Demand Forecasting & Inventory Optimization — Cut Simulated Stockouts from 20% to 11% The Problem

A recurring pain point for retail and e-commerce businesses: how much stock to hold per product, and when to reorder — without either running out (lost sales) or over-ordering (wasted capital). I built this project end-to-end on a real 2-year, 1M+ row UK online retail transaction dataset to show how that decision can be made with data instead of guesswork.

What I Did

Cleaned 1M+ raw transactions: resolved cancellations, removed non-product entries, and separated one-off wholesale bulk orders from genuine recurring retail demand — a distinction that changed the shape of the demand data significantly.

Classified 1,760 products by revenue impact (ABC/Pareto analysis) and demand behavior (smooth, erratic, intermittent, lumpy — using the Syntetos-Boylan method from supply-chain statistics).

Backtested 9+ forecasting models — from simple moving averages to Croston's Method (for intermittent demand) to a global LightGBM model with quantile regression — using proper rolling-origin backtesting across 3 time periods, including the toughest case: the Q4 holiday demand surge.

Built a full inventory policy: safety stock and reorder points using both classical and non-parametric (quantile-based) methods, simulated against real held-out demand.

Delivered an interactive dashboard (built in HTML/JS, deployed live) so a non-technical stakeholder can explore any product's forecast, view inventory recommendations, and see model performance — no spreadsheet required.

Key Results

Identified that 21.6% of products drive 80% of revenue (classic Pareto effect) — informing where forecasting effort should focus.

Found that unrepeated wholesale bulk orders were distorting demand patterns for a subset of products, and built a cleaning method to separate wholesale noise from real retail demand.

Compared machine learning (LightGBM) against classical time-series methods honestly — reporting where ML wins, where it doesn't, and why (a well-tuned moving average performed comparably at the individual-product level).

Simulated the recommended inventory policy against real Q4 demand: cut stockout weeks from ~20% (naive rule) to ~11%, while explicitly weighing that against a fair (matched-service-level) comparison to a classical method — an honest, non-oversold conclusion.