Statistics Glossary → Exponential Smoothing (Holt-Winters)
Exponential Smoothing (Holt-Winters)analysis
A family of methods (SES, Holt, Holt-Winters) that produce short-term forecasts by continually updating a series' level, trend and seasonal pattern with weighted averages that give more weight to the recent past.
Concept
When forecasting next month's value, the latest observations usually say more than the distant past. Exponential smoothing formalises this intuition: the forecast is a weighted average of past observations, with weights that decline exponentially the further back one goes. Simple exponential smoothing (SES) tracks only a level; Holt's method adds a slope (trend), and Holt-Winters adds a seasonal component as well.
Each component is updated whenever a new observation arrives. For the level, for example: ℓ_t = α·y_t + (1 − α)·ℓ_{t−1}. The smoothing parameters lie between 0 and 1: α sets how quickly the level reacts to new information, β* the slope and γ the season; values near 1 make a component follow the latest observations closely, values near 0 keep it almost fixed. With a damped trend, the parameter φ makes the trend gradually flatten over a long horizon. The parameters are estimated from the data so that the one-step-ahead forecast errors are as small as possible.
The season may be additive (the same amount each year) or multiplicative (proportional to the level); the choice depends on whether the seasonal wave grows with the level of the series. The method is strong for short-term forecasting but includes no explanatory variables and cannot foresee structural changes. Models can be compared with AICc, and prediction intervals widen as the horizon lengthens.
When to use it
- To forecast a few periods ahead for a single series with trend or seasonality using a quick, intuitive model (demand, stock, occupancy).
- If the series has neither trend nor season, use SES; if it has only a trend, Holt; if it has both trend and season, Holt-Winters.
- If the autocorrelation structure is to be modelled in more detail or an external variable added, use ARIMA/ARIMAX; if the components are to be interpreted separately within a probability model, use an unobserved components model.
An everyday example
A hotel in a seaside town wants to plan next summer's staffing from the past six years of monthly guest nights. Guest nights peak every summer, drop in winter and rise slowly over the years. A Holt-Winters model tracks the level, the upward trend and each month's seasonal effect separately; a high γ shows that the seasonal pattern quickly reflects recent changes (for example, September becoming busy too). The model might forecast 3,100 guest nights for next July, and the hotel decides how many extra staff to hire based on that forecast's interval.
Parameters in MerQur
Method (Automatic (best fit) / Holt-Winters (trend+season) / Holt (with trend) / SES (simple) / Simple seasonal (no trend))
Concept: The method sets which components the model tracks: SES only the level, Holt the level and slope, Holt-Winters the level, slope and season. The automatic option fits candidate models and picks the one with the lowest AICc, which rewards fit while penalising unnecessary components.
Everyday example: Because the hotel's guest-night series has both yearly growth and a summer peak, SES or Holt miss the season and produce forecasts that ignore the month. Holt-Winters, by contrast, forecasts July far higher than January.
What it does: SES level; Holt level + trend; Holt-Winters level + trend + season; simple seasonal level + season. Automatic: the smallest AICc among additive-error candidates (SES, Holt, damped Holt; additive seasonal if n ≥ 2m and multiplicative if the series is positive).
Effect on the result: Model structure and all results.
On the validation data: oil SES α = 0.8338, ℓ0 = 446.58, forecast 542.68. ausair Holt α = 0.8211, β* = 0.0001.
Seasonal type (Additive / Multiplicative / None) and Seasonal period (default 12)
Concept: With an additive season each season's effect is a fixed amount (e.g. +500); with a multiplicative season it is a proportion of the level (e.g. 30% more). The period is the number of time steps after which the seasonal cycle repeats: 12 for monthly data, 4 for quarterly data.
Everyday example: If the gap between July and January widens as the hotel grows, the season is multiplicative: July sits about 60% above the average every year. For monthly data the period must be 12; with quarterly data, entering 12 makes the model look for a three-year cycle that does not exist.
What it does: The form and length m of the season. With "None", Holt-Winters equals Holt. The period is not detected from the date frequency; 4 must be entered by hand for quarterly data.
Effect on the result: The seasonal component, SSE and forecasts.
On the validation data: austourists (m = 4): additive SSE = 135.92, forecast[1] = 76.01; multiplicative SSE = 105.17, forecast[1] = 79.43. The default period 12 builds a wrong season on quarterly data (SSE 108.59).
Damped trend (for long horizons) and Trend type (Additive / Multiplicative)
Concept: A damped trend multiplies the slope by φ (< 1) at each step so that it gradually shrinks; forecasts at long horizons therefore do not keep rising or falling at the same pace forever. A multiplicative trend assumes the level grows by a constant proportion rather than a constant amount.
Everyday example: The hotel's guest nights have grown fast in recent years, but with a limited number of rooms this growth cannot last forever. A damped trend prevents unrealistic forecasts five years ahead that exceed the hotel's capacity.
What it does: The damped trend makes the trend die out over the horizon via φ; the multiplicative trend assumes proportional growth.
