Is Holt's method the same as Holt-Winters?

Holt's linear method includes level and trend; Holt-Winters adds seasonality.

Does additive seasonality mean the series cannot grow?

No. Growth can be represented by trend while seasonal effects remain additive.

Do the smoothing recursions alone define prediction intervals?

No. Uncertainty requires an associated statistical model or another justified procedure.

AI Glossary / Time series and signal processing / Intermediate

Holt-Winters forecasting

Definition

Holt-Winters forecasting is a family of exponential-smoothing methods that represents a time series through level, trend, and repeating seasonal components. The components are updated as observations arrive and combined to produce future forecasts. Additive and multiplicative seasonal forms handle different relationships between seasonal variation and the series' overall level.

What it means

A business may have both a gradual increase in demand and a recurring busy season. Holt-Winters methods separate these patterns operationally: level describes the current baseline, trend describes its movement, and seasonality describes position within a repeating cycle.
The method does not discover every kind of pattern or explain why the pattern exists. It assumes that its chosen component structure is a useful approximation for forecasting, and that the seasonal period has been specified appropriately.

How it works

Choose the seasonal period, initialize level, trend, and seasonal states, and select smoothing parameters using training data. Each new observation updates the level after accounting for seasonality, updates the trend from the level movement, and revises the seasonal component for the corresponding season.
Additive seasonality represents seasonal effects in the original measurement units; multiplicative seasonality represents relative factors and requires care with zero or negative values. Forecasts combine the latest level, an extrapolated or damped trend, and the relevant seasonal state. Different software conventions can parameterize seasonal updates differently, so equations and parameter meanings must be matched before comparing numerical values.

Three practical examples

A distributor forecasts quarterly demand with recurring high and low quarters, comparing additive seasonality with proportional seasonal variation.
A service organization forecasts monthly request volumes while allowing its baseline and trend to adapt gradually.
A facilities team models a single regular seasonal pattern in consumption, but checks whether additional daily or weekly patterns require a more suitable model.

Think of it this way

Think of estimating a runner's position on a gently rising circular track. There is an overall height, a gradual change in that height, and a repeated rise or dip at each part of the circuit. Holt-Winters forecasting keeps track of those components, although actual demand patterns need not remain as regular as the analogy suggests.

Worked example

Use an original additive quarterly example with period four. Immediately before observing the current quarter, suppose level is 100, trend is 2 units per quarter, and the old seasonal component for this quarter is −10. Other seasonal components are stipulated as +5, +15, and −10 in subsequent-quarter order. The new observation is 96. Set alpha = 0.3, beta-star = 0.2, and gamma = 0.4.
Using the FPP3 seasonal-update convention, updated level is 0.3 × [96 − (−10)] + 0.7 × (100 + 2) = 103.2. Updated trend is 0.2 × (103.2 − 100) + 0.8 × 2 = 2.24. Updated season is 0.4 × (96 − 100 − 2) + 0.6 × (−10) = −8.4.
The next-quarter forecast uses that next quarter's existing seasonal component +5: 103.2 + 2.24 + 5 = 110.44. The forecast four quarters ahead returns to the current quarter's updated seasonal component: 103.2 + 4 × 2.24 − 8.4 = 103.76.
These states and parameters are stipulated to demonstrate one update, not estimated from a dataset. A formulation using the newly updated level in the seasonal equation requires a correspondingly different seasonal smoothing parameter; substituting the same gamma would not reproduce this calculation.

Illustrative case study

Illustrative fictional case: Fir Homeware uses an additive Holt-Winters model for quarterly sales. As the business expands, seasonal swings appear to grow in proportion to the baseline rather than by a fixed number of units.
The analyst compares a multiplicative alternative and a seasonal naive benchmark on later periods, checks the treatment of zero-sales products, and evaluates the ordering horizon. Management retains different methods for different product groups where evidence supports them. This fictional case illustrates model comparison, not a claim that proportional seasonality is universally superior.

Common misconceptions and mistakes

Holt-Winters is not simply a moving average repeated three times. Its components have distinct update roles. A seasonal period of twelve for monthly annual seasonality also differs from a twelve-step forecast horizon.
Do not mix additive and multiplicative equations, or assume that similarly named parameters have identical meanings in every implementation. Undamped trends can extrapolate unrealistically over long horizons. Outliers and structural changes can distort updates, and one seasonal component may be inadequate for multiple overlapping seasonal cycles. Accurate-looking fitted values do not replace out-of-sample evaluation.

Business use

The method offers a practical baseline for recurring business demand because its level, trend, and seasonal contributions can be explained to non-specialists. Efficient updating can support many related series, provided each series has enough relevant history and an appropriate seasonal specification.
Document initialization, parameter constraints, seasonal form, and trend damping. Compare forecasts against simpler methods at operational lead times, and evaluate interval coverage when a statistical ETS formulation supplies uncertainty. Review products or locations with breaks, intermittent activity, or insufficient seasonal history separately.

Check your understanding

Why is the next-quarter seasonal value +5 rather than −8.4 in the example?

Answer: The next forecast refers to a different position in the seasonal cycle.

What is the updated trend?

Answer: 2.24 units per quarter.

Why does the four-quarter forecast use −8.4?

Answer: Four steps ahead returns to the same quarter of the seasonal cycle.

Would the example demonstrate parameter estimation?

Answer: No. All starting states and parameters were supplied deliberately.

Related terms

Exponential smoothing
Levels
level
trend
seasonality
additive seasonality
multiplicative seasonality
damped trend
ETS model
seasonal naive forecast

Learn next

reproduce the additive update and compare it with a multiplicative example using positive data. Then examine how trend damping changes longer forecasts and how rolling-origin evaluation informs selection among seasonal methods.

Sources

Sources: Hyndman and Athanasopoulos, Forecasting: Principles and Practice, third edition, Holt-Winters seasonal methods and parameter conventions: https://otexts.com/fpp3/exponential-smoothing.html . Glossary of Forecasting Terms: https://robjhyndman.com/mwh3/FG4.pdf . Primary textbook equations consulted. Original fractions independently verify the update and forecasts; the case is fictional.

Previous: Holm - Bonferroni method Next: Homogeneity of regression