Changes in shopping habits during COVID-19

Abstract

By the start of 2020, the daily and business world had to undergo a radical change with the widespread pandemic known as COVID-19. Many people had to replace their everyday purchase medium to meet the enforced restrictions, and local businesses had to adjust their operations to accommodate the negative impacts brought upon by the disease’s rapid spread. Groceries and FMCG sub-sectors of the retail industry were forced to adapt to consumers’ stockpiling and panic-buying behaviors. We studied the impact of similar purchase attitudes for various product groups during the COVID-19 and probed the differences between sales of online and physical markets. Initially, a cluster analysis identifies which product groups were affected by similar shopping behaviors during the pandemic. Subsequently, the impact of the number of COVID cases on sales levels was measured using stepwise, lasso, and the best subset models. All the models were applied to both physical and online market datasets. The results showed a significant shift from the physical to the online markets during the pandemic. These findings can provide an essential guideline to retail managers in adapting to the new world.

Introduction

On the 31st of December 2019, the Wuhan local government warned the World Health Organization (WHO) that the cause of an outbreak of a new virus in Hubei remained unclear. COVID-19 has influenced every aspect of daily life. Masks have become necessary as people learned to be more conscientious about personal hygiene. Rules and regulations, such as curfews and lockdowns, travel restrictions, and social distancing measures, were implemented in various countries to help the population cope with the pandemic in the most effective way possible. Since the first news of the outbreak in China, world economies have been impacted. Thus, it obligated companies to identify the potential dangers that might arise if they cannot run their supply chains during the crisis (Saenz et al. 2020). As a result, businesses began to concentrate on making their supply chain distribution more resilient than ever before.

The pandemic affected both the supply and demand sides of all industries. The retail industry was no exception. It was one of the most affected sectors (TÜİK. 2023). Limitations imposed by governments led to panic buying of consumers, which left the market mainly unprepared. Due to panic buying and the surge of delivery issues on the manufacturers’ end, store shelves remained empty for several days (Hall et al. 2020). During this period, the preference for online shopping grew. Although most markets already had an infrastructure for online shopping, supply and logistics issues rendered the capacity of both the online and physical stores insufficient to meet rapidly increasing demand. Especially, the online channels of many supermarket chains require 3 to 4 days of the delivery time to deliver the customers’ orders (Islam et al. 2021).

To better understand the impact of the pandemic during this period, this research aims to analyze the effect of COVID-19 on the consumer behavior in the retail industry. For this purpose, this study focuses on answering the following questions:

1. Which product groups were affected by similar purchase attitudes during the COVID-19 pandemic?

2. Compared to 2019, which product groups were affected by a significant change in shopping behavior during the pandemic?

3. Are there any differences between the online market and physical store sales levels during the pandemic and in 2019? Is there a shift from physical to online markets?

This case study investigates the sales data of Migros, one of Turkey’s largest supermarket chains, for the 2019–2020. The supermarket chain provides its service with a total of 2,370 stores spread over 81 provinces in Turkey and holds over 2000 varieties of product groups and brands. In addition to perishable and food-related groups, glassware, textile, toys, and electronic products are also provided. According to official figures for the year 2020, the retailer’s sales volume increased by 24% YoY.

The remainder of this paper is organized as follows. The “Literature review” provides a literature review based on the effect of the pandemic in general, its impact on supply chain disruption and consumer behavior, and the research conducted to analyze the effects of COVID-19, in particular on consumer shopping behavior, and the contribution of the paper is underlined. The “Proposed methodology” provides the proposed methodology. “Discussion of the results” presents the analysis of the Turkish retailer as a case study, and the “Discussion of the results” discusses the findings. Finally, the “Conclusion and further suggestions” is based on conclusions and further suggestions.

Literature review

The effect of pandemic in general

The Justinian Plague of the 6th Century and the Bubonic plague of the 14th Century were the first global pandemics that spread through trade routes. Although these outbreaks caused substantial trade disruption, it was impossible to observe the supply chain and consumption relations during medieval times and assess their effects on the economy and trade-related industries. On the other hand, the first global health crisis observed in the modern age was the Spanish Flu of 1918–1920. However, during this pandemic, the supply chain and consumption side were already heavily distressed due to World War I.

The effect of flu outbreaks from 1950 to 2008 remained confined to Asia. However, in 2009, the Avian Flu (H1N1) pandemic became the first global outbreak spread across different continents. More importantly, this outbreak was the first pandemic that directly impacted the global economy and was interconnected to the information age. Early studies on the potential macroeconomic effects of the worldwide pandemic were published a few years before the start of the H1N1 pandemic. In a working paper submitted to the European Commission (Jonung and Roeger 2006), an empirical model-based approach was utilized to simulate the effects of a possible pandemic on the European economy. Under the assumption of an epidemic hitting Europe in 2006, the authors suggested that a disturbance in supply due to social distancing would cause a 1.6% contraction in the European economy. Moreover, Buetre et al. (Buetre et al. 2006) examined the same situation for Australia and estimated that its regional economy would curtail by 6.8% in the wake of the Avian Flu endemic. Finally, (Keogh-Brown et al. 2010) probed the effect of a major epidemic scenario using the UK Economic and Social Research Council’s quarterly Macro-Economic Model and estimated at least a 2.52% and at most a 6.05% decrease in UK’s GDP with consumption impact.

