Syntha Lab RU

Holt-Winters, seasonal decomposition and Croston: forecasting with trend, season and intermittent demand

Three methods that cover what plain smoothing and a seasonal index cannot do: triple exponential smoothing (level, trend, season), decomposition of a series into trend, seasonality and remainder, and Croston’s method for items that do not sell every week. The decision: which method to use for a growing or declining line and for rare sizes, and how to show an interval rather than one point.

Math and analytics Brand and retail Verified: 2026-11-14

Open interactively →

Why this document

The planner and the analyst read this when simple smoothing and a seasonal index stop working. Holt-Winters is for series that have both a rising or falling trend and a repeating season: simple smoothing lags such a series, a seasonal index misses the trend. Decomposition into trend, season and remainder shows where growth came from. The Croston method is for items that sell rarely and in ones: ordinary methods give false precision there, and stock is either frozen or suddenly runs out.

At a glance

What is this document and why is it needed?
Three methods that cover what plain smoothing and a seasonal index cannot do: triple exponential smoothing (level, trend, season), decomposition of a series into trend, seasonality and remainder, and Croston’s method for items that do not sell every week. The decision: which method to use for a growing or declining line and for rare sizes, and how to show an interval rather than one point.
Who owns the document and how often is it rebuilt?
Demand planning analyst; parameters are reviewed once a season, the forecast is updated monthly (lines) and weekly (rare items); read by purchasing, supply planning and finance.
What mistakes are the most common?
α, β, γ are fitted on the whole series and tested on it → false accuracy → a hold-out window excluded from tuning. A linear trend is extended 18-24 months → the forecast flies up or down → limit the horizon or damp the trend. Sales rather than demand are forecast with frequent stock-outs → trend and season are understated → reconstruct demand on out-of-stock days.

For whom

Brand
A brand forecasts a line or category at the order horizon (4-9 months): Holt-Winters and decomposition give a baseline with trend, and Croston serves rare SKUs in the top-up warehouse.
Retail
A retailer applies them to a category and store at the season purchasing horizon and to rare sizes at the replenishment horizon (1-4 weeks), where ordinary WAPE misleads.

What it consists of

When a plain index is not enough

  • A seasonal index and a smoothed level (the demand forecasting document) fail when a line has a trend: store growth, range expansion, a declining category. In the example a 48-month series grows by about RUB 4.4 thousand a month on top of an annual pattern; a method without seasonality (Holt) errs by 38.3% WAPE on the hold-out because it does not know the December peak.
  • Selection rule: no trend and no season, plain smoothing; season without trend, a seasonal index; trend and season together, Holt-Winters or decomposition; many zeros in the series, Croston. A method that does not beat seasonal naive is not adopted.
  • The multiplicative form fits when the seasonal swing grows with the level (December is 35% higher, not RUB 300 thousand higher); the additive form when the swing is constant in roubles. For apparel sales it is usually multiplicative.

Holt-Winters triple smoothing

  • Three states are updated at every step: the level L (deseasonalized volume), the trend T (level gain per period) and the seasonal coefficients S by position within the year. The parameters α, β, γ from 0 to 1 set how fast the method forgets the past for each state.
  • Parameters are picked by a grid minimizing the sum of squared one-step errors on the training window (36 months). The example selects α = 0.2, β = 0.01, γ = 0.05: the level reacts moderately and trend and season barely twitch. Small β and γ protect from noise but slow the reaction to a real break.
  • Forecast h steps ahead: (L + h·T)·S. On the 12-month hold-out the method gives WAPE 4.4% against 6.2% for seasonal naive. The 80% interval is built from one-step residuals with width growing as √h; in the example 11 of 12 months fell inside, so the interval is conservative.
  • Retailer: for a new store without history trend and season are taken from analog stores of the format and city; its own Holt-Winters is switched on after 24 months.

