Evidence of the time-varying impacts of the COVID-19 pandemic on online search activities relating to shopping products in South Korea

The COVID-19 pandemic, with its various contagious variants and multiple waves of infection, significantly altered public risk perception and social behavior. The absence of initial vaccines exacerbated the situation, leading people to avoid in-person activities, including shopping. Consequently, online shopping experienced a surge in popularity as people increasingly preferred digital platforms. This shift provides a valuable opportunity to study the pandemic's socioeconomic impacts using big data from online information-seeking activities. Previous research has examined the impact of the COVID-19 pandemic on online activity, but fewer studies have considered the time-varying effects across multiple waves due to data collection and processing challenges. Online social monitoring data, however, offers a cost-effective and real-time way to track such changes. While existing studies have analyzed changes in online activity patterns, the detailed socioeconomic impact across various COVID-19 waves remains largely unexplored. Existing literature explores online shopping patterns during the pandemic. The "Reacting Coping Adapt" (RCA) framework suggests that consumer behavior goes through three phases: reaction, coping, and adaptation. The reaction phase involves immediate behavioral changes based on perceived risk; the coping phase sees adaptation to public policies; and the adaptation phase leads to the establishment of new, less reactive purchasing habits. This framework has been validated in France, but broader application remains limited. Internet service providers collect valuable data on online search activities, allowing analysis of user interest and optimization of search algorithms. Increased search volume for specific products indicates emerging demand, vital for inventory and supply chain management. Studies have shown the utility of online data in predicting economic indicators, consumption, and epidemics. However, such data have been underutilized to study the impact of multiple COVID-19 waves on socioeconomic activities. NAVER, South Korea's most popular search engine, provides the NAVER DataLab Shopping Insight (NDLSI) dataset containing weekly online search activity volumes for over 1,800 shopping products. This dataset enables the investigation of the COVID-19 pandemic's impact on Korean online shopping habits within the RCA framework, addressing key questions regarding the major components of online search activities before and after the pandemic, the impact of virus waves on shopping behavior, and the influence of prevention policies.

Several studies have explored consumer behavior changes during the COVID-19 pandemic. Sheth (2020) discussed the potential for the pandemic to disrupt existing purchasing habits and create new ones, shaped by socioeconomic factors and public policy. The RCA framework (Kirk and Rifkin, 2020; Guthrie et al., 2021) offers a useful lens for understanding these changes, identifying distinct phases of reaction, coping, and adaptation. Other research has examined the use of online social network data (e.g., Twitter) to predict stock market fluctuations (Almehmadi, 2021), and the use of Google Trends data to forecast economic indicators (Carrière-Swallow and Labbé, 2013; Choi and Varian, 2012), private consumption (Vosen and Schmidt, 2011), and epidemics (Carneiro and Mylonakis, 2009; Teng et al., 2017). The use of online data to study responses to natural disasters has also been explored (Gizzi et al., 2020; Kam et al., 2021; Kam et al., 2019; Kim et al., 2019). However, the application of these methods to analyze the dynamic effects of multiple COVID-19 waves on e-commerce remains relatively novel. Studies at the individual level have shown relationships between decision-making and consumer behavior during the pandemic (Birtus and Lăzăroiu, 2021; Smith and Machova, 2021; Vătămănescu et al., 2021), but national-level patterns were less explored. This study addresses this gap by using the NDLSI dataset to analyze national-level behavior changes in South Korea.

This study leverages the NAVER DataLab Shopping Insight (NDLSI) dataset, which contains weekly click counts for 1,837 shopping products from the NAVER Shopping platform between July 31, 2017, and August 30, 2021. The data are categorized into three levels: 11 first-level categories (e.g., Fashion Clothing, Digital/Home Appliance, Food), 204 second-level categories, and 1,837 individual products. These categories are based on merchant category codes (MCCs) commonly used by credit card issuers. Six COVID-19 metrics from the Johns Hopkins University Center for Systems Science and Engineering (JHU CSSE) dataset were used: new confirmed cases, stringency index, residential index, vaccination index, new death cases, and fatality rate. Weekly sums of new cases and weekly averages of the other metrics were calculated to align with the NDLSI data's weekly frequency. Weekly average temperatures from 95 stations across South Korea were obtained from the Korea Meteorological Administration (KMA) to account for seasonal effects. Singular Value Decomposition (SVD)-based Principal Component Analysis (PCA) was employed to extract major modes of online search activity patterns from the NDLSI data. The SVD algorithm decomposes the covariance matrix of the data into three matrices (U, Σ, V^T), where U represents orthogonal eigenvectors (principal components or PCs), Σ contains eigenvalues representing variance explained, and V^T is the transpose of the right singular vectors. The analysis was performed across five periods: before the pandemic (Wave 0), and Waves 1-4, defined based on surges in new confirmed cases. Linear trends and seasonality were removed from the data before SVD analysis to isolate the pandemic's impact. Spearman's rank correlation was used to assess relationships between the principal components and individual product search activities, and between the principal components and the COVID-19 metrics. This non-parametric method was chosen due to the non-normal distribution of the search activity data. Quantile-Quantile (QQ) plots were used to assess the stability and reliability of the correlation distributions across different numbers of shopping products. A threshold of 0.45 for Spearman's correlation coefficient was used to identify products strongly associated with the principal component capturing pandemic-related changes.

