Integrating Unstructured Social Media Sentiment, User Clickstream, and Historical Sales Data: A Multi-Source Fusion Model for Predictive E-Commerce Demand Analytics

• Ada John

  ladoke Akintola university of technology

  Author

Background: The integration of heterogeneous data sources for demand forecasting has emerged as a critical frontier in e-commerce analytics, yet existing approaches remain fragmented—either relying solely on transactional data or treating social media sentiment and clickstream behavior as isolated signals. Research gap: No validated multi-source fusion framework exists that systematically combines unstructured social media sentiment, granular user clickstream sequences, and historical sales records within a unified predictive architecture tailored for high-variance e-commerce demand environments. Purpose: This study develops and validates a Multi-Source Temporal Attention Fusion (MSTAF) model that integrates these three data modalities to improve demand forecasting accuracy and lead time. Methodology: Using a retrospective dataset comprising 2.1 million clickstream events, 487,000 social media mentions, and 36 months of transactional records from 150 product categories, we compared MSTAF against six baseline models (ARIMA, XGBoost, LSTM, GRU, LightGBM, and single-source transformers) using stratified 5-fold cross-validation. Key findings: MSTAF achieved 89.4% directional accuracy and 12.7% MAPE, outperforming the strongest baseline by 14.2 percentage points in directional accuracy and reducing forecast error by 31.4% during promotional periods. Sentiment volatility and cart-recovery sequences emerged as the strongest cross-modal predictors. Main conclusion: Multi-source fusion substantially outperforms single-source and pairwise approaches, with sentiment and clickstream data providing complementary predictive signals that compensate for each other's missingness. Implications: The framework offers practitioners a replicable architecture for inventory optimization and a 48-hour lead time advantage over traditional models.

• pdf

Copyright (c) 2026 Ada John (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.