Every purchasing decision is a bet on future demand. A good forecast doesn't need to be perfect — it needs to be unbiased, regularly updated and connected to buying decisions such as reorder points and order quantities. This guide covers the methods most small and mid-sized teams actually use.
Start with clean history
Forecast from demand, not just sales: add back orders you couldn't fill because of stockouts, and remove one-off events such as a single bulk order or a promotion you won't repeat. Twelve to twenty-four months of weekly or monthly data is enough to start.
Three simple methods
| Method | How it works | Use when |
|---|---|---|
| Naive | Next period = last period | A baseline to beat |
| Moving average | Average of the last n periods | Stable demand, no trend |
| Exponential smoothing | New forecast = α × latest actual + (1 − α) × previous forecast | Most SKUs — responsive yet smooth |
The smoothing constant α (between 0 and 1) controls responsiveness. Around 0.1–0.3 suits most products; higher values react faster but chase noise.
Worked example: exponential smoothing
Monthly demand: January 100, February 120, March 110, April 130. Use α = 0.3 and start with January's actual as the first forecast.
- February forecast = 100
- March forecast = 0.3 × 120 + 0.7 × 100 = 106
- April forecast = 0.3 × 110 + 0.7 × 106 = 107.2
- May forecast = 0.3 × 130 + 0.7 × 107.2 = 114.0
Absolute percentage errors: February 16.7%, March 3.6%, April 17.5%. The mean absolute percentage error (MAPE) is about 12.6%. Every forecast here was below actual, so the method is lagging an upward trend — a sign to try a trend-adjusted method (Holt's double exponential smoothing) or a higher α.
Trend and seasonality
If demand grows steadily, add a trend term (Holt's method). If it repeats a yearly pattern — holidays, weather, back-to-school — use seasonal indices: average each month's demand as a ratio of the yearly average, forecast the deseasonalised level, then multiply back. Holt-Winters combines level, trend and seasonality and is built into most planning tools and spreadsheets.
Measuring accuracy
- MAPE — average absolute percentage error. Easy to explain, but distorted by very low-volume items.
- Weighted MAPE (WMAPE) — total absolute error ÷ total actual demand. Better for comparing across many SKUs.
- Bias — the average of (forecast − actual). Persistent positive bias builds excess stock; persistent negative bias causes stockouts. Bias is often more damaging than random error.
- Forecast value added — compare your forecast's accuracy with the naive method. If manual overrides don't beat it, stop making them.
Connect the forecast to decisions
- Use the forecast as average daily demand in the reorder point calculation.
- Use the standard deviation of forecast error to size safety stock — a better input than raw demand variability.
- Use the annual forecast in EOQ.
- Share the forecast with key suppliers so they can plan capacity, especially before peak season and factory holidays.
Segment first with ABC analysis: spend modelling effort on A items and let simple methods run C items.
Ready to put this into practice?
Post a free RFQ and get competing quotes from vetted suppliers — or run the numbers with our free calculators.