In addition to these scenario-based simulation studies, a new research stream that blends data-driven studies with epidemiology disciplines came into existence in the early 2000s. One of the topics explored by the researchers in this discipline is assessing the impact of the health crisis while employing internet search data. The first studies based on the search trends came from the community healthcare field. Eysenbach (2006) suggested a relationship between the 2004 flu epidemics in Canada and Google searches well before the “Google Trends” service was launched and became publicly available. Consequently, Ginsberg et al. (2009) became the first publication in the area of info-demiology to employ Google Trends data to examine epidemic trends.

The effect of pandemic on supply chain disruption

Understanding the changes in shopping habits and consumer behavior requires us to investigate the initial disruption caused by the pandemic that occurred in the supply chain. Disturbances in the supply chain, which start with the producer and end with the consumer, also affect the final step in the process, which is the customer purchasing and consuming the product. COVID-19 highlighted the inefficiency of all industries in reacting to large-scale disruptions promptly. The importance of resilience was highlighted as various industries still build more adaptability towards unanticipated situations, such as the Covid-19 pandemic. Resilience is defined as the ability to resist disruptions and recover performance. Supply chains mostly affected by the lack of resilience mainly include the life sciences, health care, and food industries (Simchi-Levi and Simchi-Levi 2020). COVID-19 also led enterprises to transition toward more sustainable supply chains. Most companies in these industries needed to implement new sustainability strategies, such as building community trust, faster than they had expected. In return, including these strategies could aid companies in building their supply chain resilience. Risk response and crisis management techniques allow companies to transform themselves through sustainability by helping them reduce risks and strengthen resilience (Sarkis 2020).

In fact, during the crisis period, supply chains were divided into two groups. Some were faced with an extreme demand they could not meet, while others experienced a severe decrease in demand and supply, prompting them to halt production. Many companies dealt with the danger of bankruptcy and did not get any governmental support during this time (Ivanov 2020). The food industry was among the sectors that flourished the most with the sudden increase in demand. Most consumer goods were no longer as easily accessible as before. Demand for food continues to increase even if people have sufficient supplies because of the extreme financial conditions. If supply chains continue to be disrupted, further shortages in the food industry will be expected (Sarkis et al. 2020). Supply chains are susceptible to disruptions caused by disasters. Depending on the level of preparedness, the effect of disturbances may even last for several cycles.

Along with disruptions caused by disasters, stockpiling creates a more complex environment for inventory management. When restocking becomes a problem and companies cannot meet the demand, people become more eager to look for substitute products. This behavior affects the whole structure and creates double-sided problems for both the retailer and the consumer. Several manifestations of this behavior were observed in the UK and Australia at the beginning of the pandemic. For example, when people could not find toilet paper, they bought baby nappies and kitchen towels even though they are not usually considered substitutes for one another (Chen et al. 2020).

The effect of pandemic on consumer behavior

The commodity theory claims that scarcity may explain stockpiling behavior (Brock 1968), which arises as an individual’s response to stress, fear, and a panic-inducing environment. People started to purchase or order more than their average consumption and stocked them in cases of an emergency (Micalizzi et al. 2021). According to the prospect theory, risk aversion motivates this behavior if food sources are perceived to be scarce by the consumer, even though the possibility of this happening is very low (Tversky and Kahneman 1992). As a result, this behavior cannot be considered entirely irrational because it reflects human nature. However, stockpiling has serious adverse effects on the economy and society. As a social behavior, it negatively affects supply chains by disrupting them and creating shortages (Micalizzi et al. 2021).

Apart from disruptions in the supply chain, panic buying also affects social life by creating a chaotic environment and a scarcity of goods for others. This condition applies mainly to the elderly, disabled, and working-class people since they cannot shop at their convenience and are much more affected by the chaos caused by panic buying. When a product becomes scarce, its value also starts to increase. Another adverse effect of stockpiling is that it creates a competitive environment that leads to price increases (Chen et al. 2020). Empty shelves, intermittent restocking problems, and the chaos from panic buying led people to shop online. Most retailers were not ready to respond to the increased demand in online markets, which resulted in delayed deliveries that lasted for 3–4 days. Late deliveries in online markets, empty shelves in local stores, and increasing prices led people to spend more money during this period. Consumers who could not find the product they were looking for had to go farther away to see what they were seeking (Chen et al. 2020). One of the recent studies in this field also examined the changing purchasing behaviors in the United States during the COVID-19 pandemic. The authors used data gathered from a survey conducted in 2021 and utilized logistics regression to derive their results. They highlighted that the concerns about health and financial issues influence the change in shopping behavior (Truong and Truong 2022).

