Forecast-Driven Inventory for Coffee Vending

IBM 6520 • Cal Poly Pomona

Min Gong • Jarrod Griffin • Ceren Unal • Eunice Won

Agenda

  1. Business Problem
  2. Data & Assumptions
  3. Exploratory Insights
  4. Modeling Approach
  5. Forecast Results
  6. Recommendations

1 · Business Problem

Pain points

  • Stock-out → lost sales
  • Overstock → spoilage + tied-up cash
  • Decision horizon 8 weeks

Goal → forecast cups & ingredients for a coffee vending machine.

2 · Data & Assumptions

  • Kaggle data: 3,637 transactions (Mar 2024 → Mar 2025)
  • 8 drink SKUs → decomposed into 5 ingredients
  • Aggregated to weeks (matches reorder cadence)
  • Incomplete week 23 Mar 2025 removed

Weekly Cups vs Ingredients

Milk drives most of the volatility

Revenue Trend

Weekly revenue (proxy)

3 · Exploratory Insights

  • Revenue shows “three-hill” trend, no strong seasonality
  • Americano w/ Milk + Latte = 55 % of cups
  • Milk is the key stock-out risk (largest variance)

4 · Modeling Approach

  • Ingredients: auto-selected SARIMA (fable::ARIMA) — Ljung-Box p > 0.05
  • Coffee types: ARIMA on unit sales (auto differencing)
  • Revenue: ETS (A,N,N) → handles trend & changing variance
  • Forecast horizon h = 8 weeks

5 · Forecast Results

Ingredients (SARIMA)

Coffee Types (ARIMA)

Revenue (ETS)

Stable short-term revenue outlook

6 · Recommendations

  1. Prioritise milk — carry +20 % buffer Nov-Feb
  2. Keep Americano + Latte always in stock
  3. Rotate low sellers (Cortado, Espresso) to free space
  4. Automate coffee-ground reorders (low variance)
  5. Refresh models quarterly as tastes evolve

Q & A

Thank you

IBM 6520 • Spring 2025