Statistical forecasting sounds intimidating. It's not. At their core, these methods are just systematic ways of using your sales history to predict future sales. The math has been packaged into tools—you don't need to derive equations.
What you do need is understanding of which method works for which situation.
Method 1: Simple Moving Average
The most intuitive method. Average recent periods to predict the next period.
How It Works
Formula: Forecast = Average of last N periods
Example (4-week moving average):
- Week 1: 100 units
- Week 2: 110 units
- Week 3: 95 units
- Week 4: 105 units
- Forecast for Week 5: (100 + 110 + 95 + 105) / 4 = 102.5 units
Strengths
Simplicity: Anyone can calculate and understand it.
Smoothing: Averages out week-to-week volatility.
No assumptions: Doesn't require fitting complex models.
Weaknesses
No trend capture: If demand is consistently increasing, a moving average lags behind.
No seasonality: Treats all periods equally, even if December is different from July.
Lookback choice: How many periods to include? More smoothing vs. more responsiveness is a tradeoff.
When to Use It
Best for: Stable demand products without strong trends or seasonality. Short-term forecasting where you want smoothing.
Avoid for: Growing or declining products, highly seasonal items.
Method 2: Weighted Moving Average
A variation that gives more weight to recent periods.
How It Works
Assign weights that sum to 1, with higher weights on recent periods.
Example (4-week weighted average):
- Week 1: 100 units × 0.10 = 10
- Week 2: 110 units × 0.15 = 16.5
- Week 3: 95 units × 0.25 = 23.75
- Week 4: 105 units × 0.50 = 52.5
- Forecast for Week 5: 10 + 16.5 + 23.75 + 52.5 = 102.75 units
Strengths
Responsiveness: Reacts faster to recent changes than simple average.
Flexibility: You choose the weights based on your judgment.
Weaknesses
Weight selection: What weights are best? Often determined by trial and error.
Still no seasonality: Like simple moving average, treats time periods as equivalent.
When to Use It
Best for: Products where recent history is more relevant than older history, but you're not ready for more complex methods.
Method 3: Exponential Smoothing
A sophisticated approach that automatically weights recent data more heavily.
How It Works
Formula: Forecast = α × (Last Actual) + (1 - α) × (Last Forecast)
Where α (alpha) is a smoothing parameter between 0 and 1.
Example (α = 0.3):
- Last actual: 110 units
- Last forecast: 100 units
- New forecast: 0.3 × 110 + 0.7 × 100 = 33 + 70 = 103 units
What Alpha Means
High α (e.g., 0.5): Responsive—forecast reacts quickly to changes.
Low α (e.g., 0.1): Smooth—forecast changes slowly, less sensitive to noise.
Most tools find optimal alpha automatically based on your data.
Variants
Simple exponential smoothing: As described above. No trend or seasonality.
Holt's method (double exponential smoothing): Adds trend component. Good for products with consistent growth or decline.
Holt-Winters (triple exponential smoothing): Adds both trend and seasonality. The workhorse method for seasonal CPG products.
Strengths
Adaptive: Automatically adjusts based on recent performance.
Handles trend: Holt's method captures growth or decline.
Handles seasonality: Holt-Winters incorporates seasonal patterns.
Widely available: Built into Excel and almost every forecasting tool.
Weaknesses
Initialization: Needs sufficient history to calibrate parameters.
Assumes continuity: Doesn't handle structural breaks (major changes in demand pattern).
When to Use It
Simple: Stable products with no trend or seasonality.
Holt's: Products with clear growth or decline trend.
Holt-Winters: Seasonal products with predictable peaks and valleys.
Method 4: Seasonal Decomposition
Breaks down your time series into components: trend, seasonality, and residual.
How It Works
- Calculate seasonal indices (how each period compares to average)
- Remove seasonality to see underlying trend
- Forecast the trend
- Apply seasonal indices to the trend forecast
Example:
- Your deseasonalized trend forecast for December: 1,000 units
- December seasonal index: 1.8
- December forecast: 1,000 × 1.8 = 1,800 units
Strengths
Transparency: You can see and understand each component.
