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

The study examines how the multi-year COVID-19 pandemic altered public risk perception and online shopping-related search behaviors in South Korea. Motivated by the multiple waves of coronavirus variants and evolving prevention policies, the authors aim to quantify time-varying impacts on e-commerce information-seeking. They identify gaps in prior work, noting that most studies did not capture temporal changes across multiple waves due to data collection and preprocessing costs. Leveraging NAVER DataLab Shopping Insight (NDLSI) weekly search volumes for 1,837 products (2017–2021), they pose three questions: (1) What are the major components of dynamic patterns of online search activities before and after COVID-19? (2) How did social behavior patterns in online shopping searches change over multiple waves of virus variants? (3) Which prevention policies were key factors for temporal changes in online shopping searches during the pandemic? The study situates findings within the React–Cope–Adapt (RCA) behavioral framework to interpret phase transitions in consumer behavior across waves, highlighting the importance of big social monitoring data for real-time, cost-effective insights into socioeconomic changes during public health crises.

Prior research documents increased online activity and changes in consumer behavior during COVID-19, including shifts from in-person to online shopping driven by risk perception and policy strictness (Dryhurst et al., Grashuis et al., Li et al., Mouratidis and Papagiannakis, Pham et al.). The RCA framework (react, cope, adapt) has been proposed and validated in some contexts (Kirk and Rifkin; Guthrie et al.) but lacks broader cross-country and behavior applications. Social monitoring data (e.g., Google Trends, Twitter) have been used to forecast economic indicators, consumption, epidemics, and to monitor responses to disasters (Carrière-Swallow and Labbé; Choi and Varian; Carneiro and Mylonakis; Teng et al.; Kam et al.). NAVER is the dominant Korean platform, and its Shopping data can indicate emerging demand (Woo and Owen). The paper also notes that rumor impacts on online activities can diminish during nationwide crises (Park et al.; Kam et al.), and that individual-level decision-making and sentiment changes during COVID-19 have been studied (Birtus and Lăzăroiu; Smith and Machova; Vătămănescu et al.; Liu et al.), but national-level effects of multi-wave policies on social behavior remain underexplored. The study positions NDLSI as a high-quality, timely dataset aligned with merchant category codes (MCCs) similar to card spending studies (Darougheh; Dunphy et al.), enabling a novel analysis of time-varying socioeconomic impacts across waves.

Data: NDLSI provides weekly relative search volumes (0–100 scaled to each item’s maximum) for 1,837 shopping products from NAVER Shopping, spanning 214 weeks (2017-07-31 to 2021-08-30). Categories are hierarchical: 11 first-level, 204 second-level, and 1,837 third-level items (translated from Korean). Missing values were replaced with zeros. COVID-19 and auxiliary metrics: Six weekly metrics derived from JHU CSSE and related sources were used: (1) new confirmed cases (weekly sum), (2) stringency index (weekly average; 0–100), (3) residential (stay-at-home) mobility index (weekly average), (4) vaccination index (partial vaccination rate; weekly average), (5) new deaths (weekly sum), and (6) fatality (new deaths/new confirmed; weekly average). Weekly mean temperature from 95 KMA stations was computed to characterize seasonality. Wave definitions: Five analysis periods were set using cumulative windows to capture time-varying impacts and enable comparison to previous waves: pre-COVID (Wave 0: 2017-07-31 to 2019-12-31), Wave 1 (through 2020-05-25), Wave 2 (through 2020-10-19), Wave 3 (through 2021-03-01), and Wave 4 (through 2021-08-31). Waves were defined by surges in new confirmed cases. Preprocessing and PCA: Singular Value Decomposition (SVD)-based Principal Component Analysis (PCA) was applied to the covariance matrices of NDLSI data for each period. Before PCA for Waves 1–4, the authors removed the overall increasing linear trend and temperature-related seasonality previously identified in pre-COVID analysis, to focus on pandemic-related variability. SVD yields U (principal components), Σ (eigenvalues/variance explained), and V^T (loadings). Variance explained by leading PCs quantifies dominant modes. Association analyses: To relate search dynamics to product categories and pandemic metrics, Spearman’s rank correlations were computed: (a) between PC1 (time series) and each product’s detrended search series to identify associated items (threshold ρ ≥ 0.45 selected to capture up to ~20% of items based on KDE-derived distributions); (b) between COVID-19 metrics and selected product series (e.g., 31 items consistently associated with PC1 across waves). Kernel density estimates (KDE) characterized correlation distributions. Stability checks using Quantile–Quantile (QQ) plots assessed sensitivity of distribution shapes to sample size, guiding selection of top 50 items when constructing metric–item correlation distributions. Sankey diagrams visualized category flows of PC1-associated items across waves.