Supply chains are looking for more permanent solutions for these impending threats. On the other hand, governments are working on preventing panic-buying behavior in case of future disasters. Online market options are continuously increasing. Many retailers started to launch their online applications, which allow consumers to place their orders more conveniently. They strive to provide more product options than physical stores by establishing warehouses dedicated solely to online orders and circumventing the shelf capacity constraints. In reality, online shopping, especially grocery shopping, was a broadly discussed topic among researchers even before the pandemic. Many papers on this subject have focused on its driving factors and changing trends (Chen et al. 2020). Retailers should focus on changing consumer behavior and the customer experience and adoption process (Hand et al. 2009). In the online adoption of the consumers, there is also the issue of consumer perception and behavior concerning single versus multi-brand retailers (Rahnamaee and Berger 2013).

Several studies have shown that people’s adoption rate of online grocery shopping is deliberate and slow. While online shopping consumers make decisions based on their habits (Frank and Peschel 2020), the trust and recommendations from friends and family play a significant role in their shopping and adoption decisions. (Pauzi et al. 2017) and (Hand et al. 2009) showed that lifestyle changes, such as relocating to a new house, having children or newborns, taking care of the elderly at home, working late, and changing jobs, are driving factors behind online shopping (Hand et al. 2009). However, one study comparing several countries showed that Turkish citizens find online shopping less satisfying, even if the internet makes purchasing consumer goods effortless and cheaper. This is mainly a result of Turkey’s insufficient technology infrastructure (Turan 2012). Another reason why Turkish citizens find online grocery shopping less satisfactory is that those who opt for shopping from local stores are usually skeptical about product quality because it is not accurately reflected in online markets (Muhammad et al. 2016). Household Information and Communication Usage Survey (ICT) of 2021 (Turkish Statistical Institute 2021) shows that computer and internet usage in Turkey is at 61.7% and 48.1%, respectively, for those aged between 25–34 and 35–44. Overall, 92% of homes have access to the internet, and 96% of individuals own cell phones. However, the same survey points out that the percentage of individuals who do online shopping is only 44.3%. All those types of research related to the above-given consumer retail behavior changes during crisis periods such as the pandemic necessitates appropriate data collection, integration and aggregation and the use of suitable business analytics models (Petrescu and Krishen 2020).

The effect of COVID-19 on shopping behavior

Although the COVID-19 pandemic has well affected many aspects of the majority of people, the academic research analyzing its impact on consumer behavior and shopping habits are tenuous. In March 2020, shopping habits and consumer behavior began to change. The retail industry experienced a shock in the online market and the orders increased. As the number of infected cases rose, people avoided shopping in person, which resulted in a rapid increase in online orders, compelling retailers to focus more on their online services. As this was the case for past natural disasters and crises, panic buying, stock-pilling, and hoarding became rampant worldwide. Shelves of local stores were emptied out by people stockpiling their daily used and necessary items, such as food, medicine, and toilet paper (Chen et al. 2020), (Park et al. 2020). Soares et al. (2022) analyze the impact of COVID-19 on online shopping behavior of 1052 Brazilian online consumers. They used PLS-SEM and revealed that the perceived risk of being infected by the pandemic making in-person shopping positively influenced the perceived usefulness and ease of online purchases. However, the proposed model is solely based on data collected during winter where important mobility restrictions influence the online buying decisions and should be repeated to see the purchasing behavior changes that will occur at different stages of the pandemic.

Szymkowiak et al. (2021) adopted structural equation modeling on answers to 914 questions to analyze the relation between the risk of in-store infection and consumer behavior. They showed that customers had decreased pleasure during shopping due to the risk. In addition, their approach measured how the pandemic influenced the behavioral response of customers within a store and provided recommendations for stationary retailers to become more resilient. However, since the study was conducted during the pandemic, the results could not be compared with pre-COVID periods.

(Chiu et al. 2022) also emphasized the change in consumer behavior during the pandemic. They collected 608 responses from USA consumers and employed a partial least square structural equation model on the data. They concluded that COVID-19 increased the consumers’ fear and, in turn, led to a surge in the purchase of online fitness products. However, based on their results, the income level of the customers had a negative influence on the relationship between fear and online purchase of the products, as mentioned earlier. Their study provides a guideline for sports retailers in preparing strategies for different customer segments. However, as was the case with Szymkowiak et al. (2021) paper, this study could not compare its results with the pre-COVID period. In addition, the analysis does not provide insight into changes in the consumer behavior of other frequently purchased products, such as those sold in the superstores.

Boyle et al. (2022) analyzed the impact of COVID-19 on grocery shopper behavior. They used transactional data and market basket analysis with different indicators to shed light on consumption and prioritization changes. However, their focus on a single store limits the representativeness of the study’s results.

