Prediction of Daily Sales of Individual Products in a Medium-Sized

Brazilian Supermarket Using Recurrent Neural Networks Models

Jociano Perin

, Lucas Dias Hiera Sampaio

, Marlon Marcon

and Andr

e Roberto Ortoncelli

Postgraduate Program in Informatics, Federal University of Technology, Paran

a (UTFPR), Paran

a, Brazil

Keywords:

Sales Prediction, Supermarket Sales, Time Series Forecasting, Long Short-Term Memory, Gated Recurrent

Unit.

Abstract:

Accurately predicting daily sales of products in supermarkets is crucial for inventory management, demand

forecasting, and optimizing supply chain operations. Many studies focus on predicting the total sales of large

stores and supermarkets. This study focuses on forecasting daily sales of individual products across various

categories. In the experiments, we used Linear Regression and two types of Recurrent Neural Networks:

Long Short-Term Memory and Gated Recurrent Unit. One of the contributions of the work is the database

used, which is made available for public access and contains daily sales records (between January 2019 and

December 2024) of 250 products in a medium-sized supermarket in Brazil. The results show that the pre-

dictors’ performance varies signiﬁcantly from product to product. For one semester, the average of the best

25% resulted in a Root Mean Squared Error (RMSE) of 1.55 and a Mean Absolute Percentage Error (MAPE)

of 17.20, and for the average of all products, the best RMSE was 2.12, and the best MAPE was 43.94. We

observed similar performance variations for all analyzed semesters. With the results presented, it is possible

to understand the performance of the predictors in ten semesters.

1 INTRODUCTION

Organizations achieve success by adapting quickly to

changes in their business environment. Accurate and

timely sales forecasting is especially crucial for com-

panies operating in production, logistics, marketing,

trade, and retail (Meulstee and Pechenizkiy, 2008).

For retailers, sales forecast errors can lead to in-

correct stocking of products, reducing proﬁts. A man-

ager’s ability to predict sales patterns that determine

when to order and replenish stocks and plan for future

labor and sales is a signiﬁcant challenge to increasing

sales and proﬁts in a supermarket (Jeyarangani et al.,

2023)

Data generated from previous sales records are

valuable for predicting upcoming sales. These data

contain signiﬁcant patterns and information that can

be modeled using a Machine Learning (ML) algo-

rithm, which can accurately predict sales with high

precision. ML has become a signiﬁcant subﬁeld of

https://orcid.org/0009-0006-4936-0615

https://orcid.org/0000-0003-1644-3634

https://orcid.org/0000-0002-3698-8570

https://orcid.org/0000-0001-9622-8525

Data Science that has gained popularity because of

its superior predictive and forecasting abilities. An

ML model must be trained on data to identify pat-

terns from which it can accurately predict future sales

(Chen and Lu, 2017).

Recent work in the literature has explored tradi-

tional machine learning techniques (Almufadi et al.,

2023; Raizada and Saini, 2021) and also time series-

based methods (Jeyarangani et al., 2023; Huo, 2021)

for retail sales forecasting. These works have focused

on predicting total sales volume and presenting results

that could be suitable to aid retailers’ planning.

In contrast, in this study, we focused on forecast-

ing daily sales of individual products. We conducted

a case study using data from a medium-sized super-

market in Paran

a, Brazil

. Our results show that the

results vary signiﬁcantly between different products,

highlighting that for some products, we obtained re-

sults that seem adequate, while for most products, the

results are in worse ranges. For the 25% of prod-

ucts with the best prediction results, we obtained, on

average, in a semester, a Root Mean Squared Error

According to the Brazilian Supermarket Association,

supermarkets are medium-sized if they have between 5 and

19 checkouts.

Perin, J., Sampaio, L. D. H., Marcon, M., Ortoncelli and A. R.

Prediction of Daily Sales of Individual Products in a Medium-Sized Brazilian Supermarket Using Recurrent Neural Networks Models.

DOI: 10.5220/0013649800003967

In Proceedings of the 14th International Conference on Data Science, Technology and Applications (DATA 2025), pages 739-747

ISBN: 978-989-758-758-0; ISSN: 2184-285X

739

(RMSE) of up to 1.55 and a Mean Absolute Percent-

age Error (MAPE) of up to 17.20. Considering the

average for all products we obtained, the best RMSE

was 2.12, and the best MAPE was 43.94 (also for one

semester), with the results worsening considerably in

some experimental instances. We obtained forecast

results close to sales records for some of the products.

Still, the predicted values varied substantially for the

majority, indicating that new studies should be carried

out.

