Original Research Article

A Comprehensive Study of Walmart Sales Predictions Using Time

Series Analysis

Abstract:
This article presents a comprehensive study of sales predictions using time series analysis, focusing on a case study of
Walmart sales data. The aim of this study is to evaluate the effectiveness of various time series forecasting techniques in
predicting weekly sales data for Walmart stores. Leveraging a dataset from Kaggle comprising weekly sales data from
various Walmart stores around the United States, this study explores the effectiveness of time series analysis in
forecasting future sales trends. Various time series analysis techniques, including AutoRegressive Integrated Moving
Average (ARIMA), Seasonal AutoRegressive Integrated Moving Average (SARIMA), Prophet, Exponential Smoothing,
and Gaussian Processes, are applied to model and forecast Walmart sales data. By comparing the performance of these
models, the study seeks to identify the most accurate and reliable methods for forecasting retail sales, thereby providing
valuable insights for improving sales predictions in the retail sector. The study includes an extensive exploratory data
analysis (EDA) phase to preprocess the data, detect outliers, and visualize sales trends over time. Additionally, the article
discusses the partitioning of data into training and testing sets for model evaluation. Performance metrics such as Root
Mean Squared Error (RMSE) and Mean Absolute Error (MAE) are utilized to compare the accuracy of different time series
models.

The results indicate that Gaussian Processes outperform other models in terms of accuracy, with an RMSE of 34,116.09
and an MAE of 25,495.72, significantly lower than the other models evaluated. For comparison, ARIMA and SARIMA
models both yielded an RMSE of 555,502.2 and an MAE of 462,767.3, while the Prophet model showed an RMSE of
567,509.2 and an MAE of 474,990.8. Exponential Smoothing also performed well with an RMSE of 555,081.7 and an
MAE of 464,110.5. These findings suggest the potential of Gaussian Processes for accurate sales forecasting. However,
the study also highlights the strengths and weaknesses of each forecasting methodology, emphasizing the need for further
research to refine existing techniques and explore novel modeling approaches. Overall, this study contributes to the
understanding of time series analysis in retail sales forecasting and provides insights for improving future forecasting
endeavors.

Keywords: Sales prediction, Time series analysis, Walmart, ARIMA, SARIMA, Prophet, Exponential Smoothing,
Gaussian Processes, Forecasting, Model evaluation.

1. Introduction
In many fields, predictive modeling has become an effective tool for predicting future trends and helping decision-makers
make data-driven choices [1]. In particular, time series analysis is essential for forecasting future values based on
previously collected data points arranged chronologically [2]. In this paper, we explore the use of Walmart sales data from
Kaggle, a well-known platform for data science competitions and datasets, to construct predictive models through the
application of time series analysis. The Walmart sales dataset was collected from various Walmart stores around the
United States from the period of February 5, 2010, to October 26, 2012.
Along with extra features like holiday flags, temperature, fuel prices, the Consumer Price Index (CPI), and unemployment
rates, the dataset comprises weekly sales data from various Walmart stores [3]. Our goal is to create precise predictive
models that can predict future sales based on historical trends and pertinent predictors by utilizing this dataset.

In order to identify underlying patterns, trends, and seasonal variations, time series analysis entails evaluating and
modeling data points gathered at regular intervals over an extended period of time [4]. It includes a range of approaches,
including autocorrelation analysis, decomposition, smoothing techniques, and forecasting strategies like seasonal ARIMA
and ARIMA (AutoRegressive Integrated Moving Average) [6]. With the help of these methods, we are able to identify the
temporal dependencies in the data and forecast future trends in sales.

The first part of our study focuses on employing time series analysis techniques to model and forecast Walmart sales
data. We will explore methods to preprocess the data, identify seasonality and trends, and select appropriate models to
generate accurate forecasts. By evaluating the performance of our time series models against historical data, we aim to
assess their effectiveness in capturing the inherent patterns and variability present in Walmart sales.

We aim to show in this paper the applicability and effectiveness of time series analysis for predictive modeling tasks, with
a focus on retail sales forecasting. Furthermore, our objective is to offer an understanding of the real-world uses of time
series methods and how they affect Walmart-style decision-making.

1.1. Importance of Time Series Analysis in Sales Forecasting

Time series analysis plays a crucial role in sales forecasting, providing valuable insights into past trends and patterns that
can inform future predictions. Some importance of time series analysis in sales forecasting are as follows.

Identification of Trends and Seasonal Patterns:

Time series analysis helps identify trends, seasonal variations, and cyclical patterns in sales data [4]. These insights
enable businesses to anticipate demand fluctuations and adjust their strategies accordingly [7].

Forecasting Accuracy:
By analyzing historical sales data using time series techniques, businesses can develop accurate forecasting models [6].
These models take into account past sales performance and external factors, leading to more reliable predictions [8].

Resource Allocation and Inventory Management:

Sales forecasts derived from time series analysis help businesses allocate resources efficiently and manage inventory
levels effectively [9]. This optimization minimizes stockouts and excess inventory, enhancing operational efficiency [10].

Marketing and Promotion Strategies:

Understanding sales patterns over time enables businesses to optimize their marketing and promotion strategies [11].
Time series analysis helps identify peak sales periods and consumer behavior trends, guiding targeted marketing efforts
[12].

Budgeting and Financial Planning:

Accurate sales forecasts derived from time series analysis facilitate budgeting and financial planning processes [13].
Businesses can use these forecasts to set realistic revenue targets, allocate budgets effectively, and make informed
investment decisions [14].
1.2. Background
Based on past data observations, time series analysis predictions provide insightful information about potential future
trends and patterns. In order to predict future values, time series forecasting entails evaluating sequential data points
gathered over time [4]. Applications of this methodology can be found in a number of disciplines, including epidemiology,
economics, finance, and climate science, where the ability to recognize and anticipate temporal patterns is essential for
making decisions.
For example, time series analysis was used in a study by [5] to predict COVID-19 trends in Coffee County, Tennessee,
United States. To forecast future infection rates, the researchers used time series forecasting techniques and historical
data on COVID-19 cases. They were able to offer insightful analysis of historical trends and patterns, which helped
policymakers and public health experts create efficient plans for handling the pandemic.