Sohn et al. (2022) studied consumer purchasing behavior, specifically turning their attention toward purchasing organic food. Their collected data includes periods before the COVID-19 and during the first wave of the pandemic from 429 German consumers. They analyzed the data using structural equation modeling and deduced that although the pandemic increased consumers’ quality and health consciousness and consequently motivated them to buy organic products, it did not influence their environmental consciousness. They also demonstrated that the results portray differences between customer characteristics, such as income, age, and education. However, the scope of this study is limited to organic products, and it only considers and analyzes the information from the first wave of the pandemic.

Table 1 provides a summary of a comparison of this paper’s proposed methodology with the current literature.

Table 1.

Comparison of our proposed methodology with the literature

Author	Aim of the paper	Data/methods	Results	Differences with this paper
Soares et al. (2022)	Impact of COVID-19 on online shopping behavior	1052 online consumers PLS-SEM	Pandemic has positive impact on online shopping	(-)Based on data collected during winter
Szymkowiak et al. (2021)	The relation between the risk of in-store infection and consumer behavior	914 questions PLS-SEM	Customers had decreased pleasure during shopping due to the risk	(-)Based on data only during the pandemic
Chiu et al. (2022)	Change in consumer behavior during the pandemic	608 responses PLS-SEM	COVID-19 increased the consumers’ fear and, in turn, led to a surge in the purchase of online fitness products	(-)Could not compare its results with the pre-COVID period (-)Only online fitness products
Boyle et al. (2022)	The impact of COVID-19 on grocery shopper behavior	Transactional data Market Basket Analysis	Consumption and prioritization change	(-)Focus on a single store
Sohn et al. (2022)	Consumer purchasing behavior	Periods prior and during the first wave of the pandemic 429 German consumers PLS-SEM	Pandemic increased consumers’ health consciousness and motivated them to buy organic products	(-)Limited to organic products (-)First wave of the pandemic

The differences are shown with “-”

As can be seen from the table, the research which analyzes the impact of COVID-19 on shopping behavior are either limited to a specific type of products or investigates only the COVID-19 period and, thus, does not provide a comparison concerning the pre-COVID period.

The primary contribution of this paper is to analyze the sales data for both physical and online stores of one of the biggest supermarkets in Turkey and to evaluate the changing consumer behavior of customers in Turkey during the COVID-19 pandemic by comparing pre-COVID period. As a case study, the data of one of the biggest supermarket chains in Turkey are gathered, cleansed, and analyzed using a CRISP-DM methodology.

Proposed methodology

This research investigates changes that occurred in different product groups’ sales before and during COVID-19, based on data from a Turkish retailer. The dataset contains the sales levels of the product categories from January 2019 to December 2020. To explore the proposed research questions—which product groups were affected by similar purchase attitudes, which affects significantly changed shopping behavior, and are there any differences in sales levels between the online and physical stores during the pandemic in 2019—we utilized the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology as the main approach (Shearer 2000). The steps in this methodology are business understanding, data understanding, data preparation, modeling, evaluation of the models, and deployment (Provost and Fawcett 2013).

To answer the questions given in the introduction, initially, a business understanding and data cleaning activities are realized based on several meetings with the executives of the supermarket chain, subsequently, a hierarchical cluster analysis is conducted to specify which groups of products were showing similar shopping behavior. At the next stage, for each cluster, Stepwise, Lasso and the Best-Fit regression models are conducted, and the best-performing method to estimate the sale levels for each product group in both the physical and online market is selected. Finally, we can derive the transitions in shopping behavior in terms of product groups and study the level at which their attitudes shifted. By examining the impact of COVID-19 on these changes, we provide suggestions on policies that can be adopted to help retailers better adapt to the new world. Figure 1 provides the flowchart of the methodology.

Fig. 1.

Flowchart of the methodology

Consumer data

The discussed methodology is applied to one of the largest supermarket chains in Turkey. Datasets used for the case study contain the daily transactional records with unit prices and quantities sold for all product groups from 01 January 2019 to the end of 31 December 2020, with more than 5 million rows for each month. Both online transaction and physical transaction data with the same timeline were used to see the difference between the physical market and the online market. Since unit prices change regardless of the spread of COVID-19, this research focuses on the changes in the sold quantities. Covid case numbers starting from March 2020 to the end of December 2020, since the first Covid-19 case was first seen in March in Turkey, were also used in the following analysis to examine the relationship between sales and case numbers.

Analysis

Business understanding and data cleaning

As seen in Fig. 3, curfews, several restrictions, and procedures were implemented as of April 2020, which resulted in a substantial paradigm change in the Turkish Retail Sector. Some of the restrictions in the workplace for both employees and customers were mandatory use of masks, social distancing, a limited amount of person per store, and regular disinfection of stores (Ministry of Health 2021; Wikimedia Foundation 2022).

Fig. 3.

Dendrogram results for the products showing similar sales patterns

Survey-based retail sector indicators, such as Consumer Confidence and Retail Trade Confidence, contracted in the first month of the pandemic, and both of them dipped in April. However, they both recovered quickly in the first months of the summer of 2020. As can be seen from Fig. 2, the negative effect of acceleration in the trend of inflation in 2021 was more evident.

Fig. 2.

