Supply Chain & InventoryFree referencePDF, Excel, Word export

Statistical Forecast Methods Reference: Which Model for Which Demand

Choosing a forecast method starts with the shape of the demand, not the software menu. This reference matches common statistical methods to the demand pattern they handle, what data they need and where they usually fail.

Download:

Method	Demand pattern it suits	History needed	Key parameters	Pitfall
Naive (last value)	Any, used as a baseline	1 period	None	Copies noise and last month's spike
Seasonal naive	Strong repeating seasonality	1 full season	Season length	Repeats one-off events from last year
Simple moving average	Stable level, no trend or season	n periods	Window length n	Lags behind trends; every period weighs the same
Weighted moving average	Stable level where recent periods matter more	n periods	Window and weights	Weights are subjective unless tested
Simple exponential smoothing	Stable level with noise	Around 6 or more periods	Alpha between 0 and 1	Flat forecast, no trend or season
Holt (double exponential)	Linear trend, no season	Around 10 periods	Alpha, beta	Trend can run away on long horizons, damping helps
Damped trend	Trend expected to level off	Around 10 periods	Alpha, beta, phi	Phi near 1 behaves like undamped Holt
Holt-Winters additive	Trend plus seasonal swing of constant size	2 full seasons or more	Alpha, beta, gamma, season length	Poor when seasonal amplitude grows with level
Holt-Winters multiplicative	Trend plus seasonal swing proportional to level	2 full seasons or more	Alpha, beta, gamma, season length	Needs strictly positive history
Classical decomposition	Visible trend and seasonal index	2 full seasons or more	Season length, trend method	Seasonal index fixed, does not adapt
Linear regression on time	Steady linear growth or decline	Around 12 periods	Slope and intercept	Extrapolates the line forever
Regression with drivers	Demand driven by price, weather, orders or capacity	Driver history aligned to demand	Selected drivers	Driver must be known in advance for the horizon
ARIMA / SARIMA	Autocorrelated series, optional seasonality	Around 3 seasons or 50 points	p, d, q and seasonal terms	Hard to explain to sales, needs stationarity checks
Croston	Intermittent demand with many zero periods	Several non-zero demands	Smoothing constant	Biased upward; Syntetos-Boylan fixes this
Syntetos-Boylan approximation (SBA)	Intermittent demand	Several non-zero demands	Smoothing constant	Still weak on lumpy demand with extreme sizes
TSB (Teunter-Syntetos-Babai)	Intermittent demand that may become obsolete	Several non-zero demands	Two smoothing constants	Probability decays between demands, confusing to read
Life cycle or analog profile	New product with no history	History of similar items	Chosen analog and scaling	Analog choice is judgment, document it
Judgmental / consensus	Events, launches, contracts not in history	Sales and marketing knowledge	Process rules	Optimism bias unless tracked with forecast value add

History lengths are rules of thumb, not strict limits. Always compare any method with a naive forecast on a holdout period before adopting it.

Model selection on your data

Want a method suggested per SKU from your ERP history? Ask for an optional expert review.

Optional. The tool and every export work without this. No spam.

How to use this reference

Plot the item first. Is there a level only, a trend, a repeating season, or lots of zeros? Search the table for that pattern and shortlist two or three methods. Run each on a holdout of the last few periods, compare error against a naive baseline, and keep the simplest method that clearly beats it. Export the table to Excel and add a column for the method you chose per family, so the decision is written down.

Matching pattern to method

Stable demand with noise suits simple smoothing or a moving average. A steady trend needs Holt or a damped trend. A repeating seasonal shape needs Holt-Winters or decomposition, and at least two full seasons of history. Intermittent demand, where most periods are zero, needs Croston or its corrections, because ordinary smoothing keeps forecasting a small number every period that never actually occurs.

Level only: SES or moving average
Trend: Holt or damped
Season: Holt-Winters or decomposition
Intermittent: Croston, SBA or TSB
No history: analog or life cycle profile

Worked example

A spare part sells 0, 0, 4, 0, 0, 0, 6, 0, 3, 0, 0, 5 units over twelve months. Simple smoothing forecasts about 1.5 units every month, which is never right. Croston forecasts the average size of a demand, about 4.5 units, divided by the average gap between demands, about 3 months, giving roughly 1.5 per month of rate but with an explicit interval. Use the rate for stocking, not for monthly production orders.

Common mistakes

Using Holt-Winters with less than two seasons of data lets the model invent a season. Applying smoothing to intermittent parts produces a forecast that looks smooth but fits nothing. Forgetting to clean one-off events such as a lost customer or a strike lets the model learn them as the pattern. Always review outliers before fitting.

Frequently Asked Questions

What is the simplest forecast method that works?

A seasonal naive or simple moving average is often a surprisingly strong baseline. Any more complex model should beat it on a holdout period before you trust it. If it cannot, keep the simple method.

How much history do I need for Holt-Winters?

At least two full seasons so the model can estimate the seasonal shape, for example 24 months for monthly seasonality. With less, the seasonal component is poorly estimated and the forecast can be unstable.

When should I use Croston instead of exponential smoothing?

Use Croston or its variants when a large share of periods show zero demand. Ordinary smoothing treats zeros as small values and gives a flat number that is wrong every period. Croston models demand size and interval separately.

Should I let the software pick the best model?

Automatic selection is useful at scale, but check it against a naive baseline and review the largest items by hand. Selection on a very short holdout can pick a model that fit by luck.

If you want a model selected per SKU and checked against a naive baseline automatically, Netray can run that on your ERP history on your own servers.

Talk to an Expert Explore More Free Tools

Related Tools

Supply Chain & Inventory

More Supply Chain & Inventory tools

See all

Back to All Tools