As stated by [6], time series forecasting techniques range from straightforward approaches like exponential smoothing to
more intricate models like ARIMA and machine learning algorithms like LSTM (Long Short-Term Memory) networks. By
identifying underlying patterns, seasonality, and trends in the data, these techniques enable precise forecasting and well-
informed decision-making.

Time series analysis predictions are therefore essential in many fields because they use past data to predict future trends
and patterns. Organizations and policymakers can make well-informed decisions to address challenges and capitalize on
opportunities in their respective fields by having a thorough understanding of past behaviors and trends.

In the realm of retail sales forecasting, various methodologies have been explored to enhance prediction accuracy and
efficacy. a comparative study focusing on linear and nonlinear models for aggregate retail sales forecasting was conducted
by [15]. Their research delved into the effectiveness of traditional seasonal forecasting methods alongside neural
networks, highlighting the significance of seasonal adjustments and deseasonalization techniques in improving
forecasting accuracy. This study laid the groundwork for understanding the benefits and limitations of different modeling
approaches in the retail domain.

In a recent development, a novel approach to retail sales forecasting was introduced by [16] through meta-learning. Their
study proposed a meta-learning framework based on deep convolutional neural networks, which automatically learn
feature representations from raw sales time series data. By combining these learned features with a set of weights, the
framework integrates a pool of base-forecasting methods, ultimately leading to superior forecasting performance
compared to conventional benchmarks. Ma and Fildes' work underscores the importance of leveraging advanced machine
learning techniques to enhance forecasting capabilities in the retail sector.

Furthermore, evaluating the performance of state space and ARIMA models in consumer retail sales forecasting was
contributed to the field by [17]. Their research provided insights into the comparative effectiveness of these models,
demonstrating their applicability in predicting sales trends for various product categories. Through the analysis of different
methodologies, Ramos et al. shed light on the strengths and weaknesses of state space and ARIMA models, offering
valuable guidance for researchers and practitioners seeking to optimize forecasting strategies in retail environments.

Integrating findings from these seminal works into the background of our study enriches our understanding of the diverse
methodologies employed in retail sales forecasting. By building upon the insights gleaned from examination of linear and
nonlinear models[15], innovative meta-learning approach[16], and comparative analysis of state space and ARIMA
models[17], our research aims to contribute to this evolving field by exploring the performance of various time series
models in predicting Walmart sales data.

Predictive modeling has become a vital tool across various fields, aiding decision-makers by forecasting future trends
based on historical data. This study leverages Walmart sales data from Kaggle, covering weekly sales from February 5,
2010, to October 26, 2012. The goal is to develop accurate predictive models to forecast future sales by analyzing
historical trends and relevant predictors using time series analysis techniques. These techniques, including ARIMA,
SARIMA, Prophet, Exponential Smoothing, and Gaussian Processes, help identify temporal dependencies and forecast
future sales trends. This paper aims to demonstrate the applicability and effectiveness of time series analysis in retail
sales forecasting and its impact on decision-making in Walmart-like retail environments. Additionally, the importance of
time series analysis in sales forecasting is highlighted, showing how it aids in identifying trends, improving forecasting
accuracy, optimizing resource allocation, and enhancing marketing strategies. By incorporating insights from past studies
on time series forecasting, this research aims to further the understanding of predictive modeling in retail sales.

2. Materials and Methods

This section outlines the comprehensive methodology adopted for the sales prediction study, including data collection and
preparation, exploratory data analysis, model building, model evaluation and comparison.

2.1. Materials
Data Collection:
The dataset utilized in this study was sourced from Kaggle, encompassing weekly sales data from multiple Walmart stores
across the United States. The dataset contains comprehensive information on sales trends over a specified timeframe,
providing a rich source for time series analysis.

It's important to note that the dataset has a limited time range, covering the period from 2010 to 2012. This restriction
stems from the unavailability of more recent data due to constraints imposed by Walmart. The company, understandably,
restricts the dissemination of its recent sales data for proprietary and competitive reasons. Consequently, researchers
and analysts are constrained to rely on historical data for their analyses, limiting the ability to capture and analyze more
recent trends or changes in consumer behavior hence its applicability to current market conditions may be limited,
necessitating caution in extrapolating findings to present-day contexts. Researchers should consider this limitation when
interpreting the findings and extrapolating them to current market conditions.

Despite its age, the Walmart sales dataset offers a comprehensive and valuable resource for understanding sales
dynamics in the retail industry. By leveraging the rich store-level data and information provided in the dataset, researchers
and practitioners can gain insights into sales trends, identify key drivers of sales performance, and develop effective
strategies for maximizing revenue and enhancing business outcomes.

2.2. Method
2.2.1. Exploratory Data Analysis (EDA):
Before model implementation, an extensive exploratory data analysis (EDA) phase was conducted. This involved
detecting outliers and addressing them using winsorization, a robust technique to handle extreme values without removing
them entirely. By winsorizing the dataset, outliers were replaced with more moderate values, ensuring that they did not
unduly influence the analysis.
Data Import and Sorting:
 The Walmart sales data was imported into R-Studio from the provided CSV file.
 The Date column was converted to the Date format, ensuring consistency and ease of manipulation.
 The dataset was sorted by date to arrange the observations chronologically.

Decomposition plot:
A decomposition plot is essential for gaining a deeper understanding of the underlying structure of the Walmart sales data,
identifying key trends and patterns. By examining these components of the decomposition plot, we can gain insights into
the underlying patterns, trends, and seasonal variations present in the data, which can be useful for forecasting, analysis,
and decision-making.

Figure 1: Decomposition plot for Walmart Sales

The trend component of the decomposed time series plot (Figure 1) exhibits a pattern of going up and down over time, it
suggests that the overall direction of the data is not consistently increasing or decreasing but rather fluctuating cyclically.
This observation aligns well with the presence of cyclical patterns in the Walmart sales data. Seasonal effects, such as
changes in consumer behavior, purchasing patterns, and demand for certain products, often lead to fluctuations in sales
over time.