Retail indicators of turkey

Initially, we filtered out the unrelated product groups to focus on the essential consumable goods for both online and physical market datasets. For this purpose, nine meetings were conducted with the data provider retail company’s Business Intelligence (BI) Team. After cleaning the data, 30 out of 72 main product groups were kept; however, since selling alcoholic beverages and tobacco products through online channels is illegal in Turkey, these two groups were also excluded from our analysis, and only the remaining 28 groups were used in the final comparison.

Cluster analysis results

For the second stage of our methodology, we perform Hierarchical Clustering to cluster the product groups according to the similarities between their individual sales trends from 2019 to 2020. The clustering was done in SPSS Statistics software. For each of the six stores’ formats, the monthly YoY change in sales from 2019 to 2020 is revealed for the 30 main groups.

The results from cluster analysis using a complete linkage with a threshold of 10 yielded six clusters where two of them consist of only one group of products. Accordingly, the retailer’s BI team also validated these findings by pointing out that these two categories follow their characteristic trend apart from other product groups in the retailer’s inventory.

The cluster analysis results are depicted in Fig. 3.

Each cluster name represents the product groups in each cluster (Table 2).

Table 2.

Cluster results based on the similarities of sales patterns

Alcoholic beverages	Daily needs	Protein-based products	Basic consumption goods	Cold-chain products	Electronics
Alcoholic beverages	Seasonal products	Meat-Deli	Chips and snacks	Meat	Electronics
	Glassware	Olive oil and butters	Dairy products	Packaged meat
	Textile	Cheese	Imported fruit	Produce (vegetables and fruits)
	Tobacco	House care products	Eggs	Frozen goods
	Cookies, chocolate, and candy	Poultry	Canned goods, beverages, and breakfast food products
	Bread and baked goods	Fish and seafood	Grains, pasta, and sides
	Non-alcoholic beverages		Paper and baby products
	Take outs		Cleaning products
	Toys, pet care, and media		Beauty products

For the physical market, there was a significant decrease in April 2020 (Panel A of Fig. 4). Soon after the number of cases per day exceeded 5000 on the 10th of April 2020, the Ministry of Interior Affairs started to impose curfews. As a result, people began to spend more time at home, leading them to prepare their meals themselves. This habit resulted in a decrease in sales levels for this cluster of products in physical stores. However, the decrease in physical channels led to increased online sales. While many people started to cook at home, others preferred to place orders for their daily needs through online channels.

Fig. 4.

Cluster’s percentage change compared to 2019

Items in this Protein-Based Products cluster (portrayed in Panel B of Fig. 4) mainly consist of fish, meat, and dairy products. For products in the protein-based cluster, it can be seen that sales levels were increased due to the imposition of curfews.

Grains, pasta, sides, paper and baby products, and cleaning supplies were among the groups in the Basic Consumption Goods cluster that had the most purchase and consumption due to the spread of COVID-19, which led to panic buying. As a result, many retailers needed help to meet the demand for these products during this period (Panel C of Fig. 4). Instead of the product groups mentioned above, online market consumers mainly preferred “Chips and Snacks” during this period.

The sale of cold-chain products (Panel D of Fig. 4) stayed at normal levels from March 2020 to early June 2020, when the normalization period began in Turkey. During this period, sales levels decreased, suggesting that people started to cook less at home and dine outside more. However, when curfews were again imposed in September, people began to spend more on these products. The online market also showed the same trends as the physical market. For the online market, clustering indicated that shopping behavior related to this product group was similar to that of the physical market. That is why the same cluster groups are also used for the online market.

For physical and online markets, a graph was created for each cluster to show different shopping habits from 2019 to 2020 (Fig. 4).

Regression analysis results

For the third stage, Stepwise, Lasso and Best Subset Regression models with sales values aggregated at the described cluster levels were used to evaluate the effect of the COVID-19 pandemic on the sales levels of each cluster. Using time series models, we aimed to analyze whether the number of COVID-19 cases affected the sales levels of a specific cluster group and, if so, how strong its impact is and for which clusters. The variables found that could explain the sales levels in the regression analysis are given in Table 3.

$$\text{Ln}{({\text{Sales}}_j)}_t={\beta }_0+{\beta }_1\text{Ln}{(\text{Covid}\text{case}\text{number})}_t+\sum _{i=1}^7{\alpha }_{i}Ln{\left({\text{Sales}}_j\right)}_{t-i}+\sum _{k=1}{\delta }_k{Z}_{\mathit{kt}}+{ϵ}_{\mathit{jt}}.$$ 1

Table 3.

