu=
st
2024.
Krizhevsky, A., Sutskever, I., & Hinton, G. E.
(2012). ImageNet classification with deep convolutional neural networks.&nb=
sp;Advances in neural information process=
ing
systems, 25. Available in: https://proceedings.neurips.=
cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html Accessed at: 26 August 2024.
LeCun, Y., Bengio, Y.,=
&
Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444. Available in: https://www.nature.com/artic=
les/nature14539 Accessed at: 26 August 2024.
Lima, R. P., & Bezerra, F. A. (2022). Gestão fiscal e a eficiênci=
a do
gasto público em educação e saúde nos estados brasileiros. Revista do Serviço Público, 73 (2), 359-378. Available in: https://repositorio.enap.gov=
.br/handle/1/7423 Accessed at: 24 April 2024.
McCulloch, W. S., & Pitts, W. (1990). A logical calculus of the i=
deas
immanent in nervous activity. =
Bulletin
of Mathematical Biology, 52, 99-115. Available in: https://link.springer.com/ar=
ticle/10.1007/bf02459570 Accessed at: 27 August 2024.
Nascentes, R. F. (2020). Modelagem da dinâmic=
a de
herbicidas no solo e palha utilizando redes neurais artificiais.
Peixoto, A. C. P., Neves, C., & Melo, E. =
F.
L. (2016). Comparação de Modelos Tradicionais para Previsão de Taxas de
Mortalidade. Revista Brasileira de =
Risco
e Seguro, 22(1), 1-25. Available in: https://www.rbrs.com.br/arqu=
ivos/rbrs_22_1.pdf Accessed at: 23 April 2024.
Rodella, V. G. (2023). Estudo de caso: aplica= ção de machine learning para a previsão de tendências das ações das bolsas de valores brasileira e norte americana. [Trabaho de Conclusão de Curso, Unive= rsidade Estadual Paulista]. Available in: = https://repositorio.unesp.br= /items/00964d9f-7217-4187-81a6-7cc35ff600db Accessed at: 24 April 2024.<= o:p>
Rosenblatt, F. (1958). The perceptron: a
probabilistic model for information storage and organization in the brain. =
Psychological review, 65(6), 386. =
Available in: h=
ttps://psycnet.apa.org/journals/rev/65/6/386/ Accessed at: 27 August=
2024.
Rumelhart, D. E., Hinton, G. E., &
Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323(6088), 533-536. Available in: https://www.nature.com/artic=
les/323533a0 Accessed at: 26 August=
2024.
Schuch, K. E. F. (2021). Análise preditiva com
redes neurais artificiais para o planejamento de sistemas de irrigação.
[Trabalho de Conslusão de Curso, Pontifícia Universidade Católica De Goiás]=
. Available in:
Souza, J. C. (2011). Previsão do Índice BOVES=
PA
por Meio de Redes Neurais Artificiais: Uma análise comparada aos métodos
tradicionais de séries de tempo. [Dissertação de mestrado, Universidade Fed=
eral
do Rio Grande do Norte]. Available in: https://repositorio.ufrn.br/=
bitstream/123456789/12197/1/PrevisaoIndiceBovespa_Souza_2011.pdf=
Accessed at: 23 Abril 2=
024.
Zhang, G. P., & Hu, M. Y. (1998). Neural network forecasting of the British Pound/US Dollar exchange ra=
te. OMEGA, 26(4), 495-506. Available i=
n:
https://www.sciencedirect.com/science/article/pii/S0305048398000036 Accessed
at: 27 August 2024.
Innovative Applications of Artificial Neural Networks in Tax Forecas=
ting
Bruno Couto de Ab=
reu
Rodolfo; Bruno Miguel Ferreira Gonçalves
|
IS=
SN
2237-4558 •<=
/span> Navus • <=
/span>Florianópolis
• SC • <=
/span>v.
16 • p. 01- |
|
|
ISSN 2237-4558 • Navus • &n=
bsp;Florianópolis •
SC • v.9
• n.2 • <=
/span>p.
XX-XX • abr./jun. 2019 |
|