The following can be deduced from the cyclical pattern presented in the Walmart data.
Seasonal Variation:
Seasonal changes, such as holidays, back-to-school seasons, summer vacations, and winter holidays, significantly impact
consumer spending habits and preferences.
For example, sales of certain items like outdoor furniture, grills, and swimwear may peak during the summer season, while
sales of winter clothing, holiday decorations, and cold-weather accessories may increase during the winter season. These
seasonal variations create cyclical patterns in sales data, with sales rising and falling in a predictable manner as seasons
change throughout the year.
Impact on Sales:
The cyclical pattern underscores the influence of seasonal factors on Walmart's sales performance.
As seasons alternate, customers' needs, preferences, and purchasing behavior shift accordingly, leading to fluctuations
in sales volumes and revenue. Understanding these seasonal patterns is therefore crucial for Walmart to optimize
inventory management, pricing strategies, marketing campaigns, and staffing levels to meet customer demand effectively
during peak seasons.

Outlier Detection
 Box Plots and Scatter Plots
Box plots and scatterplots were utilized to visualize potential outliers in the Weekly_Sales column.

Figure 2: Boxplot showing potential outliers in the Weekly Sales.

Figure 3: Scatterplot Showing Potential Outliers in the Weekly Sales.

Based on the analysis of boxplots and scatterplots, there is evidence suggesting the presence of outliers in the
Weekly_Sales column. The boxplot (Figure 2) reveals the existence of data points that fall significantly outside the
whiskers, indicating values that lie far from the median and quartiles. Additionally, the scatterplot (Figure 3) illustrates
several data points that appear to deviate substantially from the overall trend, suggesting potential anomalies in the sales
data.
 Quantile-based outlier detection
Quantile-based outlier detection method was employed to identify outliers using lower and upper bounds. The following
outliers were determined based on this method.
2789469, 2939946, 2766400, 2921710, 2811634, 2752122, 2740057, 2811647, 2771647, 2762861, 2819193, 3436008,
3526713, 2727575, 3749058, 3595903, 3818686, 3766687, 2734277, 3078162, 3004702, 2950199, 2864171, 2906233,
2771397, 2760347, 2762817, 3224370, 3676389, 3487987, 3556766, 3369069, 3555371, 2739020

Outlier Treatment (Winsorization and Visualization of the Outliers):

Winsorization was chosen as the method to handle outliers due to its ability to mitigate their impact on subsequent
analyses while preserving the overall distribution of the data. When outliers are present, they can disproportionately
influence statistical measures such as means, variances, and correlations, potentially skewing the results and leading to
inaccurate conclusions. Winsorization addresses this issue by capping extreme values at predetermined percentiles (e.g.,
the 95th and 5th percentiles), effectively reducing their influence without removing them entirely from the dataset.
Furthermore, winsorization is a relatively straightforward and transparent method, making it well-suited for addressing
outliers in a wide range of analytical contexts.

By winsorizing the outliers, we retained the information contained in these extreme values while minimizing their disruptive
effect on subsequent analyses. This approach ensured that the data remained representative of the underlying distribution
while simultaneously improving the robustness and reliability of statistical inferences drawn from the dataset. By opting
for winsorization, we were able to effectively manage outliers in our dataset while preserving its integrity for further analysis
and interpretation.

Quantiles for winsorization were calculated to replace extreme values with their corresponding percentiles.
The winsorized dataset was visualized by plotting weekly sales over time, revealing trends and patterns in the data (See
Figure 4). Regular fluctuations are evident, suggesting a seasonal pattern in sales. Peaks and troughs occur at consistent
intervals, likely corresponding to seasonal shopping events and holidays.

Figure 4: Plot of Winsorized Weekly Sales Over Time

The winsorized dataset was also visualized using a boxplot (See Figure 5). The boxplot displayed the distribution of the
winsorized sales data, indicating the median, interquartile range (IQR). No outliers were observed after winsorization.
 Figure 5: Boxplot of Winsorized Weekly Sales

The EDA process facilitated a deeper understanding of the dataset, allowing for the identification and treatment of outliers
and the visualization of sales trends over time. These insights serve as a foundation for building time series models to
forecast Walmart sales accurately.

2.2.2. Model Implementation:

This comprehensive exploration and preprocessing of the Walmart sales data set the stage for subsequent time series
modeling, which involved fitting ARIMA, SARIMA, Prophet AND Gaussian Processes models to generate forecasts and
evaluate their performance.

i. Justification for model selection

Each of the selected models – ARIMA, SARIMA, Prophet, Exponential Smoothing, and Gaussian Processes – offers
unique advantages and is suitable for different aspects of time series analysis. Here's a justification for the use of each
model:

ARIMA (AutoRegressive Integrated Moving Average):

 ARIMA is a widely used and well-established model for time series forecasting.
 It can capture linear dependencies between past observations and forecast future values.
 ARIMA is versatile and can handle a wide range of time series data, making it a suitable choice for initial
exploration and benchmarking in time series analysis.
 The model's parameters can be automatically selected based on criteria like the Akaike Information Criterion (AIC)
or Bayesian Information Criterion (BIC), simplifying the modeling process.
ARIMA is a widely used model for time series forecasting. It combines three components: autoregression (AR),
differencing (I), and moving average (MA). The mathematical equation for ARIMA(p, d, q) can be written as:

Yt= μ +ϕ1Yt−1+ϕ2Yt−2+...+ϕpYt−p+θ1εt−1+θ2εt−2+...+θqεt−q+εt

Where:
 Yt is the value of the time series at time t.
 μ is the mean of the time series.
  ϕ1, . . . ,ϕi the autoregressive coefficients.
 θ1, . . . ,θi are the moving average coefficients.
 εt is the error term at time t.
 p is the order of the autoregressive component.
 d is the degree of differencing.
 q is the order of the moving average component.

In the AutoRegressive Integrated Moving Average (ARIMA) model, future values are forecasted by considering
the autoregressive (AR) component, the differencing (I) component, and the moving average (MA) component,
which collectively capture the temporal dependencies and trends present in the time series data.

- Autoregressive (AR) Component: The autoregressive part (ϕ1Yt−1+ϕ2Yt−2+...+ϕpYt−p ) captures the

relationship between the current observation (Yt) and its previous values (Yt−1, Yt−2,...,Yt−p ). This component
represents how the current value of the time series is influenced by its past values.
- Moving Average (MA) Component: The moving average part (θ1εt−1+θ2εt−2+...+θqεt−q) captures the
relationship between the current observation and the past forecast errors (εt−1, εt−2,..., εt−q). This component
represents the impact of past errors on the current observation.
- Mean Component: The mean component (μ) represents the average level of the time series.