In the experiments, we used three different algo-

rithms to predict the supermarket’s sales volume for

the next day: Linear Regression and two types of Re-

current Neural Networks (RNNs): Long Short-Term

Memory (LSTM) and Gated Recurrent Unit (GRU).

For all training, we trained the predictor with 1 year

of daily sales records for each product and predicted

the sales of that product in the next 30 days. To

carry out the case study, we used daily sales records

(between January 2019 and December 2024) of 250

products from 8 categories (Non-Alcoholic Bever-

ages; Alcoholic Beverages; Cookies, Sweets, and

Snacks; Hygiene and Beauty; Dairy and Cold Cuts;

Cleaning and Household Goods; Basic Grocery; Oth-

ers). Researchers interested in accessing the experi-

mental dataset may contact the last author via email.

We hope the available database and insights based on

the results will motivate new research in the area.

The rest of this paper is structured as follows: Sec-

tion 2 presents the theoretical foundations essential

for understanding this work. Section 3 provides a

comprehensive literature review. Section 4 describes

the experimental database. Details of the conducted

experiment and an analysis of its results are in Section

5. Finally, Section 6 presents the concluding remarks

and suggestions for future research.

2 THEORETICAL ASPECTS

This Section presents fundamental theoretical aspects

for understanding this project. The concept of time

series is introduced in Section 2.1. Section 2.2 de-

scribes Recurrent Neural Networks, a type of Artiﬁ-

cial Neural Network (ANN) designed to make predic-

tions on time series, describing the two main types of

RNN: LSTM and GRU.

2.1 Time Series

A time series is a set of observations made sequen-

tially over time. This means that the data are col-

lected and organized based on speciﬁc time intervals,

providing a chronological view of the variations of

the variable being analyzed. Therefore, a time series

comprises variables indexed and ordered by speciﬁc

moments in time, denoted by t. Mathematically, a

time series can be represented as a structure X, where

each element X(t) corresponds to the data observed in

the time interval t. This concept allows the analysis of

patterns, trends, and cycles within the data over time,

facilitating an understanding of dynamic behaviors in

various ﬁelds, such as economics, meteorology, and

social sciences (Chatﬁeld, 2004).

A time series consists of four fundamental compo-

nents: (i) trend, which captures the long-term move-

ment or overall direction of the data; (ii) seasonality,

representing periodic ﬂuctuations that occur at reg-

ular intervals; (iii) cyclicity, which refers to recur-

rent patterns that do not follow a ﬁxed period; and

(iv) noise, which accounts for random variations and

unexplained deviations in the data (Brockwell and

Davis, 2016).

Time series analysis involves techniques to under-

stand the stationary or non-stationary nature of data

and autocorrelation, which is the correlation of a time

series with its past values. These analyses are funda-

mental to choosing the appropriate forecasting model

or understanding the data’s structure better (Hamilton,

1994).

Traditional time series models, such as ARIMA,

moving averages, and autoregressive models, effec-

tively capture trends and seasonal patterns. Re-

searchers and analysts widely use these models to an-

alyze economic and ﬁnancial variations due to their

reliability (Wei, 2006).

The advancement of artiﬁcial intelligence has led

to the increasing application of ML techniques, such

as RNNs, in time series analysis. These approaches

effectively model non-linear complexities and inter-

actions that traditional methods may fail to capture.

In this project, we predict values in a time se-

ries with three methods: linear regression (Su et al.,

2012), which is a statistical technique used to model

the relationship between a dependent variable and

one or more independent variables, and two types

of RNNs (LSTM and GRU), described in Subsection

2.2.

2.2 Recurrent Neural Network

Traditional ANNs use a few hidden layers (one or

two), but in Deep Learning (DL), the ANN uses

more neurons and hidden layers. DL allows compu-

tational models composed of multiple processing lay-

ers to learn data representations with various levels

of abstraction. DL-based methods have contributed

to drastically improving experimental results, being

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

740

state of the art for different problems (Aldhaheri et al.,

2024; Archana and Jeevaraj, 2024).

RNNs have a structure similar to that of a standard

ANN, with the distinction that connections between

hidden units are allowed, which allows the model to

retain information about the past, allowing it to dis-

cover temporal correlations between events that are

distant from each other in the data (Sherstinsky, 2020;

Pascanu et al., 2013).

An ANN structure comprises an input layer, one

or more hidden layers, and an output layer. RNNs

have an organization similar to a chain of repeated

modules, designed to function as memory units, stor-

ing crucial information from previous processing

phases. These networks include a feedback loop that

allows the output of step t − 1 to be fed back into the

network, inﬂuencing the result of step t and, subse-

quently, of each subsequent step (Le et al., 2019).

