Demand Forecast & Seasonality
Decompose your sales forecast into Trend, Seasonality, and Promotion components. Estimate true incremental uplift by accounting for cannibalization.
Formula
Forecast = (Base × Trend × Seasonality) × (1 + PromoLift)
We use a classical multiplicative decomposition model. We start with the Baseline, apply the Trend (Growth), multiply by the Seasonality Index (e.g., 1.2 for holiday), and finally apply the Promotional Lift. We then subtract Cannibalization to find True Incremental Sales.
Worked Examples
Example 1: Holiday Sale
Problem:Base 1000, 1.5 Seasonality (Dec), 50% Promo Lift
Solution:Seasonal Base: 1500. Promo: 1500 * 1.5 = 2250.
Result:2250 Forecasted Units
Example 2: Summer Slump
Problem:Base 1000, 0.8 Seasonality (Aug), 0% Promo
Solution:Seasonal Base: 1000 * 0.8 = 800.
Result:800 Forecasted Units
Frequently Asked Questions
What is Seasonality?
Predictable, recurring fluctuations in demand (e.g., Ice cream in summer, Toys in December). It is usually expressed as an index (1.0 = average, 1.2 = 20% above average).
What is Trend?
The underlying direction of the data over time (Growth or Decline), ignoring seasonal bumps. If you grow 10% YoY, your Trend factor is 1.1.
What is Promo Lift?
The percentage increase in sales driven *specifically* by a marketing activity or discount, above the expected baseline.
What is Cannibalization?
When a promo steals sales from: 1) Future periods (Forward Buying - customers stocking up), or 2) Other products in your portfolio (Switching).
How do I calculate Baseline?
Look at sales during 'normal' periods without promos or holidays. Or use a moving average of the last 12 months to smooth out noise.
Multiplicative vs Additive?
Multiplicative models (Base * Season) are better for growth companies where seasonal spikes grow with the business. Additive (Base + Season) is better for stable, flat volumes.
What is 'True Incremental'?
Sales you wouldn't have gotten otherwise. If you sell 100 units on promo, but 20 were borrowed from next week, your True Incremental is 80.
Does this handle 'Halo Effect'?
No, Halo Effect is when a promo on Product A drives sales of Product B (e.g., cheap printer drives ink sales). That is the opposite of Cannibalization.
Why forecast decomposition?
It helps you know *why* sales happened. Was it the ad campaign? Or just Christmas? If you confuse Seasonality with Promo Lift, you'll overspend on ads.
How do I forecast revenue?
Bottom-up forecasting multiplies expected units sold by price. Top-down starts with market size and estimates market share. For existing businesses, use historical growth rates with adjustments. For SaaS: Forecast MRR = Current MRR + New MRR - Churned MRR + Expansion MRR. Always model best, expected, and worst case scenarios.
Background & Theory
The Decomposition Model
Sales = Baseline × Trend × Seasonality × Event
- Baseline: What you sell on a boring Tuesday.
- Trend: Is the brand growing or shrinking?
- Seasonality: The calendar effect (Day of week, Month of year).
- Event: The promo, price cut, or TV spot.
The Cannibalization Trap
Promotions often look profitable on the surface but destroy value hidden elsewhere.
- Temporal Cannibalization: "Pantry Loading." I buy 3 shampoos because they are 50% off. I won't buy shampoo again for 6 months. Sales didn't increase; they just moved.
- Cross-Product Cannibalization: I buy the discounted Nike shoes *instead* of the full-price Adidas shoes. Store revenue might actually drop.
Practical Tips
- Clean Your Data: Remove out-of-stock periods from your history. If you had no stock, sales were zero, but demand was not.
- Use Indices: Create a "Seasonal Index" for every week of the year. Week 52 might be 2.5x. Week 5 might be 0.8x.
- Post-Event Analysis: After every promo, calculate the "Lift." Save this variable to predict the next one.
History
The Almanac Era
Farmers were the first demand forecasters. Using almanacs, they predicted seasonal weather patterns to plan crops. Retailers adopted this "Calendar-based" planning in the 19th century—stocking coats in winter and swimsuits in summer.
Time Series Analysis
In the 1920s, economists began decomposing time series data into Secular Trend, Seasonal Variation, Cyclical Fluctuation, and Irregular Variation. The Holt-Winters method (1960) added exponential smoothing, allowing algorithms to "learn" seasonality from history automatically.
Machine Learning & Causal AI
Today, retailers like Walmart and Amazon use Gradient Boosted Trees and Deep Learning (LSTM/Transformer) models. These don't just look at history; they look at "features" (Weather, Price, Competitor Actions, Local Events). However, the fundamental concept of "Baseline vs. Uplift" remains the core language of merchandising.
Common Misconceptions
- Myth: "We sold more, so the promo worked." Reality: You might have sold more because it was Payday or Black Friday. You must strip out seasonality to measure promo effectiveness.
- Myth: History repeats exactly. Reality: COVID-19 broke all seasonal models. Modern forecasting requires "intervention" adjustments.