- Pre-COVID principal modes: PC1 captured a monotonic increasing trend in online shopping-related searches, explaining ~14.8–15% of variance. PC2 explained ~10% and reflected seasonality aligned with national temperature, leading by ~4 weeks. Summer products (e.g., fan, parasol, yeolmu kimchi, tarp) were positively correlated with PC2; winter products (e.g., brooch, beanie, neck cape) were negatively correlated. - Pandemic-era dominant mode: After detrending, PC1 closely tracked new confirmed COVID-19 cases across waves, with increasing explained variance: Wave 1 ~20.5%, Wave 2 ~20.3%, Wave 3 ~24.6%, Wave 4 ~27.3%. This indicates strengthening pandemic-related structure in search behavior. - Growth in associated items: PC1-associated items (ρ ≥ 0.45) increased from 327 (Wave 1) to 504 (Wave 2), 593 (Wave 3), and 714 (Wave 4), more than doubling from Wave 1 to Wave 4. New inflows after each wave were substantial: after Wave 1, 241 new items; after Wave 2, 124–125 items; after Wave 3, 190 items. - Category composition and inflows: Life/Health and Digital/Home Appliance consistently represented large shares of associated items; Food and Childbirth/Childcare also featured prominently. Inflows by category: after Wave 2 (N≈125), Life/Health 29%, Digital/Home Appliance 19%, Childbirth/Childcare 12%; after Wave 3 (N=190), Life/Health 22%, Digital/Home Appliance 17%, Childbirth/Childcare 19%. Duty-free category appeared after Wave 2; Leisure/Life convenience (e.g., workout classes, abroad travel) appeared after Wave 4, hinting declining perceived risk. - Persistent PC1-associated products: 31 items remained associated across all waves, over one-third from Life/Health. Two temporal groups emerged: (1) rising association over time (e.g., minidisc player, monitor arms, webcam, interphone box, fabric, handicraft supplies, processed snacks, cooking oil, bread, tuning supplies, craft, feed, seeds/seedlings, water aperture, gravel/sands/soil, landscape tree/sapling, first aid/emergency supplies); (2) declining association (e.g., gas range, microwave, toothbrush, hula hoop). - Associations with COVID-19 metrics: Correlation distributions showed strongest and tightest positive associations with Stringency and Fatality indices (often ρ > 0.8 for top items); Residential (stay-at-home), New confirmed cases, and New deaths also showed positive associations but with greater spread; Vaccination exhibited weak or insignificant associations. For the 31 persistent items, most had high positive correlations with New confirmed, Stringency, Residential, New deaths, and Fatality; Vaccination was generally weak. Some items (e.g., gas range, baby walker, toothbrush) showed comparatively lower correlations, consistent with their declining association over waves.

Findings indicate that online shopping search behavior in South Korea became increasingly structured by pandemic dynamics, with PC1 tracking infection waves and growing in explanatory power through Wave 4. Interpreted via the RCA framework: Wave 1 corresponds to the react phase (rapid behavioral shifts following heightened risk and strict policies); between Waves 2–3 reflects coping (smaller inflows when cases were relatively low, suggesting stabilization of new habits); Wave 4 signals movement toward adapt (emergence of leisure/life convenience and travel-related searches indicating reduced perceived risk). The strong links to Stringency (policy strictness) and Fatality (perceived severity) underscore the role of government measures and objective risk in shaping consumer information-seeking. Practical implications include improved inventory and supply chain planning for items tied to self-protection and home-based living, dynamic SEO and recommendation algorithms that adjust to regime shifts (pre-, mid-, and post-pandemic), and policy design that anticipates market responses to changing stringency. Integrating social monitoring with purchasing and insurance data could enhance risk management strategies and help both public and private sectors mitigate disruptions during future crises.

The study extracted dominant modes of South Korean online shopping-related search activity using SVD-based PCA on NDLSI data (2017–2021). Pre-COVID dynamics were governed by an increasing trend and temperature-driven seasonality. During COVID-19, the detrended first principal mode aligned with infection waves, with rising variance explained from Wave 1 to Wave 4. Category- and product-level analyses revealed evolving consumer interests: Life/Health, Digital/Home Appliances, Food, and Childbirth/Childcare were most associated with waves, with notable inflows of new associated items after each wave. Strong associations with Stringency and Fatality indices, and weak links with Vaccination, suggest that policy strictness and perceived severity most influenced search behavior. Framed by the RCA model, evidence points to transitions from react to cope and indications of adapt after Wave 4. The work highlights the utility of big social monitoring data for preparedness, mitigation, and recovery planning in future pandemics. Future research should integrate actual purchase records (e.g., credit card and barcode data) and survey/interview methods to establish causal drivers, validate search–purchase links, and refine predictive models of social behavior change.

- Observational correlations cannot establish causality; identified drivers are potential rather than confirmed triggers of behavior change. - Data reflect online search activity, not actual purchases; conclusions about demand are indirect and require validation with transaction data (e.g., credit card, barcode) to assess search-to-purchase conversion. - Results pertain to South Korea’s dominant platform (NAVER) and policy context; generalizability to other countries/platforms may be limited. - Item names were translated from Korean via automated tools, which may introduce minor categorization or semantic errors. - Missing values were set to zero, and detrending/seasonality removal choices may influence PCA outcomes. - Thresholds for association (e.g., ρ ≥ 0.45) and selection of top-k items for distributional analyses could affect item-level conclusions, though QQ-based sensitivity checks were performed.