Winters' Method

On this page
  1. Purpose
  2. Key Capabilities
  3. Statistical Methods
  4. Input Fields & Options
  5. Output Descriptions
  6. Worked Example

Purpose

Winters' Method (Holt-Winters, or triple exponential smoothing) smooths a time series with three running components: a level, a trend and a seasonal index for each period of the season. Its forecasts grow with the trend and repeat the seasonal pattern. Use it for monthly, quarterly or shift-pattern data with both a trend and a season, for example to forecast demand, energy use or workload for the coming year, when the pattern changes slowly over time.

Key Capabilities

  • Multiplicative method (the seasonal swing grows with the level) or additive method (the swing stays the same size).
  • Separate weights for level (alpha), trend (gamma) and season (delta), 0.2 each by default.
  • Starting values from the data or values you enter, including one seasonal index per period.
  • Forecasts with prediction limits that widen with each period ahead, from the end of the series or from an earlier row to test the method against known data.
  • MAPE, MAD and MSD accuracy measures, a plot of the data, fits and forecasts, an optional results table and four-in-one residual plots.

Statistical Methods

For the multiplicative method, with seasonal length p, the components are updated each period: level Lₜ = α·(yₜ / Sₜ₋ₚ) + (1 − α)·(Lₜ₋₁ + Tₜ₋₁); trend Tₜ = γ·(Lₜ − Lₜ₋₁) + (1 − γ)·Tₜ₋₁; seasonal index Sₜ = δ·(yₜ / Lₜ) + (1 − δ)·Sₜ₋ₚ. The fit for period t is (Lₜ₋₁ + Tₜ₋₁)·Sₜ₋ₚ, and the forecast m periods after the origin is (L + m·T) × the latest index for that period of the season. The additive method subtracts and adds the seasonal term instead of dividing and multiplying.

Starting values from the data: the initial level and trend are the intercept and slope of a least-squares line through the first season. For the multiplicative method each initial index is the average of that period over all the data divided by the overall average; for the additive method it is the average, for that period, of the residuals from a straight line fitted to the whole series. The weights are used as entered (blank counts as 0.2); they are not optimised. With Starting values = User-specified you enter the initial level, trend and p seasonal indices.

The prediction limits are forecast ± z × dₘ × MAD, where z is the normal quantile for the confidence level, MAD is the mean absolute error of the fits and dₘ grows with the number of periods ahead m (computed from the largest of the three weights). Accuracy: MAPE = mean |error / data| × 100, MAD = mean |error|, MSD = mean error². The series needs at least two full seasons and no gaps; the multiplicative method needs all values above zero.

Input Fields & Options

Field

Description

Variable (Y)

Required. The numeric series, in time order.

Seasonal length

Required, at least 2. Periods per season, e.g. 12 for monthly data with a yearly pattern.

Method type

Multiplicative (default) or Additive.

First observation is in seasonal period

Default 1. The position of the first row in the season, e.g. 6 when the season starts in January and the data start in June.

Level weight (alpha) / Trend weight (gamma) / Seasonal weight (delta)

0.0001 to 1, default 0.2 each. Larger values let that component react faster to recent data.

Starting values

From the data (default) or User-specified.

Initial level / Initial trend / Initial seasonal indices

Used when Starting values is User-specified; all three are then required. Seasonal indices: comma or space-separated, one per period of the season (Seasonal length values).

Number of forecasts

Default 0 (none), up to 1000.

Confidence level (%) for forecast interval

Default 95. Active when forecasts are requested.

Forecast origin (row)

Optional, active when forecasts are requested. Forecasts start after this row, using the data up to it; blank = end of the series.

Results to display

Summary table (default) or Summary table and results table.

Residual plots (four in one)

Unticked by default.

Title

Title of the output. Default Winters' Method.

winters-method-inputs


Figure 1 — Multiplicative Winters' Method on monthly Sales, seasonal length 12, weights 0.2 each and 12 forecasts.

Output Descriptions

  • Winters' Method (graph) — actual data, fits, forecasts and the prediction band, with MAPE, MAD and MSD in the title and the three weights above the plot.
  • Method — model type, data column and length (plus the first seasonal period and forecast origin when set).
  • Smoothing Constants — α (level), γ (trend) and δ (seasonal).
  • Accuracy Measures — MAPE, MAD and MSD.
  • Forecasts — period, forecast, lower and upper limits.
  • Results — when chosen: Period, data, Season, Level, Trend, Seasonal, Predict and Error for each period.
  • Residual Plots — when ticked.
  • PDF export via Download.

winters-method-output


Figure 2 — The 2025 forecasts repeat the yearly shape on a higher level, from a low of 1,205 in February to 1,891 in December.

Worked Example

A consumer electronics distributor needs a 12-month sales forecast for 2025 that grows with the trend and repeats the yearly pattern of weak early months and a December peak. The analyst has 60 months of unit sales (January 2020 to December 2024). The analyst selects Variable (Y) = Sales, Seasonal length = 12 and Method type = Multiplicative, keeps the three weights at 0.2, and sets Number of forecasts = 12.

The forecasts run from 1,317 in January 2025, down to the low point of 1,205 in February, up to 1,891 in December, with 95% limits of 1,756 to 2,026 for December. The intervals widen from about ±107 units in January to ±135 in December. MAPE is 4.2% (MAD 43.9 units), far better than the smoothing methods without a season (8.3% to 9.4%). December 2025 at 1,891 compares with 1,759 in December 2024: about 7.5% growth.

To test the method, Forecast origin (row) = 48 forecasts 2024 from the data up to December 2023. Eleven of the twelve actual months fall inside the 95% limits (February 2024, at 1,176, is one unit above its upper limit), and the forecasts miss by 4.9% on average; but every month is under-forecast, because growth picked up in 2024. With Method type = Additive the MAPE drops to 3.7%, so a fixed number of units per month fits this season slightly better than a percentage. Weights of 0.4, 0.1 and 0.3 give 3.8%, no better than the defaults.