RNNs perform a backward approach, layer by

layer, from the ﬁnal output of the network, adjust-

ing the weights of each unit. The information loops

are repeated, which can result in signiﬁcant updates

to the weights of the neural network model, leading

to an unstable network due to the accumulation of er-

ror gradients during the update process. Therefore,

back-propagation over time is not efﬁcient enough to

learn a long-term dependence pattern due to gradient

vanishing and gradient explosion problems, which is

one of the crucial reasons that lead to difﬁculties in

training RNNs (Rumelhart et al., 1986; Hochreiter,

1998).

There are variations of RNNs that overcome this

difﬁculty. LSTM is an evolution introduced to solve

the training problems/challenges of RNNs by adding

additional interactions per module (or cell) (Hochre-

iter and Schmidhuber, 1997). LSTMs are a special

type of RNN, capable of learning long-term depen-

dencies and remembering information for extended

periods.

In addition to LSTMs, researchers use another

type of RNN to overcome the long-term learning

problem: the GRU, which is an optimized RNN based

on LSTM. The cellular structure of a GRU resem-

bles that of an LSTM, but the combines the input and

forget gates of the LSTM into a single update gate

(Santra and Lin, 2019; Chung et al., 2014). The up-

date gate controls how much information from the

previous state is retained in the current state. In con-

trast, the reset gate determines whether to combine

the current state with earlier information (Cho et al.,

2014).

3 LITERATURE REVIEW

Regarding quantitative methods, recent studies have

presented tools based on three main groups of tech-

niques: statistics, traditional machine learning, and

deep learning (Dai and Huang, 2021). ML-based ap-

proaches are usually more powerful and ﬂexible. DL

techniques such as LSTM and GRU have recently

shown competitive results in this application domain.

In (Almufadi et al., 2023), Linear Regression was

used to predict future sales of supermarket branches,

achieving an average absolute percentage error of

27.8%. The authors used a database of 896 records

with the following data: store ID, store area (size),

variety of items available, number of customers who

visited the store, and sales per day.

In the work of (Huo, 2021), the sales volume of

10 Walmart stores distributed in 3 states (California,

Texas, and Wisconsin) is predicted using a database

with 3049 products divided into three categories and

seven departments. The authors used different algo-

rithms (Triple Exponential Smoothing, ARIMA, Lin-

ear Regression, Randon Forest, XGBost, and LSTM)

to predict sales for a 28-day window, training the al-

gorithms with sales made from 2011-01-29 to 2016-

04-04 and using as the test set the sales records for

the period 2016-04-05 to 2016-05-22. Linear Regres-

sion yielded the best predictive performance across

different experimental scenarios among the evaluated

methods.

Another study on sales forecasting in Walmart

stores was conducted by (Raizada and Saini, 2021),

utilizing a dataset containing sales records from 45 re-

tail locations. The dataset included various features,

such as historical sales, promotional events, holiday

weeks, temperature, fuel prices, the consumer price

index, and the state’s unemployment rate. The authors

applied traditional machine learning algorithms to

predict sales trends, including Linear Regression, K-

nearest neighbors (K-NN), Support Vector Machine

(SVM), and Extra Trees Regression. Extra Trees Re-

gression achieved the highest predictive performance

among these models, with accuracy exceeding 98.2%

in the experiments.

Also noteworthy is the work of (De Almeida et al.,

2022), which carries out an empirical analysis of the

sales forecast of units of a supermarket chain in the

Brazilian Northeast, applying ML techniques (Linear

Regression, Random Forests, and XGBoost) on daily

transactional data from ﬁve years (2015 to 2019) col-

lected from eight different stores. On average, the

best results were obtained with XGBoost, but other

algorithms presented superior results for some stores.

It is worth noting that this study reported the impacts

Prediction of Daily Sales of Individual Products in a Medium-Sized Brazilian Supermarket Using Recurrent Neural Networks Models

741

of the COVID-19 pandemic and seasonal events that

directly impacted the results of the prediction algo-

rithms.

In the work of (Gupta et al., 2022), Machine

Learning algorithms (Linear Regression, Decision

Tree, Random Forest, Ridge Regression and XG-

Boost) were also used to predict product sales in

stores distributed in different cities. The authors used

a database with sales data from the year 2013 for

1,559 products in different towns/stores, with the fol-

lowing data: product ID, product weight, fat level

(low-fat or regular), percentage of the total display

area of all products, item category, item MRP, store

ID, date the store was established, storage area, city

type, an identiﬁer that shows whether the product is

sold in a grocery store or supermarket, and product

sales in the store. The test and training sets consisted

of 5,681 and 8,523 records. In the case study, the best

results were obtained with the XGBoost model, which

reached an accuracy of 87%.

