Demand Forecasting Methods: Quantitative, Qualitative and Machine Learning

The three families of methods

Demand forecasting methods fall into three families according to where they get their information:

• Qualitative: from people\'s judgment. Useful when there is no history (launches, new markets) or there is information the data does not contain.
• Quantitative time series: from the series\' own history. They assume the past pattern (level, trend, seasonality) continues. They are the basis of almost every operational forecast.
• Quantitative causal and machine learning: from history plus other variables (price, promotions, weather, indicators) and, in the case of machine learning, from patterns learned across many series at once.

A mature process uses all three: the time series as a base for most items, machine learning where it adds value, and business judgment as a recorded adjustment. What follows is a review of each method with what you need to know to use it: what it assumes, where it works and where it fails.

Qualitative methods

• Sales force estimates: each salesperson estimates their territory. Cheap and with local information; biased upward if there is a quota to meet or product to secure.
• Executive opinion and the Delphi method: a group of experts converges in anonymous rounds. For long horizons and strategic decisions, not for the monthly forecast per item.
• Historical analogy: a launch is forecast with the curve of a similar product. It is the right method for items with no history, and what systems do when they assign a "profile" to new items.
• Market research: surveys and purchase intent. Costly and useful for high-impact new products.

The practical rule: judgment is applied as an adjustment on top of a statistical forecast, at the level where the information exists, and recorded separately to measure afterwards whether it improved or worsened the result.

Moving average

The forecast is the average of the last n periods. It is the most used method in Excel and the easiest to explain. It assumes no trend or seasonality, or that n is short enough to follow them with a lag.

It works as a baseline and for stable series. It fails with trend (it always lags behind), with seasonality (it averages it out) and with intermittent demand (it swings with every sale). Its greatest value today is as the naive reference method: any serious method must beat it.

Simple exponential smoothing

The forecast is a weighted average of the latest actual demand and the latest forecast. The parameter α (between 0 and 1) sets how much weight recent data gets: high α reacts fast and is noisy; low α is stable and slow. It is chosen per item by minimizing historical error.

It is the workhorse for smooth series without marked trend or seasonality, and the basis of the whole family that follows. It fails with trend and seasonality (hence Holt and Holt-Winters) and with intermittent demand (it jumps after each sale and decays; hence Croston).

Holt and Holt-Winters

Holt (double smoothing) adds a second equation that estimates trend, with its own parameter β. It suits series that grow or decline steadily; it is worth damping the trend so as not to extrapolate indefinite growth twelve months out.

Holt-Winters (triple smoothing) adds a third equation for seasonality, with parameter γ, in an additive version (seasonality adds or subtracts a fixed amount) or multiplicative (seasonality is a percentage of the level, the usual case in sales). It needs at least two full seasonal cycles of history (24 months for annual seasonality). It is what Excel\'s FORECAST.ETS does.

The full family is known as ETS (error, trend, seasonality) and has around thirty variants depending on how the three components combine. A good engine tests several per item.

ARIMA

ARIMA (autoregressive integrated moving average) models the series as a combination of its own past values (AR part) and past errors (MA part), after differencing it to remove trend (I part). Its seasonal version, SARIMA, adds the same components at a seasonal lag.

It is more flexible than ETS and often ties or wins on long, regular series. It requires choosing orders (p, d, q) per item, which auto-ARIMA algorithms do, and it is heavier to compute. It fails, like the whole family, with intermittent demand and short series.

Croston and SBA

For intermittent demand (many zero periods), Croston splits the problem in two: the size of demand when it occurs and the interval between demands, smooths each separately and forecasts their ratio: the demand rate per period. SBA (Syntetos-Boylan) corrects Croston\'s upward bias by multiplying by (1 − α/2), and TSB updates the probability of demand every period to detect obsolescence.

They are the only methods that work reasonably on a third of distribution and spare-parts catalogs. We explain them with examples in intermittent demand.

Regression and causal variables

When demand depends on something you know in advance (price, promotional spend, temperature, number of points of sale), regression estimates that relationship and forecasts from the future values of those variables. It is the way to bring in promotions and prices explicitly, and the basis of elasticity models.

Its condition: you need to know the future variables. If forecasting demand first requires forecasting the weather, one problem has been swapped for another. It works well at family or brand level and with clean promotion data; it is fragile at item level with few events.

Machine learning

Machine learning models for demand (gradient boosting such as LightGBM or XGBoost, random forests, neural networks in very large catalogs) do two things differently from the previous methods: they learn from many series at once (a global model) and they accept any number of variables (calendar, price, promotion, item attributes, lags of the series itself).

They add value when there are thousands of items with shared patterns, informative external variables and enough data to validate. They won Walmart\'s M5 competition, with intermittent demand at store and item level. They fail, or do not add value, with few items, short series or without careful validation: a flexible model with little information learns noise. And they are less transparent: explaining why the model said 340 units requires additional tools.

The sensible thing is to include them in the tournament alongside classic methods and let the error decide per item. In typical distribution catalogs, machine learning wins on a share of the series, not all of them.

How to choose the method per item

1. Classify each series by pattern (smooth, erratic, intermittent, lumpy) and maturity. The class narrows down which methods are worth testing.
2. Hide the last months (six, for example) and fit each candidate on the rest.
3. Forecast the hidden months with each candidate and measure the error (WAPE or MASE, and bias) against what really happened.
4. Pick the winner per series, always including the naive method as a reference: if nobody beats it, the series is not forecastable with that data and it is better to aggregate it.
5. Repeat every cycle or at least every quarter: the winning method changes when demand changes.

This process is the model tournament. Doing it by hand in Excel for a whole catalog is unfeasible; it is what ForecastSolve automates with 23 candidates per item, storing the metrics of each one.

Summary

Method	Assumes	Works on	Fails on	Minimum history
Moving average	Stable level	Stable series; naive reference	Trend, seasonality, intermittence	3 to 6 months
Simple smoothing	Stable level	Smooth series without pattern	Trend, seasonality, intermittence	6 to 12 months
Holt	Trend	Sustained growth or decline	Seasonality, sudden changes	12 months
Holt-Winters / ETS	Trend and seasonality	Regular seasonal series	Intermittence, short series	24 months
ARIMA / SARIMA	Autocorrelation	Long, regular series	Intermittence, short series	24 to 36 months
Croston / SBA / TSB	Intermittent demand	Spare parts, slow movers	Continuous demand	12 to 24 months
Regression	Relationship with known variables	Promotions, price, at family level	Unknown future variables	Depends on the events
Machine learning	Shared patterns and variables	Large catalogs with external data	Little data, no validation	Many series or 36+ months