The model combines these components along with differencing (if required) to forecast future values. By estimating the
parameters (ϕ1, ϕ2,...,ϕp, θ1, θ2,...,θq ), the model predicts the value of Yt+1 based on the observed values up to time t
and the forecast errors up to time t. The process is iterative, with each forecasted value being used to predict the next
one.

SARIMA (Seasonal AutoRegressive Integrated Moving Average):

 SARIMA extends the capabilities of ARIMA by incorporating seasonal components.
 Seasonality is a common feature in many time series datasets, especially in retail sales where seasonal trends
often occur.
 SARIMA allows for the modeling of both non-seasonal and seasonal components, making it suitable for capturing
complex seasonal patterns in sales data.
The SARIMA model extends ARIMA to account for seasonality in the data. The mathematical equation for SARIMA(p, d,
q)(P, D, Q)m is represented as:

Yt=μ+ϕ1Yt−1+ ϕ2Yt−2+…+ϕpYt−p+θ1ϵt−1+…+θqϵt−q+ϕs1Yt−s+ ϕs2Yt−s2+…+ϕ spYt− sp+ϵt

Where:
 Yt is the value of the time series at time t.
 μ is the mean of the time series.
 ϕ1, . . . , ϕi are the autoregressive coefficients.
 θ1, . . . , θi are the moving average coefficients.
 ϕs1, . . . , ϕsp are the seasonal autoregressive coefficients representing the effect of past seasonal values
on the current value of the time series.
 εt is the error term at time t.
 p is the order of the autoregressive component.
 d is the degree of differencing.
 q is the order of the moving average component.
 P is the seasonal order of the autoregressive component.
 D is the seasonal degree of differencing.
  Q is the seasonal order of the moving average component.

The Seasonal Autoregressive Integrated Moving Average (SARIMA) model forecasts future values by extending the
ARIMA model to account for seasonality.

- Autoregressive (AR) Component: Similar to ARIMA, the autoregressive part (ϕ1Yt−1+ ϕ2Yt−2+…+ϕpYt−p)
captures the relationship between the current observation (Yt) and its previous values (Yt−1, Yt−2,...,Yt−p).
- Moving Average (MA) Component: The moving average part (θ1εt−1+θ2εt−2+...+θqεt−q) captures the
relationship between the current observation and the past forecast errors (εt−1, εt−2,..., εt−q).
- Seasonal Autoregressive (SAR) Component: The seasonal autoregressive part (ϕs1Yt−s+ ϕs2Yt−s2+…+ϕsp
Yt−sp) captures the relationship between the current observation and its corresponding observation from the same
season in previous years. This component accounts for seasonal patterns.
- Mean Component: The mean component (μ) represents the average level of the time series.

The model combines these components to forecast future values. By estimating the parameters (ϕ1, ϕ2,...,ϕp, θ1, θ2,...,θq,
ϕs1, ϕs2+…+ϕsp), along with the seasonal period (s), the model predicts the value of Yt+1 based on the observed values
up to time t and the forecast errors up to time t. The inclusion of the seasonal autoregressive component allows the model
to capture recurring patterns at seasonal intervals.

PROPHET:
 Prophet is a forecasting tool developed by Facebook that is specifically designed for time series data with strong
seasonal effects and multiple seasonality.
 It can handle irregularly spaced data and missing values, which is beneficial for real-world datasets that may have
gaps or inconsistencies.
 Prophet offers flexibility in modeling holidays and special events, which is crucial for retail sales forecasting where
holidays often influence consumer behavior.
 While Prophet may not always outperform traditional time series models like ARIMA and SARIMA, it provides an
alternative approach with simplified implementation and tuning.

The Prophet model is based on an additive model where non-linear trends are fit with yearly, weekly, and
daily seasonality, plus holiday effects. The mathematical equation for Prophet is not explicitly defined but is
The Prophet forecasting model forecasts future values by decomposing the time series into several components:
represented as:

y(t)=g(t)+s(t)+h(t)+εt

Where:
 Trend Component (g(t)): The trend component captures the underlying direction of the time series,
representing long-term changes over time. It is modeled using a piecewise linear or logistic function that adapts
to the data.
  Seasonality Component (s(t)): The seasonality component captures periodic patterns in the data, such as
weekly, monthly, or yearly fluctuations. Prophet models seasonality using Fourier series expansions, allowing it
to flexibly capture various seasonal patterns.
 Holiday Component (h(t)): The holiday component accounts for the impact of holidays and special events on
the time series. It allows users to specify custom holiday effects, enabling the model to capture the effects of
holidays on the observed values.
 Error Component (εt): The error component represents random fluctuations or noise in the data that cannot be
explained by the trend, seasonality, or holiday effects. It is assumed to follow a Gaussian distribution with zero
mean.

The model combines these components to forecast future values y(t). By estimating the parameters of the trend,
seasonality, and holiday effects, along with their associated uncertainties, Prophet predicts future values based on the
observed historical data. The flexibility of the model allows it to capture complex patterns and variations in the time
series data, making it particularly suitable for forecasting tasks with multiple sources of variation.

EXPONENTIAL SMOOTHING:
 Exponential Smoothing is a simple yet effective method for modeling time series data, particularly when there are
clear trends and seasonal patterns.
 It is computationally efficient and requires minimal parameter tuning, making it suitable for quick exploratory
analysis or as a baseline model for comparison.
 Exponential Smoothing is robust to outliers and noise in the data, which is advantageous when dealing with real-
world datasets that may contain anomalies.
Exponential smoothing is a simple and commonly used method for time series forecasting. The mathematical
equation for single exponential smoothing is as follows:

Ŷt+1=αYt+(1−α) Ŷt

Where:
 Ŷt+1 is the forecasted value at time t+1.
 Yt is the actual value at time t.
 Ŷt is the forecasted value at time t.
 α is the smoothing parameter (also known as the smoothing factor or alpha value) which controls the weights
given to the current observation Yt and the previous forecast Ŷt. It lies between 0 and 1, where a higher value of
α places more weight on recent observations, resulting in a more responsive forecast.

