The impact of social media discourse on financial performance of e-commerce companies listed on Borsa Istanbul

The study asks how shifts in social media discourse during COVID-19 relate to the financial performance of e-commerce firms in Borsa Istanbul’s XTCRT index. It motivates a multilayer view because firms interact across price, high/low, and volume dimensions and are influenced by broader socio-economic discussions online. Using multilayer networks and topic-classified Twitter content (culture, economy, health, politics, technology, world), the paper aims to reveal how online discourse maps onto structural and topological network measures of market behavior. The work is positioned as important for investors and policymakers seeking to understand systemic risks, contagion, and the impacts of public sentiment on market dynamics, especially in crisis periods.

The review highlights COVID-19’s disruptive effects on global financial markets, including Borsa Istanbul, and notes that prior work largely relies on traditional metrics, overlooking social media’s role. It synthesizes evidence that social media sentiment can predict market movements and affect prices, volumes, and volatility. However, sector-specific analyses (e.g., e-commerce) and nonlinear modeling are scarce. The literature underscores increased investor reliance on social media during uncertainty and calls for frameworks integrating financial and non-financial (social discourse) data with advanced methods (multilayer networks, LSTM, NARDL) to capture complex, potentially asymmetric effects of sentiment on market outcomes.

Data and setting: The focus is on nine XTCRT e-commerce-related companies (BIMAS, BIZIM, CRFSA, DOAS, DOHOL, MEPET, MGROS, SOKM, TKNSA) over March 9–November 1, 2020. Financial data comprise daily close, high, low, volume. Twitter data (Turkish, COVID-19 related) were collected at 15-minute intervals using keywords (Covid/Corona/Kovid/Korona), de-duplicated, missing gaps filled via ARIMA, and synchronized to trading days. Topic series are formed using 5-day sliding windows and medians due to skewness. NLP and topic classification: Turkish-language preprocessing uses Zemberek (tokenization, stop-words, stemming, spell correction). A deep learning classifier with GloVe embeddings and LSTM layers is trained on 4,900 Turkish news items (700 per topic across six categories: culture, economy, health, politics, technology, world). The model employs Categorical Cross-Entropy; training over 50 epochs stabilizes by ~epoch 40, achieving >90% measurement accuracy on held-out tests, then applied to tweets to produce daily topic-volume series and their log differences. Multilayer network construction: A 4-layer financial correlation network is built for each 5-day window with nodes as the nine firms and layers corresponding to close, high, low, and volume. Correlation distances (stronger correlation → lower weight) are filtered using Planar Maximally Filtered Graphs (PMFG) to retain salient structure. Node strength within a layer s_il is summed across layers to obtain weighted overlap degree o_i; the multilayer participation coefficient P(i) assesses how evenly a node participates across layers (with categorization into focused, mixed, truly multilayer). Z-scores of overlapping strength z(o_i) identify hubs vs. ordinary nodes. Multilayer entropy H measures network complexity and information distribution across layers. Nonlinear ARDL (NARDL): To assess short- and long-run, and asymmetric effects, NARDL models relate multilayer metrics to topic series. Two dependent series are used: T_mp (mean multilayer participation coefficient) and T_en (multilayer entropy). z-score series are visualized but not modeled due to nonlinearity/variance issues. Topic series are decomposed into positive and negative partial sums to capture asymmetries. Lag selection uses AIC/BIC with maximum lag 4. Cointegration is tested via Pesaran, Shin, Smith bounds testing. Diagnostics include Jarque–Bera (normality), LM (serial correlation), and ARCH tests (heteroskedasticity).