It can also highlight the work of (Dai and Huang,

2021) and (Kohli et al., 2020), who explored a sales

database from a German drugstore chain, with infor-

mation that identiﬁes the store, the number of sales

and buyers in a day, variables that indicate whether

the store was closed or open on a given day and

whether the store’s sales were affected by the school

holiday period, the type of store and the level of as-

sortment, the distance to a competitor in meters, the

year in which the current store started, whether the

store was on promotion on a given day, whether the

promotion was taking place in several stores at the

same time, the period of participation in the promo-

tion and the interval between promotions. In (Kohli

et al., 2020), experiments were conducted with the

Linear Regression and KNN algorithms, obtaining a

mean absolute percentage error of up to 22.065. (Dai

and Huang, 2021), applied LSTM, obtaining better re-

sults in forecasting sales volume than those obtained

with different machine learning algorithms (used with

an Auto Machine Learning tool).

4 EXPERIMENTAL DATABASE

To construct the experimental database, we estab-

lished a cooperative agreement between a university, a

software company specializing in retail systems, and

a medium-sized supermarket in Paran

a, Brazil. The

supermarket, which uses the retail management soft-

ware provided by the software company, consented to

share its complete sales history for selected products

over ﬁve years (from January 1, 2019, to December

31, 2024). The software company was responsible

for extracting the relevant records from the database

and supplying the university with the necessary data

in .csv format.

The database includes products from 8 categories.

The following list describes these categories and

presents the number of products in each of them:

1. C

- Non-Alcoholic Beverages: (46 products) in-

cludes juices, soft drinks, teas, energy drinks, and

other non-alcoholic beverages;

2. C

- Alcoholic Beverages: (17 products) includes

beers, wines, spirits, and other alcoholic drinks;

3. C

- Cookies, Sweets, and Snacks: (38 prod-

ucts) includes cookies, chocolates, candies, sa-

vory snacks, and other snack options;

4. C

- Hygiene and Beauty: (38 products) includes

personal care items such as shampoos, soaps,

creams, and makeup;

5. C

- Dairy and Cold Cuts: (24 products) in-

cludes milk derivatives such as cheese and yogurt,

as well as cold meats like ham and salami;

6. C

- Cleaning and Household Goods: (21

products) includes cleaning supplies, detergents,

sponges, and home organization items;

7. C

- Basic Grocery: (50 products) includes es-

sential pantry items such as rice, beans, pasta,

ﬂour, oils, and canned goods;

8. C

- Others: (16 products) includes items that do

not ﬁt into the other categories;

Each product in our database has a .csv ﬁle named

with the product code followed by the product cate-

gory code (C

to C

). Each row in the database has

two columns, one showing a date and the other the

quantity of the product sold on that date. This infor-

mation represents a time series.

Figure 1 shows the quantity sold of two prod-

ucts in our database over 90 days. Supermarket cus-

tomers may exhibit similar consumption patterns ev-

ery 7 days due to characteristics of weekly shopping

cycles due to weekdays and weekends. This behavior

can be observed for the second product shown in the

Figure 1, especially in the ﬁrst few days.

Most of the .csv ﬁles in our database contain sales

records for every day in the period analyzed. When

the supermarket did not sell a product on a given date,

there is no line for that date in the .csv ﬁle for the

respective product.

5 CASE STUDY

This Section presents details about the case study

and its results. The Subsection 5.1 presents details

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

742

Figure 1: Daily sales of two products for 90 consecutive

days.

of the algorithms used and the case study methodol-

ogy. Subsection 5.2 describes the experimental met-

rics used. The results obtained and the analysis of

them are in Subsection 5.3.

5.1 Methodology

We ran the experiments with three ML algo-

rithms/models that were implemented as follows:

• Linear Regression: implemented with sliding

windows of size 7. Each window included the val-

ues of the product sales quantity in 7 consecutive

days, with the target being the prediction of the

value on the eighth day.

• GRU: The input layer of the GRU used is a 7-

day sliding window, followed by a GRU layer of

50 units, followed by a second GRU layer of 25

units, both employing the relu activation function.

The network concludes with an output layer com-

prising a single neuron.

• LSTM: With a similar architecture to the GRU

used, the LSTM has a 7-day sliding window as

the input layer, followed by an LSTM layer of

50 units, followed by a second LSTM layer of 25

units, both employing the relu activation function.

The network concludes with an output layer com-

prising a single neuron.

