Syntha Lab RU

Holt-Winters: a forecast with trend and seasonality, error on held-out periods

From a series of sales by period builds a forecast with level, trend and seasonal indices (additive or multiplicative), picks the smoothing parameters, tests the accuracy on the last held-out periods (WAPE, bias, a comparison with the naive seasonal forecast) and shows an approximate band.

Forecast and budget For whom: Planners and category managers who need a base forecast with a trend and a season for lines with history.

Calculate on your data →

At a glance

What question does it answer?
What sales will be in the coming periods if the seasonal cycle repeats and the trend holds, and whether the model beats the simple “as in the last season”.
What data is needed?
A series of sales by period in order, at least two full seasons.
How to read the result?
Example on the default values (36 months): the forecast for 12 periods is 17,919 units, the next month 1,203. On the 6 held-out months WAPE is 2.2% against 10.2% for the naive seasonal forecast; the trend is 11.6 per period. Example on the default values: Forecast for 12 periods, total: 17,919; Forecast for the next period: 1,203; WAPE on held-out periods: 2.2%; Bias on held-out periods: −0.2%. The figures are illustrative: enter your own to see your result.

Why and how to reason

A sales series consists of three parts: level (where we are now), trend (where we are heading) and seasonality (how the month affects sales). The Holt-Winters method tracks all three and updates them with each new period, giving more weight to recent data. The level is updated as α × (actual ÷ index) + (1 − α) × (level + trend), the trend as β × (change in level) + (1 − β) × previous trend, the seasonal index as γ × (actual ÷ level) + (1 − γ) × previous. The forecast h periods ahead equals (level + h × trend) × the index of the period. The calculation picks the parameters itself (α 5%, β 1%, γ 50%) and tests accuracy on held-out periods: WAPE 2.2% against 10.2% for the naive seasonal forecast “same as last year”. Level at the end of the series is 1,420, trend 11.6 per period, seasonal range 1.73.

When to use it

It comes in handy for series with clear seasonality and no sharp breaks: product groups, basic range, store traffic, where there are at least two full seasons. Do not use it for series shorter than two seasons, with rare sales and after radical events (closure, change of assortment). Assumptions: seasonality repeats, trend changes smoothly, data is cleaned of shortages and promotions.

How to read the result

The key number is the error on held-out periods (WAPE 2.2%) and the gain against the naive forecast (8 points). If there is no gain, a complex model is not needed, take the naive one. Bias (minus 0.16%) shows the model does not systematically over- or under-forecast. Parameters when fitted: low α and β mean the series is stable, and a high γ tells of changing seasonality. The seasonal range of 1.73 is the strength of seasonality: it suggests how much an order should grow toward the peak. The approximate interval shows the forecast is not a point. It is worth checking the forecast by eye on a chart and match it with the sales plan.

Common mistakes

Feeding a series with gaps and shortages (the model will learn the shortage as a season), taking the season length not matching the data period, fitting parameters on all data without held-out periods, believing the interval as a guarantee. Add known events (promotions, a shifted season) separately, the model does not know about them.

What data is needed?

  • A series of sales by period in order, at least two full seasons.
  • The season length and the seasonality type: in fashion usually multiplicative.
  • How many last periods to hold out for the accuracy test.

Inputs

FieldUnitValue in the exampleMeaning
Season lengthperiods12How many periods in one seasonal cycle: 12 for months, 52 for weeks, 4 for quarters. At least two full seasons of data are needed. Footwear, accessories and children’s have their own values: take the category data.
Seasonality typeMultiplicative: the season grows with salesIn fashion seasonality is usually proportional to the sales level, so multiplicative is taken; additive suits a season swing independent of the level. Multiplicative needs sales above zero in every period.
Smoothing parametersPick on a gridThe pick scans a grid of values and takes those with the smallest sum of squared one-step errors on the series; the result is reproducible. Set your own if you know how fast demand changes.
Level smoothing (α)%30The higher, the faster the level follows the latest sales. Used if the parameters are set by hand. The example value is illustrative.
Trend smoothing (β)%10The higher, the faster the trend estimate changes; the trend is usually changed slowly. Used with manual parameters. The example value is illustrative.
Seasonal smoothing (γ)%30The higher, the faster the seasonal indices are updated. Used with manual parameters. The example value is illustrative.
Forecast horizonperiods12How many periods ahead to forecast. The further, the wider the band.
Periods to hold out for the testperiods6The model is built on the series without these periods, forecasts them and the forecast is compared with the actual. Zero: no test. At least two seasons must remain after the hold-out.

