How do I set up sales forecasting software to simulate promo and seasonality impact across SKUs?

To set up sales forecasting software to simulate promo and seasonality impact across SKUs, you want to build a forecasting workflow that can handle:

1. baseline demand
2. promo uplift / cannibalization / halo
3. seasonality
4. SKU-level differences
5. scenario simulation

Here’s a practical setup.

---

1) Define the forecasting objective

Decide what the model should output:

• Forecast units, revenue, or both
• By SKU, store, region, channel, or customer segment
• At what granularity: daily, weekly, monthly
• Horizon: next 4 weeks, next 13 weeks, 12 months, etc.

For promo simulation, weekly SKU-level forecasting is usually a good start.

---

2) Organize the data you need

At minimum, collect:

Historical sales

• Date
• SKU
• Units sold
• Revenue
• Price
• Channel / store / region if relevant

Promotion history

• Promo flag
• Promo type: discount, BOGO, feature/display, coupon, bundle
• Promo depth / discount %
• Promo start/end dates
• Promo spend if available

Seasonality drivers

• Holidays
• Month / week-of-year
• Payday effects if relevant
• Weather, school calendar, events, etc. if relevant to category

Product hierarchy

• SKU → brand → category → subcategory

Supply / stock data

• Stockouts can distort demand, so record:
• On-hand inventory
• Fill rate
• Lost sales if available

---

3) Clean the data before modeling

This part matters a lot.

• Remove or flag stockout weeks
• Fix missing sales and price records
• Standardize SKU IDs across systems
• Separate real zero demand from out-of-stock zero sales
• Detect extreme outliers from data issues or one-time events

If you skip this, promo effects and seasonality will be distorted.

---

4) Build a baseline demand model

Start with a model that predicts normal demand without promo effects.

Common baseline approaches:

• ARIMA / SARIMA
• Prophet
• Regression with time features
• XGBoost / LightGBM
• Hierarchical time-series models
• Deep learning for large SKU portfolios

For most commercial forecasting setups, a machine learning regression model with lag features works well:

Example baseline features

• Lagged sales: t-1, t-4, t-52
• Rolling averages: 4-week, 8-week, 13-week
• Price
• Holiday flags
• Month / week-of-year
• Trend indicators
• SKU attributes

---

5) Add promo uplift modeling

To simulate promotions, don’t just use a promo flag. Model the incremental effect.

Useful promo features

• Promo binary flag
• Discount %
• Promo depth buckets
• Promo type
• Display / feature flags
• End-of-promo dummy
• Days since promo start
• Days before promo end

Optional effects to model

• Uplift: sales increase during promo
• Pre-buying: demand pulled forward before promo ends
• Post-promo dip: sales suppression after promo
• Cannibalization: one SKU steals sales from another
• Halo: promo lifts related SKUs

If you have enough history, model these separately or at least include lagged promo variables.

---

6) Model seasonality explicitly

Seasonality should be captured with:

• Week-of-year
• Month
• Quarter
• Holiday proximity
• Moving holiday indicators
• Annual seasonal indices
• Category-specific seasonality patterns

For SKUs with sparse history, borrow strength from:

• category-level seasonality
• brand-level patterns
• similar SKU clusters

---

7) Choose a modeling structure for multi-SKU forecasting

If you have many SKUs, use one of these:

Option A: One model per SKU

Good if:

• Few SKUs
• Very different demand patterns

Downside:

• Hard to maintain
• Weak for sparse SKUs

Option B: Global model across all SKUs

Good if:

• Many SKUs
• Want consistent promo effects
• Need better performance on sparse items

This is often best for SKU portfolios.

Option C: Hierarchical model

Good if:

• Need forecasts to roll up correctly from SKU to category to total

---

8) Set up scenario simulation

This is the key part for promo planning.

Your software should allow you to change inputs such as:

• Promo discount %
• Promo dates
• Promo type
• Price
• Holiday calendar
• Stock availability assumptions

Then it should output:

• Expected sales under scenario A
• Expected sales under scenario B
• Incremental lift vs baseline
• Margin impact
• Inventory implications

Example scenario workflow

1. Forecast baseline demand with no promo
2. Apply promo assumptions to the same model
3. Compare results
4. Evaluate:
   • unit uplift
   • revenue
   • gross margin
   • cannibalization
   • inventory risk

---

9) Include elasticity if possible

If price changes matter, estimate:

• price elasticity
• promo elasticity
• cross-SKU elasticity

This helps answer:

• How much does a 10% discount lift demand?
• Which SKUs are substitutes?
• Does discounting one SKU reduce sales of another?

---

10) Validate with backtesting

Before trusting the model, test it on historical periods.

Backtest setup

• Train on past data
• Predict on held-out weeks
• Compare actual vs forecast during:
  • promo weeks
  • non-promo weeks
  • peak seasons
  • holidays

Metrics

• WAPE / MAPE
• RMSE
• Bias
• Promo uplift accuracy
• Forecast accuracy by SKU tier

Pay special attention to whether the model:

• underestimates peak season
• overstates promo lift
• misses post-promo dips

---

11) Build dashboards and controls

A good forecasting tool should show:

• Baseline forecast vs promo-adjusted forecast
• SKU-level uplift curves
• Seasonal patterns by SKU
• Forecast confidence intervals
• Exception alerts for stockouts or abnormal demand

Useful controls:

• Promo calendar editor
• Discount slider
• SKU selection
• Channel filter
• Category roll-up

---

12) Operationalize the workflow

Automate:

• daily/weekly data refresh
• model retraining
• forecast generation
• scenario reruns when promo plans change
• alerting for forecast drift

Typical pipeline:

1. Ingest data
2. Clean and enrich
3. Generate features
4. Forecast baseline
5. Apply promo scenarios
6. Publish to planning system
7. Monitor error and retrain

---

Recommended software capabilities

If you’re choosing or configuring software, look for these features:

• SKU-level forecasting
• External regressors
• Promo and price modeling
• What-if scenario simulation
• Hierarchical forecasting
• Inventory-aware forecasting
• API / spreadsheet integration
• Explainability of promo impact
• Automatic retraining and backtesting

---

Simple example of the modeling logic

A common structure is:

Sales = baseline trend + seasonality + promo uplift + price effect + noise

Where:

• baseline trend = underlying demand growth/decline
• seasonality = predictable calendar pattern
• promo uplift = incremental effect from promotion
• price effect = impact of price changes
• noise = random variation

Then for a scenario, you adjust promo/price inputs and re-run the forecast.

---

Best practice starting point

If you want something practical and effective:

1. Start at weekly SKU-level
2. Use a global ML model
3. Include:
   • lagged sales
   • price
   • promo flags and depth
   • holiday/seasonality features
4. Backtest promo and non-promo periods separately
5. Add scenario simulation for discount changes and promo timing
6. Layer in cannibalization/halo later if needed

---

If you want, I can also give you:

• a step-by-step implementation checklist
• a sample data schema
• or a Python example for building promo + seasonality forecasts across SKUs.