We trained and applied each model iteratively. For

training, we used the ﬁrst 365 days (Day 1 to Day

365) from each .csv ﬁle, using a 7-day time window

to predict sales on the eighth day. After training, the

model predicted sales for the next 30 days (Day 366

to Day 396). We then shifted the training window for-

ward by 30 days and retrained the model to forecast

the following 30 days. We repeated this process until

we covered all recorded days in each ﬁle. We trained

separate models for each product.

5.2 Evaluation Metrics

To evaluate the experimental results we use the fol-

lowing metrics: Root Mean Squared Error (RMSE)

and Mean Absolute Percentage Error (MAPE).

The RMSE measures the root mean square error.

The RMSE is deﬁned in Equation 1, which has the

following terms:

• y

is the actual observed value,

• ˆy

is the value predicted by the model,

• n is the total number of observations.

RMSE =

∑

i=1

− ˆy

)

(1)

RMSE penalizes more signiﬁcant errors due to

squaring, making it useful when large errors must be

minimized.

MAPE, in turn, calculates the mean absolute per-

centage error between the actual values and the pre-

dicted values. MAPE is deﬁned in Equation 2, with

the same terms deﬁned for Equation 1.

MAPE =

100

∑

i=1



− ˆy



(2)

MAPE provides the average percentage of error

relative to the true values and helps interpret the rela-

tive error of the model. However, it can be sensitive to

values close to zero, making it less reliable in speciﬁc

contexts. In our experiments, we removed all days the

product had zero sales before calculating the MAPE.

5.3 Results and Discussion

Tables 1 and 2 present the average RMSE and MAPE

values applied to all products in the database. Each

row represents one of the algorithms used, and the

columns refer to each semester of the experiment’s

execution period. We present the average RMSE for

each semester. The database covers the period from

2019 to 2024. Of note, 2019 is not included in the

table since it is the ﬁrst year of the period and was

used exclusively for training the models in the ﬁrst

interactive training cycle.

The results presented in Tables 1 and 2 indicate

that the performance of Linear Regression (LR) was

consistently lower than that of the other two models

evaluated (which is repeated in the following tables).

Additionally, in Table 1, the two worst RMSE val-

ues were recorded in the second half of 2020, which

we believe is directly associated with the impact of

the COVID-19 pandemic. In contrast, the best results

for both evaluation metrics were obtained in 2023 and

Prediction of Daily Sales of Individual Products in a Medium-Sized Brazilian Supermarket Using Recurrent Neural Networks Models

743

Table 1: RMSE for each semester, considering the average results for all products.

2020 2021 2022 2023 2024

1º 2º 1º 2º 1º 2º 1º 2º 1º 2º

LR 4,851 9,021 5,756 6,096 7,722 7,523 6,434 5,057 4,520 3,160

LSTM 3,955 7,445 3,615 3,729 4,551 4,782 3,994 2,680 3,163 2,120

GRU 3,889 7,437 3,299 3,604 4,304 4,680 4,125 2,710 2,405 2,203

Table 2: MAPE for each semester, considering the average results for all products.

2020 2021 2022 2023 2024

1º 2º 1º 2º 1º 2º 1º 2º 1º 2º

LR 68,28 75,22 76,91 75,34 87,43 84,22 78,34 78,32 75,28 75,87

LSTM 63,16 57,59 55,54 51,68 47,85 48,28 43,94 47,26 49,08 48,46

GRU 62,62 54,44 59,06 48,75 45,81 47,98 44,77 44,10 45,92 50,98

Table 3: RMSE for each semester, considering the average results for the products by each category.

