An intelligent predictive framework for consumer returns forecasting: Leveraging social media data in the electronics service industry

Abstract

Liberal return policies can boost sales and repeat purchases, yet the rising volume of returns poses major operational challenges for service providers. Improving the accuracy of return forecasts can deliver significant cost savings and enhance strategic decision-making. This work offers a novel forecasting framework—High-Order Fuzzy Cognitive Maps (HFCMs) combined with an Improved Grey Wolf Optimization (IGWO) algorithm and Deep Ensemble Reinforcement Learning (DERL)—to predict daily customer returns. We collaborate with a consumer electronics service provider to gather (1) extensive internal operational data and (2) publicly available social media data from Twitter and Facebook. Unlike existing approaches that often employ simpler FCM or wavelet methods, our HFCM-IGWO-DERL design systematically integrates multivariate social media features, nonlinear optimization, and adaptive ensemble selection. Across both eight public benchmark datasets and real-world returns data, our model demonstrates 45–78% lower RMSE compared to leading baselines. These results underscore our theoretical contribution: higher-order FCMs capture intricate, lagged interactions between social media signals and operational data, while IGWO robustly tunes hyperparameters for non-stationary time series, and DERL adaptively refines ensemble forecasts. By detailing how social media data can be leveraged to achieve substantial performance gains, we extend existing knowledge on social media–driven forecasting strategies and provide actionable guidelines for operations managers aiming to mitigate the costs and complexities of consumer returns.