Loading...

Integrating Customer Behavior Analysis into Demand Forecasting for Fast-Moving Consumer Goods in Retail Chains

Ghaed Rahmati, Elahe | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58465 (01)
  4. University: Sharif University of Technology
  5. Department: Industrial Engineering
  6. Advisor(s): Hassan Nayebi, Erfan
  7. Abstract:
  8. Forecasting demand for fast-moving consumer goods (FMCG) is a fundamental yet challenging issue in retail management due to highly volatile demand, short product life cycles, low profit margins, and limited customer loyalty. Customer purchase behavior reflects their response to a set of concurrent product attributes in the retail environment; price, discounts, product placement, and other stimuli influence the final purchase decision not independently, but in combination and in interaction with related products. Therefore, modeling this behavior realistically requires considering the dynamic interactions among products, and relying solely on univariate time series analysis is insufficient. In this study, a novel framework combining Graph Attention Networks and Long Short-Term Memory networks is proposed, applied for the first time to FMCG demand forecasting. In this approach, each product is represented as a node in a graph, with features such as sales, price, and discounts defined as node attributes. Edge weights between nodes are learned dynamically and data-driven using an attention mechanism, allowing each node to obtain an embedded representation that captures both its own attributes and interactions with other products. These embeddings, alongside the dataset, are then fed into an Long Short-Term Memory networks to extract temporal sales patterns. The proposed model is implemented and compared with a set of commonly used baseline models using real-world data. Initially, the baseline models are augmented with engineered features reflecting the conditions of other products indirectly. Among these baselines, Random Forest shows the best performance but still has limitations. Incorporating graph-based data into model combinations yields improved results, with the highest accuracy achieved by the proposed hybrid model. Comparison with the Random Forest model shows that the mean squared error decreases from 82 to 39, corresponding to a reduction of over 52%. The mean absolute error decreases by more than 29%, and the mean absolute percentage error drops from 68% to 49%. These results indicate that the proposed framework not only captures temporal sales trends more accurately but also accounts for customer behavioral dimensions through modeling dynamic product interactions. Consequently, the proposed model provides more realistic and precise forecasts by combining time series analysis with customer behavior modeling.
  9. Keywords:
  10. Hybrid Methods ; Graph Neural Network ; Fast Moving Consumer Goods ; Consumer Buying Behavior ; Multivariate Time Series Forecasting

 Digital Object List

 Bookmark

No TOC