The equation states that the forecasted value (Ŷt+1) for the next time period (t+1) is calculated as a
weighted average of the observed value (Yt ) and the previous forecasted value (Ŷt ), with the weight
of Yt determined by the smoothing parameter α. The smoothing parameter controls the rate at which
older observations decay in influence on the forecast. A smaller value of α gives more weight to past
observations, while a larger value of α gives more weight to recent observations. As α approaches 1,
the model becomes more responsive to recent data, while as α approaches 0, the model becomes more stable
and relies more on historical data.
GAUSSIAN PROCESSES:
 Gaussian Processes offer a flexible and non-parametric approach to modeling time series data.
 They can capture complex dependencies and uncertainties in the data without assuming a specific functional
form, making them suitable for modeling nonlinear and non-stationary processes.
 Gaussian Processes provide probabilistic forecasts, allowing for the quantification of uncertainty in predictions,
which is valuable for decision-making and risk assessment.
 While Gaussian Processes may be computationally intensive and require careful parameterization, they offer
superior accuracy and flexibility, particularly in scenarios where other models may struggle to capture the
underlying dynamics of the data.
Gaussian Processes are a non-parametric Bayesian approach to regression. The mathematical equation for
Gaussian Processes can be represented as:

f(x)∼GP(m(x),k(x,x′))

Where:
 f(x) represents the forecasted value at input x.
 m(x) represents the mean function, which provides the mean value of the forecast at input x
 GP represents a Gaussian Process distribution.
 k(x,x′) represents the covariance function, which captures the pairwise covariances between the forecasted
values at inputs x and x′.
 The notation x′ denotes another input point in the feature space.

In essence, the Gaussian Process Regression model provides a distribution over possible functions that could
describe the relationship between inputs and outputs. The mean function represents the expected value of the
forecast at each input point, while the covariance function captures how the forecasted values at different input
points co-vary with each other. This allows the model to capture uncertainty in the predictions and provide a
probabilistic forecast at any given input point.

To summarize, the selection of ARIMA, SARIMA, Prophet, Exponential Smoothing, and Gaussian Processes reflects a
balanced approach that leverages the strengths of each model to capture different aspects of the sales forecasting
problem. This diverse set of models allows for comprehensive analysis and comparison, enabling insights into the
performance and suitability of various modeling techniques for retail sales data.

ii. Theoretical Framework:

In time series analysis and predictive modeling, the Box-Jenkins methodology is a cornerstone for developing ARIMA
(AutoRegressive Integrated Moving Average) models. This methodology provides a systematic approach for identifying,
estimating, and validating models for time series data. Even though our study utilizes the auto.arima function from the R-
package, this function automates the Box-Jenkins methodology, ensuring that our model selection is theoretically sound.

Box-Jenkins Methodology and Auto.ARIMA

The Box-Jenkins methodology involves three primary stages:
 Identification:
 Selecting the appropriate ARIMA model by examining the autocorrelation and partial autocorrelation functions of
the time series.
 Estimation:
Using statistical techniques to estimate the parameters of the identified model.
 Diagnostic Checking: Validating the model by analyzing the residuals to ensure they resemble white noise,
indicating a good fit.

The auto.arima function in R automates this process by performing the following:

 Automatically identifying the optimal model parameters (p, d, q) based on criteria such as AIC (Akaike Information
Criterion).
 Estimating these parameters using maximum likelihood estimation.
 Conducting diagnostic checks to ensure the selected model fits the data well.
 By using auto.arima, our study adheres to the principles of the Box-Jenkins methodology while leveraging
automation to enhance efficiency and accuracy in model selection.

Application to Current Study

Our study applies auto.arima alongside other models such as SARIMA, Prophet, Exponential Smoothing, and Gaussian
Processes to forecast Walmart sales data. This approach ensures that our ARIMA models are theoretically grounded in
the Box-Jenkins methodology, providing a robust framework for model evaluation and comparison.

Thus, while our methodology involves automated tools, the underlying theoretical principles remain integral to our
approach. This reinforces the validity of our findings and demonstrates the applicability of established time series theories
in retail sales forecasting.

iii. Data Splitting:

To ensure robust evaluation of our time series analysis models, we adopted a common practice of partitioning the dataset
into training and testing sets. This process is crucial for assessing model performance. The winsorized data was split into
training and test sets (see R-code snippets below).

# Calculate the number of rows for training and testing sets

total_rows <- nrow(walmart_data)
train_rows <- round(0.8 * total_rows) # 80% of total rows
test_rows <- total_rows - train_rows # Remaining rows for testing
# Split the data into training and testing sets
train_data_winsorized <- walmart_data[1:train_rows, ]
test_data_winsorized <- walmart_data[(train_rows + 1):total_rows, ]

In the first step, we determined the total number of rows in the dataset (total_rows). Then, we computed 80% of the total
rows (train_rows) to allocate to the training set, ensuring a sufficient amount of data for model training while retaining a
portion for testing. The remaining rows (test_rows) were designated for the testing set. Subsequently, the dataset was
partitioned accordingly, with the first train_rows rows assigned to the training set (train_data_winsorized) and the
subsequent rows allocated to the testing set (test_data_winsorized).
This systematic approach to data partitioning ensures that our time series models are trained on a representative portion
of the data while preserving unseen data for evaluation, thus enabling rigorous assessment of model generalization and
predictive accuracy.

Selecting appropriate partitions for dividing data into training and testing sets is critical for the accurate development and
evaluation of time series models. To ensure unbiased evaluation and mimic real-world forecasting scenarios, we opted
for an 80% training and 20% testing split. This division allows the model to learn from the majority of historical data during
training while reserving a smaller portion for assessing its predictive performance on unseen data. By adhering to this
standard practice, we prevent data leakage and ensure that the model's performance is evaluated rigorously on data it
has not encountered during training. The training set, comprising 80% of the data, provides ample historical information
for the model to capture underlying patterns and seasonality dynamics. Subsequently, the model's performance is
assessed on the remaining 20% of the data, enabling us to gauge its ability to generalize and make accurate predictions
beyond the training period. This approach ensures that our models are robust, unbiased, and aligned with real-world
forecasting requirements.

iv. Time Series Model Building

ARIMA Model Building:
 An ARIMA model is fitted using the training data. The auto.arima function is used to automatically select the best
parameters for the ARIMA model based on the Akaike Information Criterion (AIC) or Bayesian Information Criterion
(BIC).
 ARIMA forecasts are generated using the fitted model for the length of the test data.
 A plot is generated to visualize the ARIMA forecast for weekly sales using the winsorized data (Figure 6). The
ARIMA plot for the winsorized sales data reveals a close alignment between the predicted section and the original
sales data, indicating the effectiveness of the ARIMA model in capturing the underlying patterns and trends present
in the sales data. In the plot, the original sales data is depicted alongside the predicted values generated by the
ARIMA model. Visually, the predicted section closely follows the trajectory of the actual sales data, demonstrating
the model's ability to accurately forecast future sales based on historical patterns.
This close resemblance between the predicted and actual sales data suggests that the ARIMA model has successfully
captured the essential characteristics of the sales time series, including any seasonal fluctuations, trends, and
irregularities present in the data.

