A Data-Driven Forecasting Approach for New Product Developments: Addressing Cold-Start Demand and Product Transitions

 Overview

Company Description

Mars is an American multinational manufacturer of confectionery, pet food, and other food products and a provider of animal care services. Mars is headquartered in McLean, Virginia and is entirely owned by the Mars family.

In Veghel, Mars operates the largest chocolate factory in the world. It’s where products like Mars, Snickers, Twix, Milky Way and many others are produced. Also, Veghel is the logistics hub for chocolate products in Europe, as well as the regional supply chain office.

Project Description

New Product Developments (NPDs) present a significant challenge in supply chain forecasting due to limited or missing representative historical demand data. Within this context, two distinct types of NPDs can be identified:
•True NPDs, where entirely new products are introduced without any direct sales history (cold-start problem)
•Product switchers, where new products replace or succeed existing ones, requiring accurate modelling of demand transfer and phase-in/phase-out dynamics

Both types introduce different forecasting challenges. True NPDs rely on indirect signals such as similar product performance, market trends, and external drivers. Product switchers, on the other hand, require estimating cannibalization effects and managing demand shifts between outgoing and incoming products.

Traditional forecasting models, such as ARIMA or exponential smoothing, depend on historical patterns and are therefore not well-suited to either scenario. More advanced approaches—including machine learning, similarity-based modelling, and hybrid methods—offer the potential to better address these challenges by leveraging proxy data, capturing product relationships, and incorporating external drivers.

This thesis explores a data science-driven approach to improve forecast accuracy for both true NPDs and product switchers, while also increasing visibility into forecast accuracy and bias over time from early planning volumes till after go-live.

Goal of the Project

• To develop forecasting approaches for true NPDs (cold-start problem) using similarity learning and proxy data from existing products.
• To model product transitions (switchers), including phasing in/out effects and demand cannibalization.
• To analyse and visualize forecast accuracy and bias evolution of early planning volumes till after go-live.
• To benchmark the developed methods against current business forecasting performance, quantifying improvements in accuracy, bias, and service impact.
• To design a scalable and reusable data science pipeline for NPD forecasting in supply chain contexts.

Deliverables

•A benchmark study comparing forecasting approaches for:
oTrue NPD (cold-start) scenarios
oProduct switch (phase-in/phase-out) scenarios
oCurrent business forecasting methods vs. proposed models

•Quantitative evaluation of model performance, including:
oAccuracy and bias metrics (e.g., MAPE, WAPE, forecast bias)
oImpact on service-related KPIs (e.g., stockouts, NQCs if available)

•A framework for similarity-based forecasting, identifying which legacy products best inform new product demand.
•A reproducible end-to-end forecasting pipeline, including:
oData preprocessing
oFeature engineering
oModel selection and evaluation

•A data visualization layer/dashboard showing:
oForecasts, Accuracy and Bias of (early) planning volumes over time
oForecast evolution from launch to maturity
oComparison between business forecasts and model outputs

•Business recommendations for:
oImproving NPD forecast processes
oModel selection and deployment
oReducing service issues and improving planning reliability

Essential student knowledge

•Time Series Forecasting Fundamentals.
•Machine Learning/AI for Time Series.
•Strong coding skills (preferably Python).
•Basic understanding of supply chain planning concepts (preferred)

More information: escf@tue.nl