Hybrid Fusion for Retail Demand Forecasting & Customer Segmentation
Abstract
A hybrid machine learning framework combining LSTM, GRU, and Random Forest with early and late fusion to forecast retail demand, optimize inventory, and segment customers for personalized marketing.
Introduction
Knowing what will sell, keeping the right stock on shelves, and understanding different types of shoppers are at the heart of a profitable retail business. Software can learn from past sales, prices, promotions, seasons, and other signals to predict demand and group customers for smarter marketing.
This project brings together three related goals, forecasting demand, planning inventory, and grouping customers, into one connected approach.
Problem Statement
Getting retail decisions right is hard for a few reasons:
- Demand shifts with price changes: promotions: holidays, weather, and competitors.
- Too much stock ties up money; too little means lost sales and unhappy customers.
- Shoppers behave very differently: so one-size-fits-all marketing wastes effort.
- The useful signals are scattered across many sources and need to be pulled together.
Objectives
- Forecast future sales and stock needs from past data and outside signals.
- Group customers by how they shop: to enable targeted marketing.
- Turn those groups into tailored promotions that fit each one's preferences.
Dataset
The work is built around a set of retail records covering sales, stock, pricing, and outside signals like promotions and seasonality. Depending on the client, this can be a public retail dataset, licensed or private point-of-sale data, information from other trusted sources, or a custom set assembled for the business. The data is cleaned, standardised, and enriched with useful derived measures before modelling.
Methodology
The project prepares the retail data, adds helpful derived measures, then trains complementary models, Random Forest for structured tables, LSTM and GRU for patterns over time, with early and late fusion to combine their strengths, plus K-Means to group customers by behaviour.
The novelty is in blending these approaches so their strengths combine: table-friendly models, time-aware models, and customer grouping feed into one connected view. Rather than forecasting, stocking, and marketing in separate silos, the design links them, turning scattered signals into joined-up decisions a retailer can act on.