Figure 6: ARIMA Forecast Plot for Walmart Sales

SARIMA Model Building:

 Similar to the ARIMA model, a SARIMA model is fitted using the training data. The auto.arima function is again used
for automatic parameter selection.
 SARIMA forecasts are generated using the fitted model for the length of the test data.
 A plot is generated to visualize the SARIMA forecast for weekly sales using the winsorized data (Figure 7). The
SARIMA forecat demonstrates the model's ability to capture seasonal patterns, trends, and any remaining
irregularities in the sales data. The predicted section align closely with the original sales data, reflecting the SARIMA
model's capacity to forecast future sales accurately based on historical patterns and the incorporation of
seasonality.
 Figure 7: SARIMA Forecast Plot for Walmart Sales
PROPHET Model Building:
 The prophet library is loaded, and the winsorized data is prepared in the required format for the Prophet model.
 A Prophet model is created using the training data.
 A plot is generated to visualize the Prophet forecast for weekly sales using the winsorized data (Figure 8). The
Prophet modeling exhibits a close resemblance between the predicted and actual sales data. The plot showcase
Prophet's flexibility in handling various data components, including trend changes, holidays, and seasonality,
resulting in accurate forecasts that closely match the observed sales data.

Figure 8: Prophet Forecast Plot for Walmart Sales

Exponential Smoothing Model Building:

 Exponential Smoothing is applied to the training data to capture trend and seasonality components.
  Exponential Smoothing forecasts are generated using the fitted model for the length of the test data.
A plot is generated to visualize the Exponential Smoothing forecast for weekly sales using the winsorized data (Figure 9).
The Exponential Smoothing plot demonstrates the model's ability to capture trends and seasonality while smoothing out
any irregular fluctuations in the sales data. The predicted section closely follow the original sales data, highlighting
Exponential Smoothing's effectiveness in generating reliable forecasts.

Figure 9: Exponential Smoothing Forecast Plot for Walmart Sales

Gaussian Processes Model Building:

 The Gaussian Processes model is trained using the training data.
 Gaussian Processes forecasts are generated using the trained model for the length of the test data.
 A plot is generated to visualize the Gaussian Processes forecast for weekly sales using the winsorized data
(Figure 10). Gaussian Processes modeling plot show a high degree of accuracy in predicting future sales. The
predicted section (in red) would closely resemble the original sales data (in blue), showcasing the model's
capability to capture complex patterns and uncertainties present in the data, thus providing robust and reliable
forecasts.
 Figure 10: Gaussian Process Forecast Plot for Walmart Sales

3. Results and Discussion

3.1. Results
Table 1 compares the accuracy of different time series models in forecasting weekly sales using winsorized data. The
RMSE (Root Mean Squared Error) and MAE (Mean Absolute Error) metrics are used to evaluate the performance of each
model. Lower values indicate better accuracy.

Table 1: Comparing the Accuracy of the Different Time Series Models

Model RMSE (Root Mean Squared Error) MAE (Mean Absolute Error)
ARIMA (Winsorized) 555,502.2 462,767.3
SARIMA (Winsorized) 555,502.2 462,767.3
Prophet (Winsorized) 567,509.2 474,990.8
Exponential Smoothing (Winsorized) 555,081.7 464,110.5
Gaussian Processes (Winsorized) 34,116.09 25,495.72

3.2. Discussion
3.2.1. Model Analysis
i. RMSE (Root Mean Squared Error) :
The ARIMA and SARIMA models have similar RMSE values, both around 555,502.2, indicating their comparable
performance.

The Prophet model has a slightly higher RMSE of 567,509.2 compared to ARIMA and SARIMA, suggesting slightly less
accuracy.

The Exponential Smoothing model has an RMSE of 555,081.7, indicating its performance is similar to ARIMA and
SARIMA.
The Gaussian Processes model has the lowest RMSE of 34,116.09, indicating superior performance compared to the
other models.

ii. MAE (Mean Absolute Error):

Both ARIMA and SARIMA models have identical MAE values of 462,767.3, indicating similar accuracy in predicting the
weekly sales.

The Prophet model has a slightly higher MAE of 474,990.8, compared to ARIMA and SARIMA.

The Exponential Smoothing model has an MAE of 464,110.5, similar to ARIMA and SARIMA.

The Gaussian Processes model has the lowest MAE of 25,495.72, indicating superior accuracy compared to the other
models.

Overall Comparison:
The Gaussian Processes model outperforms all other models in terms of both RMSE and MAE, suggesting it provides
the most accurate forecasts for weekly sales.

ARIMA, SARIMA, and Exponential Smoothing models show comparable performance, with minor differences in RMSE
and MAE.

Prophet performs slightly worse than the ARIMA, SARIMA, and Exponential Smoothing models but still provides
reasonably accurate forecasts.

Based on these metrics, if accuracy is the primary consideration, the Gaussian Processes model would be the preferred
choice for forecasting weekly sales.

Our work distinguishes itself from the studies by [15],[16],[17] through several significant aspects.

First, while [15] focus on aggregate retail sales forecasting using traditional seasonal forecasting methods and neural
networks, [17] compares state space models and ARIMA models for retail sales of women's footwear, and [16] presents
a meta-learning framework using deep convolutional neural networks for retail sales forecasting, our paper specifically
addresses Walmart sales using a Kaggle dataset using a wide range of machine learning models and this dataset spans
a period from February 5, 2010, to October 26, 2012, providing a focused case study on Walmart's sales data over this
timeframe.