Before the pandemic (Wave 0), the first principal component (PC1) represented a monotonic increase in overall search activity, while the second (PC2) reflected seasonality linked to temperature, with a four-week lag. After detrending and deseasonalizing the data, the first principal component in subsequent waves (Waves 1-4) strongly correlated with new confirmed COVID-19 cases, explaining an increasing proportion of the variance (from 20% to 27%). The analysis revealed that life/health, digital/home appliance, and food products were consistently associated with COVID-19 waves. The number of products strongly correlated with the PC1 increased over time, from 241 after Wave 1 to 714 by Wave 4. After Wave 2, the number of associated products decreased, reflecting a potential shift from the 'react' to 'cope' phase of the RCA framework. The emergence of leisure/life convenience items (e.g., workout classes, international travel) after Wave 4 indicated a potential shift toward the 'adapt' phase, suggesting a reduced public perception of pandemic risk. Spearman's rank correlation analysis between the PC1 and the six COVID-19 metrics showed that new confirmed cases, stringency index, residential index, new deaths, and fatality rate were significantly associated with changes in online search activity patterns. In contrast, the vaccination index showed a weaker correlation. The stringency index and fatality rate exhibited particularly strong correlations, indicating that consumer behavior was highly sensitive to the strictness of government policies and the severity of the pandemic. QQ plots confirmed the robustness of these correlation patterns across different numbers of selected products.

This study demonstrates the utility of online social monitoring data in understanding the dynamic impacts of a pandemic on consumer behavior. The use of NDLSI data provided valuable insights into emerging purchasing patterns, suggesting that online search activity can be a leading indicator of future purchasing trends. The findings are largely consistent with the RCA framework, identifying distinct phases of reaction, coping, and adaptation in online shopping behavior across the pandemic's different waves. The increasing proportion of variance explained by PC1 in later waves highlights the growing influence of the pandemic on online shopping activity. The strong association between search activity patterns and the stringency index and fatality rate underscores the sensitivity of consumer behavior to both government interventions and the perceived severity of the health crisis. The relatively weak correlation with the vaccination index may be due to the fact that vaccination uptake lagged behind the other metrics during the study period. The study's findings can inform both public health strategies and business planning, enabling better preparation for and response to future pandemics.

This study successfully extracted major patterns of public interest in shopping products and analyzed changes in online search activities during the COVID-19 pandemic. The application of SVD-based PCA to NDLSI data revealed dynamic patterns in online search behavior across four distinct waves of the pandemic. The findings suggest that consumer behavior shifted through the 'react,' 'cope,' and 'adapt' phases of the RCA framework. The strong association between online search activity and COVID-19 metrics (except vaccination) highlights the importance of considering both public health policies and the perceived risk level. Future research could integrate online social monitoring data with actual purchase data from credit card or barcode records to further enhance understanding and prediction accuracy. This study highlights the potential of online social monitoring data in planning and responding to future pandemics.

This study uses online search activity data as a proxy for actual purchasing behavior. While search activity can be a leading indicator of future purchases, it doesn't directly measure actual sales. Future work could integrate purchase data (e.g., credit card or barcode data) to provide a more comprehensive understanding of consumer behavior. The analysis relies on correlations, which do not imply causation. While the findings suggest potential triggers for changes in consumer behavior, further research is needed to establish causal relationships. The study focuses on South Korea, and the findings may not be generalizable to other countries with different cultural contexts, pandemic management strategies, or online shopping habits.