Variables that have been used in regression analysis

DATE	Starting from March 2020 until the end of December 2020
LNCASENUM	Natural logarithm of case numbers
LNCLUSTER(A)	Natural logarithm of total sales numbers for each cluster
LNCLUSTER(A_1-7)	Values of each cluster (each of which were lagged 1 to 7 times in 7 different columns)
DUM24_5	Dummy variable for the 24th of May because there was a curfew
DUM1_6	Dummy variable for the 1st of June because it was the start of the normalization period

As can be seen from Table 3, a logarithmic transformation was applied for all the variables since the numbers of infected cases and product sale levels were not distributed normally. The model also included dummy variables for the 24th, 25th, and 26th of May and the 1st of June. These variables were added since a significant change in sales levels occurred in those days. To capture the dynamic interaction effects, in addition to level dummies, the interaction of LnCluster(A) and dummy variables representing unusual sales behavior were multiplied in the following columns to determine the difference on that day compared to 1–7 days prior.

This process was applied to both the online and physical market datasets. For the Stepwise regression, the “olsrr” (Hebbali 2020), and for the Lasso and Best subset models, “glmnet” (Friedman et al. 2010) and “leaps” (Lumley 2020) packages offered in the R-Studio were utilized. The best estimation model was selected based on Cp, Adjusted R² values, F, and t values, as well as the tenfold cross-validation test error rate (Gareth et al. 2013). The interpretation of these findings was used to propose a road map for the retailer to revise its business strategy in the face of the changing shopping behavior of consumers.

Physical market shopping regression analysis results for each cluster

As was mentioned before, three different regression models—Stepwise, Lasso, and Best subset—were used for each of the six clusters for both the physical and online market. Among these models, we selected the best-fitted model according to its adjusted R², C_p, and using tenfold cross-validation for each cluster. The results are summarized in Table 4.

Table 4.

Physical market regression results for each cluster

Cluster	R-squared			Cp			Cv test error
	Stepwise	Lasso	Best subset	Stepwise	Lasso	Best subset	Stepwise	Lasso	Best subset
Alcoholic beverages	0.825	0.824	0.826	11	10	12	2.317	2.108	0.944
Daily needs	0.728	0.713	0.735	10	6	8.31	0.215	0.19	0.142
Protein-based products	0.772	0.738	0.773	10	6	8.31	0.215	0.19	0.142
Basic consumption goods	0.753	0.726	0.753	10	7	8.24	0.184	0.167	0.128
Cold-chain products	0.764	0.729	0.764	10	7	8.27	0.207	0.179	0.135
Electronics	0.666	0.664	0.673	7	7	12	0.199	0.162	0.165

As can be seen from Table 4, the model with the lowest test error rate is the best subset selection for all clusters except Electronics. The R-squared value is greater than the other models, suggesting a better explanation of variation in the dataset. Furthermore, the CV test error value is smaller, and the R-squared and C_p values are higher for this model than the other, making the best-performing model the best subset selection model.

For the Electronics product group, the lowest test error rate is given by the Lasso regression model. However, its R-squared value is also slightly lower than the other models. When we compare the three models’ t values to the Lasso’s, the outcome was more statistically significant than the others. As a result, the Lasso model was chosen as the final model for the Electronics product group because having a significantly lower value of CV test error rate than other models.

Online market shopping regression analysis results for each cluster

The findings for online shopping are provided in the same form as used for the physical market. More specifically, the developed regression models and evaluation criteria are identical for the online market. However, since there is no data for “Alcoholic Beverages” in the online channels, it was excluded from the results.

The Best subset model has the highest R-squared, lowest CV test error, and highest C_p value among other models for each cluster, thus, making the best-performing model for each cluster the Best subset model. Table 5 summarizes the results.

Table 5.

Online market regression results for each cluster

Cluster	R-Squared			Cp			Cv test error
	Stepwise	Lasso	Best subset	Stepwise	Lasso	Best subset	Stepwise	Lasso	Best subset
Daily needs	0.886	0.883	0.886	10	7	12	0.495	0.264	0.171
Protein-based products	0.886	0.844	0.867	10	7	12	0.440	0.251	0.188
Basic consumption goods	0.870	0.776	0.870	10	7	10.04	0.608	0.358	0.210
Cold-chain products	0.870	0.865	0.882	10	7	10.30	0.410	0.225	0.177
Electronics	0.656	0.654	0.656	8	10	10	0.489	0.540	0.409

Discussion of the results

The main aim of the research was to compare the impact of the COVID-19 pandemic on the sales levels of different product groups in both online and physical markets. The coefficient of the variable representing the number of infected cases mainly tells how a 1% change in the number of COVID-19 cases affects the sales levels as a percentage. By looking at Table 6, it can be inferred that a 1% increase in the number of cases leads to a 0.38% decrease in sales in the long run. Subsequently, if the number of cases increases by 10%, the sales of the relevant category will decrease by approximately 4% in the long run. LnClust4_1 to LnClust4_7 represent the lagged days with respect to the original day sales. However, besides the example cluster provided in Table 6, none of the other clusters, physical and online, showed any significance of time effect on sales. That is why the further analysis was only continued considering the COVID case numbers. Table 6 provides an example of the final model results of how the research examined the effect of the COVID-19 pandemic on sales.

Table 6.

