🍱📈 A Data-Driven Approach for Decision Making in Food Combo Offer

Authors
Kaibo Zhang, Mingshu Liu, Yanfei Wu, David Min, and Charlie Tomlinson
Affiliation: McGill University (RETL 407 - Retail Analytics)

---

🧠 Overview

This project presents a full-stack data-driven framework to optimize multi-item food combos for a leading convenience store chain in Quebec. We integrated transaction data analysis, demand forecasting, optimization modeling, and survey-based validation to recommend relevant, profitable, and nutritionally coherent bundles.

Key contributions:

• 📊 Extracted frequent co-purchase patterns using Association Rule Mining.
• 🔮 Modeled demand with price memory, cross-price elasticity, and seasonality using NeuralProphet.
• 🔧 Implemented a Linear Programming model to optimize combos under real-world constraints.
• 🧪 Conducted external validation via survey with Net Promoter Score (NPS) analysis and NLP comment mining.
• 📦 Proposed final combo structures balancing consumer affinity and margin.

---

📁 Directory Structure

├── phase1/ # Code used for initial itemset mining and exploratory combo design
├── phase2/ # Demand modeling scripts including NeuralProphet and regression models
├── phase3/ # Linear programming formulation and combo optimization code
├── price_sensitivity/ # Category-level price elasticity analysis with brand comparisons
├── result/ # Raw combo outputs from optimization results
├── validation/ # Survey design, NPS analysis, and NLP-based feedback evaluation

---

🛒 Dataset Description

🔹 Internal Data

• Source: POS system (Jan 2021 – Feb 2025)
• Details: Over 5,000 items, store/site identifiers, price, quantity sold, transaction timestamp
• Metadata: Brand type (private/national), procurement cost, internal co-purchase affinity

🔹 External Data

• Consumer Survey: 404 valid Canadian responses via Prolific
• Task: Simulated shopping + combo rating (0–10)
• Goal: Validate proposed bundles using stated preferences

---

🧪 Methodology

1. Association Rule Mining (Apriori)

• Reduced the item space to top-performing food & drink SKUs
• Identified co-purchase patterns to guide combo design

2. Predictive Demand Modeling

• Regression + NeuralProphet-based hybrid model:
  • 🔁 Lagged price (past 3 days)
  • 🔗 Cross-product price effects (correlation ≥ 0.6)
  • 📆 Seasonal/weekly trends
• RMSE ≈ 247 units; strong performance on trend capture

3. Linear Programming Optimization

• Objective: maximize α·profit + (1–α)·bundle relevance
  • α = 0.6 (margin prioritized)
• Constraints:
  • 🥗 Nutrition (1 food, ≤2 beverages)
  • 🏷️ No same-category duplication
  • 🏭 Supplier conflict exclusion
  • 💵 At least one discounted item per combo

4. Price Elasticity Analysis

• Log-log regressions across categories (private vs. national brands)
• Highest sensitivity: energy drinks, salty snacks
• Price-insensitive: candy, soft drinks (private label)

5. External Survey & NPS Validation

• NPS computed from ratings (Promoters 9–10, Detractors 0–6)
• Best NPS combo: 🧀 Fruit & Cheese + 🥪 Steak Panini (NPS = –27.5)
• Worst NPS combo: 🧁 Muffin + 🌭 Hot Dog + ☕ Coffee + 🥤 Red Bull (NPS = –77)
• Key insight: less is more — smaller, familiar combos outperform overstuffed ones

---

📈 Results Summary

Combo	Items	Price	NPS
Balanced Combo (Best)	Fruit & Cheese + Panini	$11.59	–27.47
Control Combo (Legacy)	Baguette, Lay’s, Pepsi, Eska	$12.75	–32.43
Overloaded Combo (Worst)	Muffin, Hot Dog, Red Bull, Coffee	$10.52	–77.00

---

💡 Insights & Recommendations

• Simple Combos Win: 2–item combos maximize appeal and align with grab-and-go behavior.
• Avoid Redundancy: Multiple drinks or overstuffed bundles confuse and deter consumers.
• Target Price Zone: $5–$15 hits the sweet spot for willingness to pay.
• Cater to Vegetarians: Plant-forward combos scored better among non-meat consumers.
• Strategic Discount Framing: Highlight the discounted drink to boost perceived value.

---

🔄 Future Work

• A/B test optimized combos in select stores to validate impact
• Develop seasonal or regional combo variants
• Integrate loyalty card data to personalize recommendations
• Use NLP (e.g., BERT) for deeper comment sentiment analysis

---

📌 References

• [1] Cohen & Perakis. "Promotion Optimization in Retail", SSRN 2018
• [2] Triebe et al. "NeuralProphet: Explainable Forecasting at Scale", arXiv 2021

---

🙏 Acknowledgments

We would like to express our sincere gratitude to the partner company for their invaluable support and collaboration throughout this project.

We are also grateful to Ms. Nymisha Bandi, Ph.D. and teaching assistant, for her thoughtful guidance and constructive feedback during every stage of the project.

Finally, we deeply appreciate the supervision and mentorship of Professor Saibal Ray, whose expertise and leadership in operations management helped shape our analysis and approach.

---