AI Glossary / Time series and signal processing / Advanced
Seasonal exponential smoothing
Definition
Seasonal exponential smoothing is a family of forecasting methods that updates a series' level and seasonal pattern as observations arrive. Some variants also update a trend. Holt-Winters methods are familiar examples, combining level, trend and seasonality through weighted updating equations rather than treating every historical observation equally.
What it means
The method maintains evolving estimates. A new observation can change the estimated level, the trend and the seasonal component associated with its position in the cycle. Smoothing parameters control how strongly these components react.
Additive seasonality expresses seasonal effects in the series' units. Multiplicative seasonality uses factors relative to the level. Simple exponential smoothing has no seasonal component and is not equivalent to a seasonal method.
How it works
Specify the seasonal period, component forms, initial states and smoothing parameters. Generate a forecast using the states available before observing the next value. Once that value arrives, update the components and produce new forecasts.
In additive Holt-Winters smoothing, the level update uses the observation after subtracting its previous seasonal component. The trend update uses the change in level. The seasonal update blends new information about that season with its previous estimate. Forecasts add the projected level and trend to the most recent component for the relevant future season.
Different equation conventions use differently scaled parameters. Record the actual update equations when reproducing a calculation.
Three practical examples
1. Recurring service demand: update forecasts when each month's workload becomes available.
2. Inventory planning: compare additive and multiplicative seasonal models when seasonal variation may grow with overall demand.
3. Forecasting evaluation: use a transparent seasonal method as a benchmark for an AI forecasting system, with the same forecast dates and available history.
Think of it this way
Think of a planner who keeps three notes: the current baseline, its direction of movement and the usual contribution of each season. New evidence revises the notes gradually. A busy season is not automatically treated as a permanent jump in the baseline.
Worked example
Consider a fictional quarterly index. Before observing the next first quarter, suppose the estimated level is 100 and the trend is two index points per quarter. The most recently available seasonal components for quarters 1 through 4 are [-10,0,15,-5].
Use additive Holt-Winters smoothing with alpha = 0.3, beta-star = 0.2 and gamma = 0.1. For the seasonal update, use the convention based on the previous level plus previous trend. The states and parameters are stipulated for teaching, not fitted from a supplied dataset.
The forecast for the arriving first quarter is 100 + 2 - 10 = 92. Suppose the observed value is 98. The forecast error is therefore 98 - 92 = 6. This is a forecast made before seeing the observation; replacing it with an updated-state calculation would obscure the timing.
First update the level. Remove the old first-quarter seasonal component from the observation: 98 - (-10) = 108. Blend this with the old nonseasonal projection of 100 + 2 = 102:
New level = 0.3 × 108 + 0.7 × 102 = 103.8.
Next update the trend. The level rose by 103.8 - 100 = 3.8. Blend that change with the previous trend of two:
New trend = 0.2 × 3.8 + 0.8 × 2 = 2.36.
Now update the first-quarter seasonal component. Relative to the previous nonseasonal projection, the new seasonal information is 98 - 100 - 2 = -4. Blend this with the old first-quarter component of -10:
New first-quarter component = 0.1 × (-4) + 0.9 × (-10) = -9.4.
The seasonal component becomes less negative because the observation was higher than expected.
At this forecast origin, the next four quarters are quarter 2, quarter 3, quarter 4 and the following quarter 1. With no further observations, use the updated level and trend plus the latest component for each forecast season:
One quarter ahead: 103.8 + 1 × 2.36 + 0 = 106.16.
Two quarters ahead: 103.8 + 2 × 2.36 + 15 = 123.52.
Three quarters ahead: 103.8 + 3 × 2.36 - 5 = 105.88.
Four quarters ahead: 103.8 + 4 × 2.36 - 9.4 = 103.84.
These are four forecasts from the same origin. We have not invented intermediate observations or updated the states using the forecasts as if they were newly measured data.
There is a convention check. Another form updates seasonality using the new level, 103.8, rather than the previous projection, 102. To reproduce our calculation, its parameter must be gamma-star = gamma/(1 - alpha) = 0.1/0.7 = 1/7.
Using that parameter gives (1/7) × (98 - 103.8) + (6/7) × (-10) = -9.4, the same result. Using 0.1 in this alternative equation would instead give -9.58. The discrepancy would reflect mixed parameter conventions, not rounding.
All values are invented. The example demonstrates one update and its point forecasts. It does not establish suitable parameter values, forecast accuracy, uncertainty coverage or the persistence of a linear trend over long horizons.
Illustrative case study
Fictional Alder Planning reproduces a seasonal forecast in two tools and finds different updated seasonal components. Its analyst compares equations and discovers that one calculation uses the previous nonseasonal projection while the other uses the updated level.
After translating the parameter convention, the teaching calculations agree. This fictional case distinguishes matching an implementation from proving its operational performance.
Common misconceptions and mistakes
"Every smoothing parameter has the same meaning across formulas." Conventions can differ.
"The next observation updates every seasonal component directly." The illustrated update concerns its matching season.
"Four forecasts require four new observations." Multiple horizons can be forecast from one origin.
"A good point forecast supplies its own interval." Uncertainty requires a statistical specification.
Business use
Record the seasonal period, initialization, component forms and forecast origin. Compare performance with simple seasonal benchmarks and monitor changes in the data. Consider whether an undamped trend is plausible at the horizons used for planning, and investigate unexpected errors before assuming the seasonal pattern has permanently changed.
People also ask
Is trend mandatory?
No; seasonal smoothing variants can omit it.
Can seasonality be multiplicative?
Yes, when appropriate to the data and model.
Can parameters be estimated?
Yes; they need not be manually selected as in this example.
Check your understanding
What was the forecast before observing 98?
Answer: Ninety-two.
What is the updated level?
Answer: 103.8.
Which seasonal component appears four quarters ahead?
Answer: The updated first-quarter value, -9.4.
Related terms
- Exponential smoothing
- Levels
- Holt-Winters method
- seasonal component
- level
- trend
- smoothing parameter
- ETS model
- forecast origin
Learn next
Replace the arriving observation of 98 with 92, keeping the previous states and parameters unchanged. Recalculate all three updates and the next-quarter forecast. Explain which updates occur even when the forecast error is zero.
More in Time series and signal processing
- Seasonal Autoregressive Integrated Moving-Average - Seasonal Autoregressive Integrated Moving-Average, commonly abbreviated SARIMA, is a time-series model that combines ordinary ARIMA terms with terms...
- Seasonal difference - A seasonal difference subtracts an earlier observation from the same position in a seasonal cycle from the current observation.
- Seasonal effect - A seasonal effect is a systematic variation associated with a recurring position in the calendar or seasonal cycle.
- Seasonal index - A seasonal index summarizes the estimated contribution of a particular position in a recurring seasonal cycle.
- Seasonal variation - Seasonal variation is recurring movement in a time series associated with a known seasonal or calendar period.
- Short-Time Fourier Transform - The Short-Time Fourier Transform, or STFT, represents how a signal's frequency content changes over time.
Work in finance? Get the free CFO AI Prompt Pack: 7 prompts for finance leaders.
Sources
Rob J Hyndman and George Athanasopoulos, Forecasting: Principles and Practice, Methods with seasonality and Holt-Winters equations: https://otexts.com/fpp3/exponential-smoothing.html
Statsmodels developers, Exponential smoothing examples, model forms and initialization: https://www.statsmodels.org/stable/examples/notebooks/generated/exponential_smoothing.html
Previous: Seasonal effect Next: Seasonal index