Second, in terms of models and methodology, [15] compares linear models like ARIMA with nonlinear models such as
neural networks, emphasizing deseasonalization and seasonal adjustment techniques. [17] focuses on comparing state
space models and ARIMA models, using Akaike's Information Criteria (AIC) for model selection and producing both one-
step and multiple-step forecasts. [16] employs a meta-learning framework with deep convolutional neural networks to
combine multiple base-forecasting methods. In contrast, our paper evaluates a wider range of models, including ARIMA,
SARIMA, Prophet, Exponential Smoothing, and Gaussian Processes. Our approach also includes winsorization to handle
outliers, ensuring robust model performance. Model evaluation is done using RMSE and MAE metrics.

In other recent studies, advanced time-series forecasting models have shown superior performance compared to
traditional methods. For instance [18] compared classical techniques like SARIMA and Triple Exponential Smoothing with
advanced models such as Prophet, Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN) in
forecasting furniture sales. Their findings indicated the superiority of Stacked LSTM, with Prophet and CNN also
performing well [18]. This aligns with our exploration of advanced models like Gaussian Processes and Prophet,
highlighting the potential of these techniques in improving retail sales forecasts."

The novel contributions of our study further differentiate it from the others. [15] finds that nonlinear models outperform
linear ones and stress the importance of seasonal adjustments, while [17] concludes that state space and ARIMA models
have similar forecasting performance and emphasize the effectiveness of using an automatic algorithm for model
selection. [16] recommends a meta-learner combining multiple forecasting methods for improved accuracy, noting the
modest gains over sophisticated benchmarks and the lack of interpretability in learned features. Our paper introduces
new methodologies for retail sales forecasting, demonstrating the benefits of using Gaussian Processes among other
models. It integrates advanced techniques like winsorization to handle outliers effectively, improving model robustness.
Additionally, our detailed comparison of multiple advanced time series models provides insights into their strengths and
weaknesses, offering practical steps for retail businesses to implement more accurate and robust sales forecasting
models.

Our study therefore provides a comprehensive analysis of Walmart sales data using an extensive range of time series
models, including innovative techniques like Gaussian Processes. The detailed comparison of these models and the
inclusion of advanced outlier handling techniques make our research a valuable contribution to the field, offering practical
insights and novel methodologies for improving retail sales forecasting.

The assessment of various forecasting models for weekly sales prediction has shed light on their effectiveness and
relevance in practical contexts. While Gaussian Processes exhibit superior accuracy, it's crucial to acknowledge the
diverse strengths and weaknesses inherent in each methodology. While Gaussian Processes excel in accuracy, traditional
models like ARIMA, SARIMA, and Exponential Smoothing offer comparable performance, potentially providing benefits in
interpretability and computational efficiency. The marginally lower performance of the Prophet model underscores the
importance of refining its parameters for enhanced predictive capabilities.

These findings offer valuable insights into the performance and applicability of different time series models for sales
forecasting. While Gaussian Processes stand out for accuracy, traditional models remain viable alternatives with similar
performance levels. Understanding the strengths, limitations, and practical implications of each model is pivotal for
selecting the most suitable approach based on specific forecasting needs and organizational constraints.

iii. Analysis of Implications and Practical Considerations

The results obtained from the comparison of different time series models have several implications, including strengths,
limitations, and practical implications, which enrich the analysis:

 Strengths:
Gaussian Processes Superiority:
The standout performance of the Gaussian Processes model, as evidenced by its significantly lower RMSE and MAE
values, highlights its superiority in accurately forecasting weekly sales. This suggests that Gaussian Processes are well-
suited for capturing the complex patterns and dynamics present in the Walmart sales data.

Comparable Performance of Traditional Models:

Despite the superior performance of Gaussian Processes, traditional models such as ARIMA, SARIMA, and Exponential
Smoothing demonstrate comparable accuracy in forecasting weekly sales. This underscores the reliability and robustness
of these established methodologies in capturing temporal dependencies and seasonal variations in retail sales data.

 Limitations:
Computational Complexity: While Gaussian Processes offer superior accuracy, they may come with higher
computational costs and resource requirements compared to traditional models like ARIMA and SARIMA. The
implementation of Gaussian Processes may pose challenges in scenarios where computational resources are limited or
where real-time forecasting is necessary.

Interpretability: Although traditional models like ARIMA and SARIMA are well-understood and interpretable, Gaussian
Processes may lack interpretability due to their complex nature. Understanding the underlying mechanisms driving the
forecasts generated by Gaussian Processes may be challenging, limiting the insights that can be derived from the model
outputs.

 Practical Implications:
Decision-Making Support:
The accurate forecasts provided by the Gaussian Processes model can serve as valuable inputs for decision-making
processes within Walmart and other retail organizations. By leveraging these forecasts, decision-makers can optimize
inventory management, resource allocation, and marketing strategies to meet consumer demand more effectively and
enhance overall business performance.

Resource Allocation:
The comparable performance of traditional models like ARIMA, SARIMA, and Exponential Smoothing suggests that these
models remain viable options for forecasting weekly sales, particularly in scenarios where computational resources are
limited or where interpretability is a priority. Organizations can allocate resources judiciously based on the specific
requirements and constraints of their forecasting tasks.

Continuous Improvement:
The slightly lower accuracy of the Prophet model compared to traditional models highlights the importance of continuous
model refinement and parameter tuning. By iteratively improving model performance and incorporating domain knowledge
and expertise, organizations can enhance the accuracy and reliability of their sales forecasting systems over time.
In conclusion, the results of the analysis provide valuable insights into the performance and suitability of different time
series models for forecasting weekly sales. While Gaussian Processes emerge as the top performer in terms of accuracy,
traditional models like ARIMA, SARIMA, and Exponential Smoothing remain viable options with comparable performance.
Understanding the strengths, limitations, and practical implications of each model is essential for selecting the most
appropriate approach based on specific forecasting requirements and organizational constraints.

