Interactive guide · 11 models, 4 families
Forecasting models compared
Describe a product's demand in three questions. See how each model in a family forecasts it, where each one is strong, and where it lets you down.
1. Describe the demand
Three years of monthly demand built from your answers.
2. The models, on the same history
Actual demandWhat the model forecast at the timeForecast, next 12 months
3. Same family, what is different
Which family to start with
| If the demand looks like this | Start with | Why |
|---|---|---|
| Sells every month, roughly level | Simple exponential smoothing or a moving average | There is no trend or season to model, so averaging out the noise is enough. |
| Sells every month, steadily rising or falling | Holt linear trend | It carries the slope forward. Level-only models lag behind every month. |
| Sells every month, with a yearly cycle | Holt-Winters, or seasonal naive if history is short | Both remember what the same month did last year. |
| Month-to-month changes carry over or bounce back | ARIMA(1,1,0) with drift | It forecasts the change itself, using the previous change. |
| Many months with no orders | SBA | The usual first choice for slow movers. It corrects Croston's tendency to run high. |
| Slow mover that may be dying out | TSB | Its forecast fades when orders stop. Croston and SBA stay where they were. |
| Very few orders, very uneven sizes | ADIDA | Adding months into buckets removes most of the zeros. |
| Any of the above | Naive, as the yardstick | If a model cannot beat "same as last month", it is not adding value. |
These are starting points. The reliable way to choose is to test every model on the product's own history, which is what Demand Bench does.
How this guide works
- The demand history is generated from your three answers. It is an example, not your data.
- Each model forecasts one month ahead, using only the months before it. "Ran high" or "ran low" compares the total of those forecasts with total demand.
- Baselines and exponential smoothing are scored on year 3, because seasonal models need two full years before they can start. The autoregressive and sparse-demand families are scored on years 2 and 3.
- "Closest" combines the size of the misses with how far the model ran high or low.
- Settings are fixed at common textbook values so the comparison stays simple: smoothing constants between 0.05 and 0.3, a 3-month moving average and 3-month ADIDA buckets. Tuned settings can change which model comes closest.
- For sparse demand, ask the ADI and CV² calculator which class a SKU falls in, or read the intermittent demand guide.