Handles strong seasonality: Explicitly models seasonal patterns.
Adjustable: You can override specific components based on judgment.
Weaknesses
Requires history: Needs at least 2 years of data to calculate seasonal indices reliably.
Assumes stable seasonality: If seasonal patterns are shifting, historical indices may be wrong.
When to Use It
Best for: Products with strong, consistent seasonal patterns and sufficient history.
Method 5: Linear Regression
Models demand as a function of time and other variables.
How It Works
Simple form: Demand = a + b × Time
With multiple variables: Demand = a + b₁×Time + b₂×Promo + b₃×Price + ...
Strengths
Interpretable: Coefficients tell you how much each factor affects demand.
Flexible: Can incorporate multiple drivers beyond just time.
Handles trend: Linear trend is explicitly modeled.
Weaknesses
Assumes linearity: Real demand relationships aren't always linear.
Requires variable data: To include price, you need price variation in your history.
Overfitting risk: Too many variables can fit history perfectly but forecast poorly.
When to Use It
Best for: When you understand what drives demand and have data on those drivers. Useful for modeling price elasticity or promotional response.
Choosing the Right Method
Start Simple
If you're new to statistical forecasting, start with simple moving averages or basic exponential smoothing. Get comfortable with the process before adding complexity.
Match Method to Data
| Product Characteristics | Recommended Method |
|------------------------|-------------------|
| Stable, no trend, no seasonality | Simple moving average or simple exponential smoothing |
| Clear growth or decline trend | Holt's exponential smoothing |
| Strong seasonality | Holt-Winters or seasonal decomposition |
| Multiple known demand drivers | Regression |
Let Tools Help
Modern inventory planning tools like Planster select and fit methods automatically. They test multiple approaches against your data and choose what works best for each SKU. You don't need to manually select methods for 500 SKUs.
Don't Over-Engineer
A simpler method that you understand and maintain beats a complex method that sits untouched. Forecasting accuracy comes more from regular review and adjustment than from sophisticated algorithms.
Common Pitfalls
Using Too Much History
Older data may no longer reflect current demand patterns. For most products, 2-3 years of history is sufficient. More than that can drag in outdated patterns.
Ignoring Stockout Periods
Statistical methods treat low-sales periods as low demand. If you were stocked out, that period doesn't represent true demand. Exclude or adjust it.
Not Validating
Test your method on historical data before trusting it for future forecasts. Hold out the most recent few months, forecast them, and see how close you get.
Trusting Blindly
Statistical forecasts are a starting point. Apply judgment for events the model can't see—promotions, new distribution, market changes.
Key Takeaways
- Moving averages are simple and effective for stable products
- Exponential smoothing adapts to recent data and handles trends
- Holt-Winters is the go-to method for seasonal CPG products
- Seasonal decomposition makes seasonal patterns transparent and adjustable
- Regression helps when you have identifiable demand drivers
- Start simple, match method to product characteristics, and validate before trusting
- Use tools that select methods automatically when managing many SKUs
Frequently Asked Questions
Q: Which method is best for CPG products?
Holt-Winters (triple exponential smoothing) works well for many CPG products because it handles both trend and seasonality. But "best" depends on your specific products—stable items might do fine with simpler methods.
Q: Do I need to do the math myself?
No. Excel has built-in forecasting functions (FORECAST.ETS). Planning tools like Planster handle method selection and calculation automatically.
Q: How much history do I need?
At minimum, 12 months to capture seasonality. 24-36 months is better. Beyond that, older data may be less relevant than recent patterns.
Q: What if my data is messy?
Clean it first. Stockout periods, anomalous spikes, data errors—fix these before applying statistical methods. Garbage in, garbage out.
Q: Should I use the same method for all products?
Not necessarily. Different products have different characteristics. Your hero SKU with strong seasonality might need Holt-Winters while a stable everyday item works fine with moving averages.