Example of regression results

Variable	Coefficient	Std. error	t values	p values
Intercept	7.111703	0.980449	7.254	3.80e-12	***
LnCaseNum	− 0.038233	0.012847	− 2.976	0.003168	**
LnClust4_1	0.266468	0.41211	6.466	4.32e-10	***
LnClust4_2	− 0.184126	0.043141	− 4.268	2.68e-05	***
LnClust4_3	− 0.006281	0.043976	− 0.143	0.886520
LnClust4_4	0.102171	0.044541	2.294	0.022520	*
LnClust4_5	− 0.183306	0.045945	− 3.990	8.40e-05	***
LnClust4_6	0.157594	0.046314	3.403	0.000762	***
LnClust4_7	0.38103	0.038052	10.068	< 2e-16	***
Dum24_5	− 4.200313	0.238334	− 17.624	< 2e-16	***
Dum1_6	2.994115	0.423527	7.069	1.19e-11	***

The impact of the number of COVID-19 cases on each cluster in the physical market

Table 7 illustrates the number of COVID-19 cases for each cluster according to the model’s result. In each cluster, the number of COVID-19 cases has a negative impact on sales. The least affected product group is “Daily Needs,” which also has an insignificant p value as well. Other clusters show that there is a significant negative impact of COVID case numbers on sales.

Table 7.

The impact of the number of COVID-19 cases on each cluster in the physical market

	Variable	Coefficient	T values	P values
Daily needs	LnCaseNum	− 0.01	− 1.44	0.15
Protein-based products	LnCaseNum	− 0.03	− 2.28	0.02
Basic consumption goods	LnCaseNum	− 0.03	− 2.97	0.00
Cold-chain products	LnCaseNum	− 0.03	− 2.33	0.02

The impact of the number of COVID-19 cases on each cluster in the online market

The number of COVID-19 cases in each cluster positively affects the online market; as the number of infected cases increases, sales levels also increase. The least affected product group is Basic Consumption Goods, as shown by its p value and t value.

As can be seen from Tables 7 and 8, the physical market is negatively affected by the number of infected cases, whereas online market sales are positively impacted by it. In the physical market, the least affected cluster is Daily Needs, while Basic Consumptions goods are the least affected in the online market. The decrease in sales levels in the physical market caused an increase in those of the online market for Daily Needs.

Table 8.

The impact of the number of COVID-19 cases on each cluster in the online market

	Variable	Coefficient	t values	p values
Daily needs	LnCaseNum	0.07	4.71	0.00
Protein-based products	LnCaseNum	0.05	3.18	0.00
Basic consumption goods	LnCaseNum	0.03	1.99	0.04
Cold-chain products	LnCaseNum	0.06	3.72	0.00

Case numbers for the Daily Needs category are only statistically significant, with a 0.85 confidence level for the physical market’s regression model. On the contrary, the regression model in the online market is significant at the 0.995 level. These results suggest that there is a non-dramatic decrease in the sales for the physical market and a considerable increase in demand for the online market. This circumstance suggests that the increase in online channels results in not only a decrease in sales for the physical channel but also the existence of extra demand for the online channel.

Conclusion and further suggestions

Unlike natural disasters, such as earthquakes, tsunamis, and tornadoes, pandemics spread worldwide, and their effect is not limited to a specific region. The most striking difference between pandemics and other disasters is that the economic impact is not caused by the pandemic but by how the public responds. During the COVID-19 crisis, consumers’ spending priorities quickly shifted as it fanned out and authorities took action.

As revealed by this research findings, consumers have turned to online retailers for essentials and discretionary items alike during the crisis. Therefore, the acceleration of e-commerce adoption translates into greater business-to-consumer (B2C) package volumes, and more pressure is placed on last-mile delivery performance. To thrive after the pandemic, retailers must understand and adapt to these changes. To provide customers with what they want, retailers must carry various items. Many of these changes in spending patterns are likely to be temporary, reflecting caution in times of great uncertainty; as conditions improve and consumer confidence returns, discretionary spending in some categories rebounds. However, some shifts may prove to be more permanent. Before COVID-19, retailers mainly focused on their supply chains’ efficiency, speed, and reliability. The pandemic showed how quickly global supply chains could be disrupted, slowed, and even stopped. Therefore, retailers should prioritize resiliency rather than efficiency by incorporating integrated and agile approaches.

Immediately after the first official COVID-19 case that was announced in the early hours of the 13th of March 2020, Turkey took several precautions to limit the spread of the virus. The first impact of the pandemic on food and fast-moving consumer goods (FMCG) retail sub-sectors triggered an acceleration in consumption just 1–2 days after the Minister of Health’s announcement. However, the rapid increase in the number of new cases from 3 on the 15th of March to 670 on the 20th of the same month forced the government to enact nationwide restrictions that would affect the retail sector, such as curfews for the elderly and the temporary closure of service businesses, such as barbershops, restaurants, dining places, and patisseries (other than delivery or takeaway). The first restrictions for grocery shopping came into effect by April 2020. In addition to weekend curfews and limitations on the maximum number of customers allowed in supermarkets, the opening hours of stores were also shortened.

