Dynamic pricing in e-commerce using real-time data from social media sentiment analysis

Thesis Abstract

Dynamic pricing in e-commerce leveraging real-time social media sentiment to optimize price discrimination and demand responsiveness addresses the persistent challenge of aligning online pricing with volatile consumer perceptions and market competition. The study investigates whether incorporating real-time sentiment indicators from Twitter, Reddit, and Instagram, alongside traditional sales and inventory signals, improves pricing efficiency, revenue, and margin stability in online retail platforms. The aim is to develop an empirically validated pricing framework that integrates sentiment-driven signals with conventional econometric models to enhance dynamic pricing decisions under uncertainty. Specific objectives are (i) to model the relationship between social media sentiment metrics and short-run demand elasticities; (ii) to quantify the incremental revenue and margin gains from sentiment-enhanced pricing relative to baseline dynamic pricing; (iii) to identify optimal lag structures and feature ensembles for timely price updates; (iv) to assess cross-category heterogeneity and platform-specific effects on pricing performance; and (v) to evaluate robustness against noise and manipulation in user-generated content. The methodology adopts a mixed-methods research design anchored in both quantitative econometrics and validation through qualitative insights. The population comprises monthly transactional data from a mid-sized global e-commerce retailer spanning 24 months, including 1.2 million order records across five product categories (electronics, apparel, home goods, beauty, and toys). A stratified random sample of 15,000 orders is used for model estimation, with an additional 5,000 orders reserved for out-of-sample testing. Social media data are collected from four platforms (Twitter, Reddit, Instagram, and Facebook public pages) using a streaming API and web-scraping tools, yielding approximately 10 million sentiment-bearing posts monthly. Instruments of data collection include platform-specific sentiment scores (e.g., Vader-based sentiment polarity, LIWC-inspired affect scores), topic-modeling outputs (LDA themes), and engagement metrics (mentions, shares, comment counts). The study employs a two-stage modeling approach first, a dynamic pricing model specified as a hierarchical Bayesian VAR (vector autoregression) that integrates price, inventory, promotional variables, and sentiment-derived features; second, a machine-learning rewritter layer using gradient boosting (XGBoost) to capture nonlinear interactions and optimal lag structures. Model specifications include price level, competitor price proxies, stock availability, promotional campaigns, and sentiment indicators such as aggregate sentiment score, sentiment momentum, and platform-specific bullish/bearish signals. Validity and reliability are established through back-testing on the out-of-sample set, cross-validation, and backtesting-based performance metrics (mean absolute percentage error, revenue lift, and margin stability). Ethical considerations address data privacy, platform terms of service, and transparency in pricing decisions. Key expected findings include (i) sentiment signals provide statistically significant incremental explanatory power over traditional price determinants in short-horizon demand forecasting; (ii) optimal lag structures differ by product category, with fast-moving electronics showing quicker sentiment-to-price translation than durable goods; (iii) sentiment-aware pricing yields higher revenue per visit and reduced price volatility without eroding consumer trust; (iv) model performance is moderated by platform-specific dynamics and by the quality and moderation of social discourse; (v) robustness checks confirm resilience to noise through ensemble modeling and anomaly detection. The study anticipates a measurable average revenue lift of 2–5% across categories under realistic operating conditions, with margin gains sustained over volatile sentiment periods. Contributions to knowledge are threefold (a) the integration of real-time social media sentiment into dynamic pricing frameworks within e-commerce, extending the literature on revenue management under uncertainty; (b) methodological advancement through a hybrid Bayesian–machine-learning approach that captures both linear dependencies and nonlinear sentiment interactions; and (c) practical guidance for practitioners on feature engineering, lag selection, and governance of sentiment-driven pricing to balance profitability with customer perception. The main conclusion is that sentiment-informed dynamic pricing can enhance revenue and margin stability when integrated with robust statistical controls and category-aware deployment. Recommendations include implementing a controlled rollout with continuous monitoring of pricing impact, refining sentiment data quality protocols, and developing governance mechanisms to mitigate potential manipulation or bias in social signals, along with further research into long-term effects on customer trust and competitive dynamics.