Supply Chain · Forecast

Forecast Method Picker

Choose the simplest method that fits: moving average for stability, exponential smoothing for trends, seasonal models for cycles.

Time45 min
FormatSolo
StageForecast

Forecast Method Picker: what it is and why it works

The forecast method picker matches each item or family to the simplest statistical method that fits its demand pattern. You plot at least a year of history and decide whether the series is flat, trending, seasonal or both. Flat, noisy demand suits a moving average or simple exponential smoothing, which weights recent periods more heavily through a smoothing constant. A persistent trend calls for Holt's linear method, which adds a smoothed trend term. Trend plus a repeating cycle calls for Holt-Winters, which adds seasonal indices. Very intermittent demand, with many zero periods, needs its own treatment, such as Croston-type methods or simple reorder rules.

The logic is parsimony: every extra parameter must earn its place by reducing error on data the model has not seen. Complex models fitted to short or noisy history often track past noise and forecast worse than a simple baseline. Comparing each candidate against a naive forecast, such as last period's actual or the same month last year, keeps the choice honest. The picker works hand in hand with a forecast accuracy audit, which measures whether the chosen method performs, and with the monthly demand review, which catches events no method can predict. Machine learning approaches become worth testing only once simple methods are well tuned and there are enough drivers and data to exploit.

What you need

At least twelve months of clean demand history per item or family; two full years for seasonal items
Demand cleaned of known one-off events, stockout periods and data errors
A forecasting tool or spreadsheet able to run moving average, exponential smoothing and Holt-Winters
An agreed error metric, such as weighted MAPE, to compare methods on held-out months

What you get

A pattern label for each item or family: flat, trend, seasonal, trend-seasonal or intermittent
An assigned method and its parameters for each group
Out-of-sample error for the chosen method compared with a naive benchmark
A quarterly review date to check whether patterns have shifted

When to use it

When one forecasting model is used for everything from stable to chaotic demand.

How to do it, step by step

Plot twelve months of demand for the items in question.
Check stability: flat, trending, or seasonal?
Match the method: moving average for flat, Holt-Winters for trend and season.
Never use a complex model where a simple one performs as well.
Re-evaluate the choice quarterly as demand patterns drift.

Worked example: Choosing methods for a pool chemical line

Illustrative scenario — figures are realistic but not from a real company.

A regional producer of water treatment chemicals sold about 60 packaged products to pool service companies and municipalities. The planning system applied a six-month moving average to everything, which left the company short every spring and overstocked every fall.

The planner plotted two years of monthly shipments for each product family and sorted them visually into flat, trending and seasonal groups.
Municipal bulk products were flat with random noise; pool chlorine tablets showed a strong summer peak about three times the winter level; a newer enzyme product was growing steadily.
She fitted simple exponential smoothing to the flat group, Holt's method to the growing product and Holt-Winters to the seasonal families, holding back the last six months to test.
For each family she compared the held-out error to a naive forecast using the same month last year. For two small seasonal items Holt-Winters did no better than the naive seasonal forecast, so she kept the simpler one.

Result. On the seasonal families, weighted error over the test months fell from about 45% with the moving average to around 22%. The spring shortage did not recur the next season. The planner noted that matching method to pattern delivered most of the gain, and that tuning parameters further added little.

Open it in the FieldKit Supply Chain app → Get the Supply Chain template pack — join the waitlist

Common pitfalls and how to avoid them

Selecting the method with the best fit on historical data.Judge methods on held-out periods they were not fitted to; in-sample fit rewards overfitting.
Using Holt-Winters with less than two full seasonal cycles.Collect at least two years for seasonal items, or borrow seasonal indices from a similar family.
Feeding raw history that includes stockout months and one-off orders.Clean or flag abnormal periods before fitting so the method learns true demand, not supply problems.
Forcing a smoothing method on intermittent items with many zero periods.Treat intermittent demand separately with methods designed for it or with simple min-max rules.

Frequently asked questions

What is the difference between a moving average and exponential smoothing?

A simple moving average gives equal weight to the last n periods and ignores older ones. Exponential smoothing gives every past period some weight, with weights decreasing geometrically as data gets older, controlled by a smoothing constant between 0 and 1. Exponential smoothing reacts more smoothly to change and needs to store only the previous forecast, which made it popular for large item counts.

When should you use Holt-Winters forecasting?

Use it when demand has a clear repeating seasonal pattern, with or without a trend, and you have at least two full seasonal cycles of history. It maintains a level, a trend and a set of seasonal indices. Choose the multiplicative form when seasonal swings grow with volume and the additive form when they stay roughly constant in size.

How do you choose a forecasting method?

Start by plotting the data and identifying the pattern. Shortlist methods that fit it, test each on recent periods held out of the fitting, and compare them against a naive baseline. Keep the simplest method that performs about as well as the best. Revisit the choice regularly, since products move through launch, growth, maturity and decline.

Origin

Exponential smoothing — Robert G. Brown, 1956; Holt-Winters — Charles Holt & Peter Winters, 1960.

Related methods

More in “Forecast”

Anticipate demand so the chain plans from facts, not hope.