In the first month of the pandemic, YoY growth in retail sales with constant prices contracted to 0.7% from 11% of the previous month. Hence, the situation's urgency was not acknowledged until the last week of March; retail indicators for the third month of the year failed to reflect the real impact of the COVID-19 pandemic on retail economics in Turkey. However, in April and May 2020, retail sales recorded a yearly 20% and 17.4% YoY decline, respectively. On the one hand, the textile and apparel sub-group of the retail sector suffered the worst decline in sales with 41% YoY terms due to public health restrictions, while the grocery sub-sector remained robust, with a 12% yearly increase in April 2020. Finally, the online retail sector continued its expansion during the pandemic, and the average yearly increase in online monthly sales stood at 95% for the second half of 2020. However, soon after restrictions were eased in June 2020, the annual trends in the retail sector returned to their normal levels.

This research aimed to investigate whether the COVID-19 pandemic changed consumer shopping behavior in supermarkets. The most affected product groups were also revealed by comparing the changes in shopping habits in physical versus online markets. One of the largest retailers in Turkey was selected for the case study, and their data from 2019 to 2020 were used to address the question mentioned earlier. After cleaning the data, a hierarchical clustering technique was used to identify the product groups that showed similarities in shopping behaviors for both physical and online markets. The sales levels of each product group were then estimated using three different regression models—Stepwise, Lasso, and Best-fit—separately for each market. The best model for each cluster per market was selected, and the changes in shopping behavior for each cluster group in both physical and online markets were highlighted.

The cluster analysis showed that all six product groups saw a significant change in sales levels compared to 2019, the pre-COVID period. The most in-demand product group in both physical and online markets is Basic Consumption Goods (i.e., grains, pasta, and sides; cleaning and paper products). In online markets, sales for all product groups increased compared to 2019. Even after curfews, physical market sales for all product groups were still significantly higher.

All of the clusters in both physical and online markets saw a sudden increase in sales at the beginning of the COVID-19 period, suggesting stockpiling and panic-buying behavior in both shopping channels. At the end of 2020, behaviors turned into normal purchases since sales numbers began to decrease after several months. Online markets still had a significantly higher sales ratio, signaling that people started to get used to online shopping because various products were made accessible to the consumer. Based on the physical market’s regression model, the number of cases parameter for the Daily Needs category is only significant with a 0.85 confidence level, which translates to a non-dramatic decrease in sales for this category. On the other hand, the parameter indicating the infected daily cases in the online market’s regression model is significant at 0.995 level and indicates a substantial increase in demand. This phenomenon reveals that the increase in the case of the Daily Needs cluster in online channel sales not only stemmed from demand reverted from the physical channel but also the extra demand attracted for the online channel.

As can be seen from the regression analysis conducted, the number of COVID-19 cases affected shopping habits. Curfews also had a significant effect on sales levels. The included dummy variables that indicated the dates where a significant change was observed afterward—the 25th of May, 1st of June, and 12th of December—showed that the sales levels on these dates were the most affected. Limitations, curfews, and the number of cases made people shift from physical shopping to shopping through online markets.

In line with the results derived above, while the effects of the pandemic de-escalated in the final quarter of 2021 thanks to the worldwide vaccination programs, the retailer company investigated in this research announced that they would focus on enhancing their online channel penetration in the future. According to the company’s recent press releases, the increased sales from its online channel acceleration continued through 2021. By the third quarter of 2021, the YoY change in online sales reached 99% (CPI YoY 20%). Company officials also reported that the share of online sales had reached 15% of sales, excluding alcoholic beverages and tobacco. In addition, the company plans to establish a new online grocery section to support its expanding online channel. This attempt is totally in parallel to the evolving digital age, which provides significant opportunities to retailers in the face of challenges (Villanova et al. 2021).

Based on the methodology proposed in this research, changes in consumer purchasing habits among retailers having similar product configurations in Turkey can be evaluated in future research. This will allow more generalized results to be obtained for Turkey’s retail sector. In addition, the proposed methodology can also be utilized to investigate the changes in consumer shopping habits in retailers in other countries. This can specify whether the changes in consumer shopping habits of a country show any similarities or differences to other countries.

Author contributions

ÖZ: provided the dataset, FÜ and BÜ: designed the model and computational framework, EHG: analyzed the dataset, carried out the implementation and performed the calculations. The study was supervised by FÜ as Thesis Advisor. The first draft of the manuscript was written by EHG and all the authors commented on previous versions of the manuscript. All the authors read and approved the final manuscript.

Funding

No funding was received for conducting this study.

Data availability

The data that support the findings of this study are available from Migros but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are, however, available from the authors upon reasonable request and with permission of Migros.

Declarations

Conflict of interest

All the authors declare that they have no conflicts of interest.

Ethical approval

This article does not contain any studies with human participants or animals performed by any of the authors.

Consent for permission

No permission required. All figures, diagrams, schemes, tables, and text in this manuscript are original, unpublished work.