2020 2021 2022 2023 2024

1º 2º 1º 2º 1º 2º 1º 2º 1º 2º

LR 5,642 7,642 5,526 6,488 11,391 9,353 9,259 5,800 5,609 5,024

LSTM 4,848 5,091 3,175 3,484 7,133 5,367 5,844 3,594 3,378 3,193

GRU 4,891 4,41 3,027 3,169 6,150 4,628 4,952 3,352 3,400 2,854

LR 13,27 62,31 22,69 21,12 22,44 30,01 18,21 10,25 8,35 6,89

LSTM 11,25 58,08 17,96 17,68 19,70 28,73 16,73 9,090 9,630 7,180

GRU 11,19 59,11 17,63 19,16 20,28 30,21 23,23 9,411 9,737 9,756

LR 3,461 4,016 3,839 4,334 5,443 5,064 4,468 4,050 4,306 3,146

LSTM 2,822 3,088 2,187 2,257 2,829 2,206 2,211 1,893 2,259 1,610

GRU 2,725 2,482 1,738 1,802 2,768 1,929 1,658 1,774 2,148 1,726

LR 2,368 2,507 2,855 2,402 6,093 4,630 3,336 2,650 2,392 2,110

LSTM 2,146 1,938 1,978 1,497 3,055 3,189 2,214 1,633 1,508 1.214

GRU 2,058 1,572 1,518 1,253 2,911 2,954 2,015 1,324 1,339 1,099

LR 4,214 4,666 6,222 8,477 7,363 5,984 6,552 7,988 6,432 4,606

LSTM 3,295 2,743 3,012 4,824 2,274 2,719 3,388 3,943 3,040 1,956

GRU 3,103 2,413 2,609 4,896 3,068 3,413 3,288 4,029 2,775 2,060

LR 4,140 4,276 4,435 4,435 4,775 5,065 6,120 4,292 3,793 3,429

LSTM 3,168 2,539 2,063 2,053 1,751 2,099 3,114 1,851 1,565 1,378

GRU 3,102 2,242 1,799 1,808 1,718 1,62 2,851 1,385 1,343 1,412

LR 3,808 3,404 3,9,7 3,804 4,325 4,649 4,204 4,338 3,637 2,862

LSTM 2,962 2,127 1,849 1,784 1,632 1,844 1,758 1,779 1,334 1,075

GRU 2,892 1,887 1,597 1,57 1,443 1,802 1,617 1,719 1,303 1,087

LR 6,464 6,679 6,112 5,417 4,234 5,156 4,484 3,584 3,142 2,189

LSTM 4,110 3,178 2,910 2,827 2,351 3,123 2,307 1,792 1,744 1,368

GRU 4,113 3,012 2,805 2,645 2,190 2,843 2,872 1,717 1,549 1,772

2024, particularly with the LSTM and GRU mod-

els. This improvement may also be linked to the pan-

demic, as the models were trained using data from

previous years. Consequently, the predictive perfor-

mance has improved from 2023 onwards when the

training dataset comprises only post-pandemic peak

data (i.e., from 2022 onwards). Furthermore, as seen

in Tables 1 and 2, in 2024, RMSE values decreased

to approximately 2. However, the MAPE values re-

mained relatively high.

We also evaluated the experimental metrics con-

sidering each of the product groups described in Sub-

section 4. Tables 3 and 4 present the RMSE and

MAPE results for each of these subsets, with a struc-

ture similar to that of Tables 1 and 2, but every three

rows, the results refer to one of the product groups.

In Tables 3 and 4, a pattern of results similar to

that described in Tables 1 and 3 can be observed,

except for the products in groups C

and C

(alco-

holic and non-alcoholic beverages), which presented

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

744

Figure 2: Real and predicted sales values for two products over 120 day.

Table 4: MAPE for each semester, considering the average results for the products by each category.

