Holt-Winters Seasonal Forecast: Worked Example

What is the Holt-Winters method

The Holt-Winters seasonal forecast smooths three parts of demand at once: level, trend and seasonality. It is the standard choice when the series repeats a pattern every year (or week, or month) and moving averages and simple exponential smoothing keep missing in the same direction.

Each new observation corrects the estimate of each component, weighted by α (level), β (trend) and γ (seasonality), all between 0 and 1. It extends Holt's linear exponential smoothing with one seasonal term per position in the cycle. The method is credited to Charles Holt (1957) and Peter Winters (1960).

Holt-Winters visual map

The map summarizes the four equations, the quarterly worked example below and the rule for choosing between the additive and multiplicative models.

The equations (additive model)

With a cycle of m periods (m = 4 for quarters, 12 for months) and actual demand y_t:

Forecast: F_t+h = L_t + h × T_t + S_t+h−m

• Level: L_t = α × (y_t − S_t−m) + (1 − α) × (L_t−1 + T_t−1)
• Trend: T_t = β × (L_t − L_t−1) + (1 − β) × T_t−1
• Seasonal term: S_t = γ × (y_t − L_t) + (1 − γ) × S_t−m

In the multiplicative version the seasonal term is a ratio: the forecast is (L + h × T) × S, and the updates use division (y ÷ S and y ÷ L) instead of subtraction. Use it when the size of the seasonal peak grows with the demand level.

Worked example: three years of quarterly demand

Fictional example, in units of a seasonal item with a peak in Q3:

Step 1: initialization. The year 1 average is (120 + 150 + 180 + 130) ÷ 4 = 145. The year 2 average is 159, so the initial trend is (159 − 145) ÷ 4 = 3.5 per quarter. The level at the end of year 1 is 145 + 1.5 × 3.5 = 150.25 (the average is centered between Q2 and Q3, 1.5 periods before Q4). Initial seasonal terms are y − 145: Q1 = −25, Q2 = +5, Q3 = +35, Q4 = −15.

Step 2: parameters. α = 0.3, β = 0.1, γ = 0.2.

Step 3: first forecast (Y2 Q1). F = 150.25 + 3.5 + (−25) = 128.75. Actual = 132, error = +3.25. Update: L = 0.3 × (132 + 25) + 0.7 × (150.25 + 3.5) = 154.725; T = 0.1 × (154.725 − 150.25) + 0.9 × 3.5 = 3.5975; S = 0.2 × (132 − 154.725) + 0.8 × (−25) = −24.545.

Repeating through Y3 Q4:

Values are rounded for display only; calculations keep full precision. The mean absolute deviation (MAD) of these 8 one-step forecasts is 5.55 units and the MAPE is 3.6%. Both are explained in forecast error: MAPE, MAD and bias.

Forecast for year 4

At the end of year 3: L = 176.78 and T = 3.30. Applying F_t+h = L + h × T + S:

Forecast demand for year 4 totals 739 units, against 692 in year 3. The Q3 peak (about 222) and the Q1 trough (about 156) stay in the right place, and the level rises roughly 3.3 units per quarter.

How to choose α, β and γ

• High values (near 1) react fast but chase noise. Low values (0.05 to 0.3) smooth more and are more stable.
• Test a grid of combinations and keep the one that minimizes error (MAD, MAPE or RMSE) on recent periods, without using hold-out data to choose. Spreadsheet Solver and statistics libraries do this automatically.
• β and γ are usually smaller than α: trend and seasonality change slowly.
• Require at least two full cycles of history (24 months for a yearly cycle in monthly data). With less, the seasonal initialization is fragile.

When to use it and how it feeds planning

Use Holt-Winters for items with regular demand, stable seasonality and long history: beverages, fans, school supplies, harvest packaging. For intermittent items or items with frequent stockouts, the series cannot support the model.

The forecast feeds the master production schedule, aggregate planning and the S&OP meeting. The forecast error, in turn, sets the safety stock: the larger the deviation, the bigger the buffer.

Common mistakes and limitations

• Using too little history: fewer than two cycles gives poor seasonal terms.
• Confusing sales with demand: lost sales from stockouts, one-off promotions and shortages distort the series. Clean explained spikes before fitting.
• Choosing additive when multiplicative is right (or the reverse): if the seasonal swing grows with the level, use multiplicative.
• Ignoring damped trend: projecting a linear trend over many periods overstates growth; for long horizons, damp or cap the trend.
• Measuring error only on fitted data: also test on a period not used in fitting.
• Forgetting known events (launches, price changes, moving holidays): the model only sees the past.

Frequently asked questions

What is Holt-Winters?

An exponential smoothing method with three components, level, trend and seasonality, updated with each new actual value.

What is the difference between Holt and Holt-Winters?

Holt handles only level and trend. Holt-Winters adds a seasonal term, so it fits series with a repeating cycle.

When should I use the multiplicative model?

When the seasonal swing grows in proportion to the demand level. If the peak is always the same size in units, use the additive model.

How much data do I need?

At least two full cycles, for example 24 months for yearly seasonality in monthly data. Three or more cycles are better.

How do I pick α, β and γ?

Test combinations and keep the one with the lowest error (MAD, MAPE or RMSE) on recent periods not used in fitting. Values between 0.05 and 0.3 are common starting points.

What is the Y4 Q1 forecast in the example?

L + T + S = 176.78 + 3.30 + (−23.96) ≈ 156.1 units.

About the author

Vagner Soares

Lean Manufacturing & Behavioral Management Specialist

Over 20 years in the automotive and metalworking industries (GM and Dana), Lean Manufacturing practitioner since 2006. SENAI instructor and mentor in Brazil’s Brasil Mais Produtivo program, delivering consulting, training and audits for 50+ companies, combining quality, productivity and people development.

LinkedInInstagramFull profile →