ADS1002 Group Project · Final Presentation · Monash University
Can historical sales data support a reliable four-week sales plan?
Forecasting Rossmann sales: why it matters → focus the data → test fairly → plan four weeks.
Presenters and roles
| Member | Slides | Responsible part |
|---|---|---|
| Li Yiming | 1–2 | Business context, data and research questions |
| Lim Ying Yu | 3–4 | Data wrangling and store statistics |
| Tay Chen Zhe | 5–6 | Calendar patterns, customers and promotions |
| Huang Qingyue | 7–8 | Forecast design and model results |
| Yanyu Zhu | 9–10 | Scenarios, recommendations and next steps |
Five presenters, two slides each, one connected story from data to decision.
Slide 2 · Speaker: Li Yiming
Business problem, data and four research questions
Start with three things: the business decision, the available data, and what can be known at forecast time.
Why forecasting matters
- Inventory planning
- Staffing decisions
- Budgeting and short-term resource allocation
- Promotion-aware demand planning
Four questions guide the project
Understand the data first, then test a four-week forecast under realistic information limits.
Data: the Rossmann training file named in the project brief (Kaggle Rossmann Store Sales). No other competition files are used.
Slide 3 · Speaker: Lim Ying Yu
Data wrangling and business meaning
No missing cells does not mean a complete calendar.
Quality checks
How zeros and gaps are handled
- Dates parsed, weekdays checked, records sorted by store and date
- Holiday codes kept as categories
- Closed-day zeros retained as real zeros
- 7 selected open-day zero records retained (cause unknown)
- Missing dates never converted to zero demand
Missing cell, missing date and zero Sales are not the same thing.
Slide 4 · Speaker: Lim Ying Yu
Sales differences across the selected stores
Even among the top-selling stores, daily sales levels differ a lot.
Distribution of 223 store means
Mean Sales per store (thousand units) against number of stores. Most stores sit at the lower end; a few sell far more.
Use store-specific history: one network average is not enough.
Slide 5 · Speaker: Tay Chen Zhe
Calendar patterns and coverage
Sales move over time, and the same broad pattern appears with or without shorter calendars.
Weekday means
December peaks occur in both series. Coverage alone is unlikely to explain the broad movement.
Slide 6 · Speaker: Tay Chen Zhe
Customer and promotion associations
Customers strongly track realised Sales; Promo can be planned in advance.
Mean open-day Sales
No promotion 9,016 vs promotion 12,161 units.
Realised Customers are unavailable for future forecasts. Promotion differences remain associations.
Slide 7 · Speaker: Huang Qingyue
Forecast design: how we tested models fairly
Use only information available at each forecast date, then keep July untouched until the end.
Four 28-day horizons (select one)
Leakage controls
- Re-rank stores using past data at each origin
- Build profiles from earlier Sales only
- Fit transformations on training rows only
- Freeze profiles for the full 28-day horizon
Information allowed at forecast time
- Historical Sales profiles
- Weekday, month and trend
- Known Open, Promo and holiday schedules
- Realised future Customers excluded
Models compared: historical profile, OLS, Ridge regression, decision trees and random forests. Settings chosen by RMSE, with MAE also checked.
Try it: which inputs are allowed at the August origin (31 July)?
Choose models on earlier windows. Use July only once as an independent test.
Slide 8 · Speaker: Huang Qingyue
The forest wins validation, but loses in July
A small 2.2% validation edge does not survive the independent July test.
Eight candidates ranked by mean RMSE over three 28-day validation windows (lower is better). Select a model to highlight it, then reveal the July test.
Validation RMSE: forest 1,335 vs profile 1,366 (−2.2%). Validation MAE: profile 882 vs forest 910.
Slide 9 · Speaker: Yanyu Zhu
August scenarios: promotion schedules change the plan
August outcomes are unavailable, so these are conditional planning scenarios, not accuracy claims.
Pick a scenario, or toggle weekdays to build your own. Each day shows the forest’s predicted Sales (million units, 223 stores) with or without promotion that day.
These scenarios support planning; they do not prove causal campaign returns.
Slide 10 · Speaker: Yanyu Zhu
Recommendations: what managers should do next
July supports a simple reference model, confirmed business inputs, and monitored use.
1. Confirm the business inputs
WHYForecast totals depend on Open, Promo and holiday inputs.
DOReplace assumptions with confirmed schedules.
2. Historical profile as the reference
WHYThe forest wins validation but loses the July test.
DOUse the profile as reference; keep the random forest as challenger.
3. Monitor errors by store and period
WHYLarge misses and promotion bias affect operations (forest bias on promoted open days: +731 units).
DOMonitor by store and period; compare the business cost of over- vs under-forecasting.
A useful four-week forecast needs realistic inputs, fair testing and continuous monitoring.
Questions welcome. Open “Q&A extras” for prepared answers.