2020 2021 2022 2023 2024

1º 2º 1º 2º 1º 2º 1º 2º 1º 2º

LR 80,11 72,59 84,13 87,27 95,00 97,81 96,07 102,77 95,30 106,3

LSTM 76,26 62,24 60,46 58,61 58,48 58,57 59,71 62,83 60,38 64,18

GRU 77,05 57,70 53,85 50,87 54,95 49,13 47,10 57,39 62,08 60,37

LR 82,61 195,8 168,9 122,7 125,7 192,5 110,7 88,24 74,98 72,78

LSTM 75,78 116,3 144,2 97,75 92,90 99,36 94,70 94,92 118,0 84,84

GRU 75,50 143,7 144,9 112,7 125,9 170,3 116,4 94,91 104,3 154,5

LR 66,69 67,48 66,66 65,34 76,94 74,63 72,41 73,41 72,87 77,51

LSTM 62,59 53,12 49,37 47,93 44,51 40,61 42,83 40,56 41,09 46,71

GRU 61,74 45,02 39,99 39,10 38,12 35,63 39,61 38,05 41,28 39,44

LR 61,16 60,37 63,50 62,26 127,6 71,75 71,10 66,34 66,58 65,31

LSTM 60,32 55,24 49,65 48,51 46,74 47,22 48,31 45,26 50,27 47,29

GRU 59,32 46,73 39,81 40,34 39,90 42,98 41,58 41,06 45,05 45,23

LR 58,89 64,63 70,93 84,37 73,15 67,75 77,09 80,55 81,56 80,83

LSTM 54,96 53,49 47,74 48,35 38,50 40,98 37,92 36,95 37,00 40,51

GRU 55,12 46,94 41,41 50,75 33,26 35,95 35,09 34,40 32,39 41,48

LR 58,89 64,63 70,93 84,37 73,15 67,75 77,09 80,55 81,56 80,83

LSTM 55,22 44,82 38,37 42,45 38,12 39,45 40,40 41,11 40,24 39,98

GRU 61,07 41,57 46,95 49,27 44,86 45,62 50,50 45,45 49,30 49,09

LR 62,37 57,60 61,89 61,31 61,89 65,50 65,74 64,12 64,76 61,86

LSTM 56,66 46,12 40,05 39,34 33,12 33,63 36,21 33,10 31,22 32,13

GRU 55,12 40,54 34,24 34,40 28,37 29,52 32,29 32,22 28,93 29,8

LR 86,62 86,46 81,05 71,85 72,92 81,00 78,20 72,53 78,05 73,67

LSTM 60,51 45,60 48,45 52,63 50,64 49,34 50,25 44,18 53,86 54,26

GRU 54,94 43,37 32,69 34,75 36,82 32,26 30,11 33,31 33,05 39,49

the worst results. The results for the other categories

were slightly better than the general average for all

products, but the problem remains that the average

MAPE values are high, consistently above 28.93.

We also evaluated the experimental results, con-

sidering only the 25 products with the best MAPE

Prediction of Daily Sales of Individual Products in a Medium-Sized Brazilian Supermarket Using Recurrent Neural Networks Models

745

Table 5: RMSE for each semester, considering the average results for the products with better results.

2020 2021 2022 2023 2024

1º 2º 1º 2º 1º 2º 1º 2º 1º 2º

LR 5,682 5,437 5,711 6,241 6,072 7,916 7,301 6,522 6,086 5,160

LSTM 3,998 2,767 2,123 2,486 1,783 2,726 2,356 1,740 1,911 1,550

GRU 3,749 2,149 1,819 2,187 1,699 2,724 2,287 2,052 2,019 1,536

Table 6: MAPE for each semester, considering the average results for the products with better results.

2020 2021 2022 2023 2024

1º 2º 1º 2º 1º 2º 1º 2º 1º 2º

LR 65,91 59,38 60,81 71,70 65,74 60,75 69,99 72,74 66,19 75,73

LSTM 55,27 36,84 28,30 32,50 22,14 19,32 22,06 19,38 18,97 18,66

GRU 52,69 27,55 20,69 23,91 18,87 17,20 19,76 18,71 19,80 19,22

results with LSTM or GRU. The RMSE and MAPE

results for the 10% of products in the database that

had the best MAPE results are in Tables 5 and 6.

Regarding the products with the best MAPE, the

RMSE is close to that obtained in several product cat-

egories, and the average MAPE, as of 2022, is always

lower than 22.14, reaching 17.20. Considering the

sales volume of a medium-sized supermarket, an error

close to 17% can be considerable.

Considering 2 of the 25 products with the best pre-

diction results, Figure 2 is a graph of the predicted

value (in orange) compared to the actual value sold

(in blue) over 120 days for two products of this set.

In the graph, it is possible to observe that for some

of the products, the methods based on RNNs could

follow the sales pattern.

6 CONCLUSION

In this study, we aimed to forecast daily sales of indi-

vidual products in a medium-sized supermarket using

different machine learning algorithms. By focusing

on a diverse set of products across multiple categories,

we demonstrated that forecasting accuracy varies sig-

niﬁcantly from product to product, emphasizing the

complexity of predicting retail sales at such a granular

level. The database used in this work, which includes

ﬁve years of sales data from a Brazilian supermarket,

is a valuable resource that can contribute to future re-

search in sales forecasting.

The experimental results show that RNNs (LSTM

and GRU) outperformed traditional Linear Regres-

sion models in terms of accuracy. For some of the

products analyzed, the results produced can be con-

sidered adequate; however, the performance varied

widely, with some products yielding higher predic-

tion errors, especially those in the non-alcoholic and

alcoholic beverage categories. The high MAPE val-

ues, particularly for products with lower sales vol-

ume, suggest room for further improvement in the

models.

Future studies could explore integrating additional

features, such as promotional periods or external fac-

tors like holidays and weather, which may help im-

prove the models’ generalization. Future work may

also involve adjusting the hyperparameters of the al-

gorithms and experimenting with other ML models.

In conclusion, this work contributes to understanding

sales forecasting at the product level in medium-sized

supermarkets, highlighting both the potential and the

challenges of using deep learning techniques for this

task. The insights gained can be used to improve in-

ventory management and enhance supply chain oper-

ations in the retail industry.

ACKNOWLEDGEMENTS

We want to thank the supermarkets that kindly pro-

vided the data used in the experiments and who chose

not to have their identities disclosed.

REFERENCES

Aldhaheri, A., Alwahedi, F., Ferrag, M. A., and Battah, A.

(2024). Deep learning for cyber threat detection in iot

networks: A review. Internet of Things and cyber-

physical systems, 4:110–128.