Effect on the result: Parameters (φ) and forecasts.
On the validation data: livestock damped α = 0.9999, φ = 0.98; multiplicative trend ausair SSE 139.15.
Forecast horizon and Confidence Level
Concept: The forecast horizon sets how many periods ahead to forecast, and the confidence level sets how much confidence the prediction interval carries. A prediction interval is the plausible range of a single future observation, and it widens as the horizon lengthens.
Everyday example: A 12-step forecast gives the hotel a guest-night figure for every month of the next year. The 95% interval is narrow for next month but much wider for July a year from now; staffing plans should allow for this uncertainty.
What it does: Forecast horizon and interval level. In class 1 models (additive error, additive trend/season) the intervals are analytic; in multiplicative models they come from simulation.
Effect on the result: Number of forecasts and interval width.
On the validation data: Holt α = 0.8, β* = 0.1, h = 10: 95% interval [76.92, 113.29]. austourists: 90% [72.58, 79.21], 95% [72.06, 80.08], 99% [70.72, 80.99].
Fixed parameters (α, β*, γ, φ; lowest value = free)
Concept: Fixes one of the smoothing parameters by hand instead of estimating it from the data. α sets how quickly the level reacts to new observations, β* the slope, γ the season, and φ how strongly the trend is damped; a fixed value usually fits slightly worse than the estimated one.
Everyday example: If the hotel manager knows the seasonal pattern hardly changes from year to year, γ can be fixed at a small value such as 0.1 so that a single unusual summer does not distort the pattern. How much the fixed value worsens the fit can be seen from the error measures.
What it does: Fixes the selected parameter; the others and the initial states are estimated.
Effect on the result: SSE, forecasts and intervals.
On the validation data: oil with α = 0.5 fixed: SSE = 15387.9 (free 14235.6); ausair α = 0.8 + β* = 0.1: SSE = 138.078; austourists all fixed (α 0.3, β* 0.05, γ 0.3, φ 0.95): SSE = 171.783.
Holdout set (last k observations); Additional output: Initial states (ℓ0, b0, s), Save to dataset
Concept: A validation set holds back the last k observations when fitting the model so that the forecasts can be compared with these actual values; this is a more honest measure of performance than in-sample fit. The initial states are the level, slope and seasonal values before smoothing begins.
Everyday example: If the last 12 months are held out and the model is fitted to the previous five years, one sees by what percentage the forecasts miss that year's actual guest nights on average. The hotel can use this error to decide how much of a margin to leave when planning.
What it does: The last k observations are not used for fitting and test accuracy is reported. The initial states ℓ0, b0, s1..sm are given in chronological order. Save adds the forecast and residual columns.
Effect on the result: Additional output.
On the validation data: austourists, last 8 observations: test RMSE 2.499, MAPE 3.385.
Simulation seed (prediction interval)
Concept: For models with a multiplicative component, the prediction interval is computed not from a closed formula but by simulating the model thousands of times. The seed is the starting point of this random simulation; the same seed reproduces the same interval.
Everyday example: If two analysts forecasting July with a multiplicative-season model use the same seed, their intervals come out exactly the same. With a different seed the point forecast does not change, and the interval limits move only slightly.
What it does: For models with a multiplicative trend or season, the prediction interval is computed with 1000 simulations; the seed makes the simulation reproducible. The default is 20240607.
Effect on the result: Only the simulation interval of multiplicative models changes; the point forecast and analytic intervals do not. With a non-default seed, "(Simulation seed = N.)" is added to the interval note.
On the validation data: Entering 20240607 explicitly gives a result equal to the default; with seed 7 the interval differs and the point forecast is the same.
Example on real data
Data: austourists 2005-2015, R fpp2
Variables: date date (quarter), value nights (international visitor nights, millions); Holt-Winters, period 4
- Question: How are the quarterly visitor nights of international tourists to Australia (2005-2015, n = 44) smoothed and forecast with Holt-Winters?
- Model: Additive trend and additive season (m = 4), ETS(A,A,A); α = 0.263, β = 0.0004, γ = 0.455; the level is updated slowly, the seasonal pattern quickly, and the trend is almost constant.
- Fit: SSE = 135.96, AIC = 65.64, AICc = 72.31, BIC = 79.91; RMSE = 1.758, MAPE = 2.94%, MASE = 0.449, R² = .976.
- Forecast: 76.01 for 2016 Q1 (95% interval 72.20 - 79.82) and 51.45 for Q2; the first quarters are the highest every year.
- Additional finding: Automatic selection prefers the multiplicative seasonal model in these data (SSE 105.21, AICc 61.02); the seasonal amplitude may grow with the level.
- APA: Holt-Winters additive model, α = .26, β = .00, γ = .45, SSE = 135.96, AICc = 72.31.
Related concepts
Exponential smoothing parameters (α, β, γ, φ)Forecast horizon and forecast intervalDamped trendSeasonality and seasonal periodTrain-test split (holdout)Prediction intervalSkewnessInformation criteria (AIC, AICc, BIC)
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.