4. Conclusion and Recomendation:

4.1. Conclusion
In conclusion, our study offers a comprehensive analysis of Walmart sales data utilizing a diverse range of time series
models. Through a focused examination of weekly sales trends from February 5, 2010, to October 26, 2012, we aimed to
enhance the accuracy and robustness of sales forecasting methodologies.
Our work distinguishes itself from previous studies by [15], [16], and [17] in several significant aspects. While these studies
focused on various aspects of retail sales forecasting, our research specifically addressed Walmart's sales data using a
Kaggle dataset, providing a unique case study over a defined timeframe. We expanded the scope by evaluating a wider
range of models, including ARIMA, SARIMA, Prophet, Exponential Smoothing, and Gaussian Processes. Additionally, we
introduced innovative outlier handling techniques such as winsorization to improve model robustness.
The comparative analysis of these models revealed nuanced insights into their performance and applicability. While
Gaussian Processes exhibited superior accuracy, traditional models like ARIMA, SARIMA, and Exponential Smoothing
demonstrated comparable performance levels. This suggests that while advanced techniques offer enhanced predictive
capabilities, traditional methods remain viable alternatives with distinct advantages in interpretability and computational
efficiency.
Our findings underscore the importance of understanding the strengths, limitations, and practical implications of each
forecasting model. By providing practical insights and novel methodologies for improving retail sales forecasting, our study
contributes to the advancement of predictive modeling in the retail sector.

4.2. Recommendations and Future Perspective:

The evaluation of multiple forecasting models for predicting weekly sales reveals valuable insights into their performance
and potential areas for improvement. Moving forward, it's essential to leverage these findings to enhance forecasting
methodologies and address existing challenges.

Firstly, the Gaussian Processes model emerged as the most accurate predictor, demonstrating the potential for precise
weekly sales forecasts. Further exploration and refinement of Gaussian Processes modeling techniques could lead to
even more accurate predictions, making it a promising avenue for future research.

However, it is crucial to acknowledge the complexities associated with different forecasting methodologies. While
Gaussian Processes offer superior accuracy, traditional models like ARIMA, SARIMA, and Exponential Smoothing may
provide advantages in terms of interpretability and computational efficiency. Future research should aim to elucidate the
trade-offs between these factors, aiding in the selection of the most suitable approach for specific forecasting
requirements.

Moreover, the slightly inferior performance of the Prophet model compared to traditional time series models highlights the
need for further refinement and customization. Adjusting Prophet's parameters could enhance its predictive capabilities
for weekly sales forecasting tasks, warranting attention in future research endeavors.

Moving forward, future research in the area of weekly sales forecasting should focus on refining existing methodologies,
exploring novel modeling techniques, and elucidating the trade-offs between accuracy, interpretability, and computational
resources. By doing so, researchers and practitioners can develop robust forecasting frameworks tailored to the unique
characteristics and requirements of sales forecasting tasks, thereby facilitating better decision-making and resource
allocation in retail and related industries.
5. Acknowledgement
We acknowledge the use of generative AI technology in the preparation of this manuscript. Specifically, OpenAI's ChatGPT
(version GPT-4) was utilized to assist in drafting and refining the text. The use of this AI tool included providing language
suggestions and ensuring coherence and clarity throughout the manuscript. We affirm that the final content has been
thoroughly reviewed and approved, ensuring its accuracy and adherence to the research objectives.

COMPETING INTERESTS DISCLAIMER:

Authors have declared that they have no known competing financial interests OR non-financial interests OR personal
relationships that could have appeared to influence the work reported in this paper.

References:

[1] James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning. Springer.

[2] Box, G. E., Jenkins, G. M., & Reinsel, G. C. (2015). Time series analysis: Forecasting and control. John Wiley & Sons.

[3] Kaggle. (2024). Walmart sales data. Retrieved from https://www.kaggle.com/code/yasserh/walmart-sales-prediction-

best-ml-algorithms/input

[4] Hyndman, R. J., & Athanasopoulos, G. (2018). Forecasting: principles and practice. OTexts.

[5] Cyril Neba C., Gerard Shu F., Gillian Nsuh, Philip Amouda A., Adrian Neba F., Aderonke Adebisi, P. Kibet, & F. Webnda.
(2023). Time Series Analysis and Forecasting of COVID-19 Trends in Coffee County, Tennessee, United States.
International Journal of Innovative Science and Research Technology (IJISRT), 8(9), 2358-2371. Retrieved from
https://www.ijisrt.com. DOI: 10.5281/zenodo.10007394.

[6] Shumway, R. H., & Stoffer, D. S. (2017). Time series analysis and its applications: With R examples. Springer.

[7] Nicolau, J. L. (2019). Tourism demand forecasting: A review of empirical research. Annals of Tourism Research, 78,
107833.

[8] Rogers, J., & Makridakis, S. (1999). The leading indicators of turning points. International Journal of Forecasting,
15(4), 385-401.

[9] Armstrong, J. S., & Collopy, F. (2001). Principles of forecasting: A handbook for researchers and practitioners. Springer
Science & Business Media.

[10] Chatfield, C. (2000). Time series forecasting. CRC Press.

[11] Wan, C., & Wang, F. (2018). Effects of sales forecasts on marketing mix strategy in the industrial sector. Industrial
Marketing Management, 74, 197-207.

[12] Waters, C. D. J. (2009). Principles of marketing. Routledge.

[13] Cox, D. R. (2012). Principles of statistical inference. Cambridge University Press.

[14] Chase, C. W., Jacobs, F. R., & Aquilano, N. J. (2019). Operations management for competitive advantage. McGraw-
Hill Education.
[15] Chu, C. W., & Zhang, G. P. (2003). A comparative study of linear and nonlinear models for aggregate retail sales
forecasting. International Journal of Production Economics, 86(3), 217-231. https://doi.org/10.1016/S0925-
5273(03)00068-9.

[16] Ma, S., & Fildes, R. (2021). Retail sales forecasting with meta-learning. European Journal of Operational Research,
288(1), 111-128. https://doi.org/10.1016/j.ejor.2020.05.038.

[17] Ramos, P., Santos, N., & Rebelo, R. (2015). Performance of state space and ARIMA models for consumer retail sales
forecasting. Robotics and Computer-Integrated Manufacturing, 34, 151-163.
https://doi.org/10.1016/j.rcim.2014.12.015.

[18] Ensafi, Y., Hassanzadeh Amin, S., Zhang, G., & Shah, B. (2022). Time-series forecasting of seasonal items sales
using machine learning – A comparative analysis. International Journal of Information Management Data Insights,
2(1), 100058. https://doi.org/10.1016/j.jjimei.2022.100058