Almufadi, N., Alblihed, N., Alhabeeb, S., Alhumud, S.,

and Selmi, A. (2023). Sales prediction based on

data mining techniques. In International Conference

on Emerging Smart Technologies and Applications,

pages 1–6. IEEE.

Archana, R. and Jeevaraj, P. E. (2024). Deep learning mod-

els for digital image processing: a review. Artiﬁcial

Intelligence Review, 57(1):11.

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

746

Brockwell, P. J. and Davis, R. A. (2016). Introduction to

Time Series and Forecasting. Springer.

Chatﬁeld, C. (2004). The Analysis of Time Series: An Intro-

duction. Chapman & Hall/CRC, 6 edition.

Chen, I.-F. and Lu, C.-J. (2017). Sales forecasting by com-

bining clustering and machine-learning techniques for

computer retailing. Neural Computing and Applica-

tions, 28:2633–2647.

Cho, K., Van Merri

enboer, B., Bahdanau, D., and Bengio,

Y. (2014). On the properties of neural machine trans-

lation: encoder-decoder approaches. arXiv preprint

arXiv:1409.1259.

Chung, J., Gulcehre, C., Cho, K., and Bengio, Y.

(2014). Empirical evaluation of gated recurrent neu-

ral networks on sequence modeling. arXiv preprint

arXiv:1412.3555.

Dai, Y. and Huang, J. (2021). A sales prediction method

based on lstm with hyper-parameter search. Journal

of Physics: Conference Series, 1756(1):012015.

De Almeida, F. M., Martins, A. M., Nunes, M. A., and Bez-

erra, L. C. (2022). Retail sales forecasting for a brazil-

ian supermarket chain: an empirical assessment. In

IEEE Conference on Business Informatics, volume 1,

pages 60–69. IEEE.

Gupta, G., Gupta, K. L., and Kansal, G. (2022). Megamart

sales prediction using machine learning techniques. In

International Conference on Computing, Communica-

tions, and Cyber-Security, pages 437–446. Springer.

Hamilton, J. D. (1994). Time Series Analysis. Princeton

University Press.

Hochreiter, S. (1998). The vanishing gradient problem dur-

ing learning recurrent neural nets and problem solu-

tions. International Journal of Uncertainty, Fuzziness

and Knowledge-Based Systems, 6(02):107–116.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural Computation, 9(8):1735–1780.

Huo, Z. (2021). Sales prediction based on machine learn-

ing. In International Conference on E-Commerce and

Internet Technology, pages 410–415. IEEE.

Jeyarangani, J., Sumanth, G. B., Raviram, C., and

Sreekanth, M. V. (2023). Regressor based supermar-

ket sales prediction using time series data. In Interna-

tional Conference on Futuristic Technologies, pages

1–4. IEEE.

Kohli, S., Godwin, G. T., and Urolagin, S. (2020). Sales

prediction using linear and knn regression. In Ad-

vances in Machine Learning and Computational In-

telligence, pages 321–329. Springer.

Le, X.-H., Ho, H. V., Lee, G., and Jung, S. (2019). Applica-

tion of long short-term memory (lstm) neural network

for ﬂood forecasting. Water, 11(7):1387.

Meulstee, P. and Pechenizkiy, M. (2008). Food sales pre-

diction:” if only it knew what we know”. In IEEE

International Conference on Data Mining Workshops,

pages 134–143. IEEE.

Pascanu, R., Mikolov, T., and Bengio, Y. (2013). On the

difﬁculty of training recurrent neural networks. In In-

ternational Conference on Machine Learning, pages

1310–1318, Atlanta, USA. Pmlr.

Raizada, S. and Saini, J. R. (2021). Comparative analysis of

supervised machine learning techniques for sales fore-

casting. International Journal of Advanced Computer

Science and Applications, 12(11).

Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986).

Learning representations by back-propagating errors.

Nature, 323(6088):533–536.

Santra, A. S. and Lin, J.-L. (2019). Integrating long short-

term memory and genetic algorithm for short-term

load forecasting. Energies, 12(11):2040.

Sherstinsky, A. (2020). Fundamentals of recurrent neu-

ral network (RNN) and long short-term memory

(LSTM) network. Physica D: Nonlinear Phenomena,

404:132306.

Su, X., Yan, X., and Tsai, C.-L. (2012). Linear regres-

sion. Wiley Interdisciplinary Reviews: Computational

Statistics, 4(3):275–294.

Wei, W. W. S. (2006). Time Series Analysis: Univariate and

Multivariate Methods. Pearson/Addison Wesley, .

Prediction of Daily Sales of Individual Products in a Medium-Sized Brazilian Supermarket Using Recurrent Neural Networks Models

747