- Network dynamics: Mean participation coefficients fluctuate between ~0.7–1.0 across windows, indicating periods of strong and weak cross-layer interconnectedness. Mean z-scores oscillate around zero with occasional spikes, marking atypical performance. Entropy ranges ~0.5–2.5, often high, indicating complex, interconnected market states with intermittent simpler regimes. - Classifier performance: LSTM topic classifier exceeds 90% accuracy on test data; topic tweet volumes show distinct temporal patterns (e.g., economy and health spikes consistent with pandemic news cycles). - Strong autoregressive effects: T_mp shows significant negative own-lag effects (e.g., coefficients around −0.28 to −0.32; p<0.001), indicating short-term mean reversion in participation. T_en shows strong negative own-lag effects (≈−1.07; p<2×10^−16), implying entropy is largely driven by its history. - Cointegration: Bounds tests generally indicate cointegration between multilayer metrics and topic series (e.g., F-statistics for T_en models often >40; p<2×10^−16), supporting long-run relationships despite limited short-run significance in many topics. - Culture: T_mp relation shows positive but marginal short- and long-run coefficients for positive/negative cultural changes (p≈0.06–0.08) and suggestive long-run asymmetry (W≈4.82; p≈0.09). For T_en, culture effects are not significant; cointegration present; no asymmetry. - Economy: For T_mp, economy lags are mostly not significant; cointegration present; long-run asymmetry suggested by tests, though short-run asymmetry not significant. For T_en, economy effects are small and insignificant; cointegration present; no asymmetry. - Health: Effects on both T_mp and T_en are generally insignificant in short and long run; cointegration present; no asymmetry. - Politics: For T_mp, a short-run positive effect at lag 1 is significant (t≈2.10; p≈0.037), while other terms are not. Long-run asymmetry is significant (W≈25.88; p≈0.00017). For T_en, political effects are not significant; cointegration present; no asymmetry. - Technology: For T_mp, tech terms are positive but not significant; cointegration present. For T_en, delayed positive technology effects are significant/borderline (e.g., short-run coefficient ≈0.361; p≈0.045; long-run ≈0.334; p≈0.044), indicating technology discourse raises network entropy; other tech asymmetry tests are generally not significant. - World: For T_mp and T_en, positive coefficients are observed but are not statistically significant; cointegration present; no asymmetry. Overall: Social media topics show differentiated impacts. Market structure metrics are highly path-dependent; political and especially technology-related discourse exhibit measurable effects (politics on participation; technology on entropy). Economy shows indications of long-run asymmetry for participation but lacks robust short-run effects; culture, health, and world are largely non-significant for entropy and mostly marginal for participation.

The research question concerned whether and how social media discourse across distinct topics maps onto multilayer financial network behavior of e-commerce firms during COVID-19. The results indicate that market network structure is dominantly influenced by its own recent past (strong negative autoregressive effects), consistent with mean-reverting interconnectedness and persistent complexity. Against this backdrop, selected topics exert measurable incremental influences: political discourse associates with short-run increases in cross-layer participation and exhibits long-run asymmetry, consistent with policy/news shocks affecting how broadly firms comove. Technology discourse significantly elevates network entropy with a lag, suggesting that tech-related public discussions propagate uncertainty/complexity across layers, possibly through expectations about digital adoption, regulation, and operational shifts central to e-commerce. Economic discourse shows cointegration and suggestive long-run asymmetry with participation, implying that the direction of economic news may matter differentially over horizons, even when immediate effects are muted. Topics like culture, health, and world affairs appear less directly tied to the sector’s structural metrics in this window, despite their prominence during the pandemic. For practice, this implies that monitoring topic-specific social media streams, particularly politics and technology, can enhance understanding of sectoral network dynamics, complementing traditional indicators for risk assessment and strategy. For policy, findings support efforts to manage information flows and counter disinformation that could amplify systemic fragility in sensitive sectors.

The study integrates multilayer financial networks with topic-classified social media data to assess the influence of online discourse on the XTCRT e-commerce sector during COVID-19. Contributions include: (1) a 4-layer network capturing close, high, low, and volume with PMFG filtering and multilayer metrics (participation, entropy, z-scores); (2) a Turkish-language LSTM topic classifier; and (3) NARDL models quantifying short-/long-run and asymmetric effects. Key conclusions: (a) market network metrics are highly autoregressive (mean-reverting participation; persistent entropy dynamics); (b) political discourse has tangible short-run effects on participation and long-run asymmetry; (c) technology discourse raises entropy with delayed effects; (d) economic discourse shows cointegration and potential long-run asymmetry for participation; (e) culture, health, and world topics show limited direct effects in this sample. Policy implications include strengthening transparency and anti-disinformation measures, and tailoring safeguards for sectors that are sensitive to public discourse. Future research should broaden sectors, extend time horizons, integrate multiple social platforms, refine sentiment/topic granularity, and explore causal frameworks linking discourse to network restructuring and firm-level outcomes.

- Sector and sample scope: Only nine XTCRT e-commerce-related companies, potentially limiting generalizability within Borsa Istanbul. - Time window: Approximately eight months in 2020 (pandemic onset), which may capture crisis-specific dynamics not representative of normal times. - Platform coverage: Social media discourse sourced from a single platform (Twitter/X); cross-platform differences are not captured. - Data and modeling nuances: Non-normal residuals (JB tests) in several models; topic assignment and tweet-volume proxies may imperfectly capture investor-relevant sentiment; z-score series excluded from NARDL due to variance/linearity constraints. - Potential ambiguity between training corpus (news articles) and tweets used for inference, and limited tweet coverage interruptions filled via ARIMA may introduce measurement noise.