How it is calculated

StepFormula
Forecast(level + h × trend) × the period’s seasonal index
Levelα × (actual ÷ index) + (1 − α) × (level + trend)
Trendβ × (new level − old level) + (1 − β) × the old trend
Seasonγ × (actual ÷ (level + trend)) + (1 − γ) × the old index
Startthe first season’s level, the trend from the gap between the means of the first two seasons, indices from deviations from that trend
TestWAPE = Σ|actual − forecast| ÷ Σactual on the held-out periods

Worked numbers

Default values from the app; in the interactive version you enter your own.

Forecast for 12 periods, total
17,919
Forecast for the next period
1,203
WAPE on held-out periods
2.2 %The sum of |actual − forecast| ÷ the sum of actuals; the model has not seen the held-out periods. The colour by 15 and 30% thresholds is illustrative.
Bias on held-out periods
−0.2 %Plus: the model over-forecasts, minus: under-forecasts.
MAPE on held-out periods
2.1 %
WAPE of the naive seasonal forecast
10.2 %A forecast “as in the same period of the last season”. If the model is no better than it, the complex model is not needed.
Gain against the naive, p.p.
8
Level at the end of the series
1,420
Trend per period
11.6
Seasonal swing: the largest index to the smallest
1.73×
Parameter α (level)
5 %
Parameter β (trend)
1 %
Parameter γ (season)
50 %
Actual and forecast with an approximate 80% band
Line chart: Actual: 880 → 1,950; Forecast: 1,950 → 2,162; Lower bound: 1,143 → 1,955; Upper bound: 1,263 → 2,3705001,0001,5002,0002,500Jan 2023Jan 2024Jan 2025P+1P+12Actual, Jan 2023: 880Forecast, P+12: 2,162Lower bound, P+12: 1,955Upper bound, P+12: 2,370₽Period
  • Actual
  • Forecast
  • Lower bound
  • Upper bound
Data as a table
PeriodActualForecastLower boundUpper bound
Jan 2023880———
Feb 2023910———
Mar 20231,000———
Apr 20231,000———
May 2023930———
Jun 2023910———
Jul 2023880———
Aug 2023950———
Sep 20231,150———
Oct 20231,280———
Nov 20231,380———
Dec 20231,640———
Jan 2024960———
Feb 20241,010———
Mar 20241,190———
Apr 20241,110———
May 20241,090———
Jun 20241,050———
Jul 2024940———
Aug 20241,110———
Sep 20241,310———
Oct 20241,430———
Nov 20241,630———
Dec 20241,750———
Jan 20251,070———
Feb 20251,190———
Mar 20251,300———
Apr 20251,280———
May 20251,220———
Jun 20251,100———
Jul 20251,100———
Aug 20251,240———
Sep 20251,450———
Oct 20251,650———
Nov 20251,710———
Dec 20251,9501,950——
P+1—1,2031,1431,263
P+2—1,2951,2101,379
P+3—1,4401,3361,544
P+4—1,4011,2811,521
P+5—1,3341,2001,468
P+6—1,2461,0991,392
P+7—1,2011,0421,359
P+8—1,3521,1831,522
P+9—1,5951,4151,775
P+10—1,7831,5941,973
P+11—1,9081,7092,106
P+12—2,1621,9552,370
Seasonal indices, % of the level
Bar chart: 1 84, 2 90, 3 99, 4 96, 5 90, 6 8405010015011Index: 848422Index: 909033Index: 999944Index: 969655Index: 909066Index: 848477Index: 808088Index: 898999Index: 1051051010Index: 1161161111Index: 1231231212Index: 139139%
Data as a table
MeasureIndex
184
290
399
496
590
684
780
889
9105
10116
11123
12139

Forecast by period

PeriodForecastLower bound 80%Upper bound 80%
P+11,2031,1431,263
P+21,2951,2101,379
P+31,4401,3361,544
P+41,4011,2811,521
P+51,3341,2001,468
P+61,2461,0991,392
P+71,2011,0421,359
P+81,3521,1831,522
P+91,5951,4151,775
P+101,7831,5941,973
P+111,9081,7092,106
P+122,1621,9552,370

A series of 36 periods, season 12. The parameters are picked on a grid over the whole series; the error on the held-out periods uses parameters picked only on the series before them. The band is approximate (80%): from the spread of one-step errors, growing with the root of the horizon.

For new models without history, and periods with stock-outs and advertised promotions the series is distorted: clean it first. Footwear, accessories and children’s have their own season and length. Rare sizes with zeros are counted by the intermittent demand.

Nuances and limits

  • Leaving periods with stock-outs and promotions in the series: the season is distorted.
  • Judging the model by its error on the very data it was fitted on: a test on held-out periods is needed.
  • Forecasting a new style so: there is no series, take analogs.

Terms used

Calculate on your data →