Seasonal decomposition

  • Classical decomposition: the trend is a centerd 2×12 moving average (it removes season and noise), the series is divided by the trend, ratios are averaged by month, indices are normalized to an average of 1 and a remainder is left. For the future the trend is extended by a line through the moving average. In the example hold-out WAPE is 3.6% and the remainder spread is 2.5%.
  • The remainder is read as a diagnostic: repeating spikes in one week mean a missed event (a promotion, a moving holiday), a run of one sign means the trend is wrong. Decomposition is more transparent for an owner, Holt-Winters is handier for monthly updates.
  • The method needs at least two full years, preferably three; for lines shorter than two years the season is taken from analogs, not from own history.
  • Brand: the wholesale season leads retail by 2-4 months (partner buy-in), so indices are computed separately by channel, otherwise the shipment peak reads as a demand peak.
  • Retailer: the season depends on region and weather, so indices by climate zone or store group are more accurate than a single chain index if the group has enough history.

Croston’s method for intermittent demand

  • If a series has many zeros, ordinary smoothing gives a forecast that jumps down after weeks without sales and up after sales. Croston smooths the sale size z and the interval p between sales separately and takes the forecast z/p. In the example of 300 simulated items the share of weeks with a sale averages 30%.
  • The method’s bias is known: it overstates the demand rate. The Syntetos-Boylan correction multiplies the forecast by (1 − α/2). Against the known true demand rate: Croston bias 6%, SBA 0.7%, root-mean-square error 28% and 26% against 69% for an 8-week moving average.
  • For rare items weekly WAPE is almost always near 100%: a zero forecast “errs” less than any non-zero one. Quality is checked by cumulative demand over the lead time and by the share of stock-outs, not by weekly error.

Brand and retailer: what to apply where

  • Brand: Holt-Winters by line and channel on a monthly grid for the order plan and OTB; decomposing the remainder helps separate the collection effect from the channel trend.
  • Retailer: by category and store on a weekly or monthly grid; the store trend is accounted for after an opening, a refit and a change of neighbors.
  • Brand: Croston serves top-up warehouse SKUs and edge sizes with zero weekly shipments; safety stock is computed on cumulative demand over the lead time.
  • Retailer: Croston serves store sizes and accessories with more than half zero weeks; the result feeds min-max replenishment.
  • Brand: interval bounds are turned into order scenarios (optimistic, base, cautious), not shown to an owner as a guarantee.
  • Retailer: the forecast interval sets safety stock and the size of the promotion allowance; actuals leaving the interval two months in a row start a parameter recalculation.

Where the data comes from

Data sourceFields
Sales by month and weekA series of at least 24 months (preferably 36), demand reconstructed on out-of-stock days, events flagged; for Croston weekly unit sales by SKU and size.
Calendar and eventsPromotions, holiday shifts, openings and closures, 4-5-4 calendar changes; they are removed from the series or modeled separately.
AnalogsSeasonal profiles of similar lines for short series; confidence bounds from error history.

How to build it step by step

  1. Step 1. Build the monthly series, clean it of events and out-of-stock days; check the length (at least two full years).
  2. Brand: Step 2. Take the series by line and channel from shipments and sell-out; separately flag pre-orders and partner buying-in that distort the trend.
  3. Retailer: Step 3. Take the series by category and store; drop months of closure and refit, move promotions into a separate flag.
  4. Step 4. Choose the season form: multiplicative if the swing grows with the level; check on the plot of series ÷ trend.
  5. Step 5. Keep the last 12 months as a hold-out; on the rest pick α, β, γ by a grid on the sum of squared one-step errors.
  6. Step 6. Compute the forecast on the hold-out and compare WAPE and bias with seasonal naive and decomposition; choose the method that beats the benchmark.
  7. Step 7. Build the 80% interval from one-step residuals (width grows as √h) and check how many actuals fell inside.
  8. Brand: Step 8. For rare top-up warehouse SKUs compute Croston with the SBA correction and demand over the lead time; hand the result to safety stock.
  9. Retailer: Step 9. For sizes and accessories with a zero share above 50% replace the moving average by Croston in min-max replenishment; track stock-outs, not weekly error.
  10. Step 10. Fix the parameter recalculation frequency (once a season) and the condition for an unplanned recalculation (actuals outside the interval two periods in a row).

Formulas

MeasureFormulasNote
Holt-Winters (multiplicative)L_t = α·y_t/S_(t−m) + (1−α)·(L_(t−1)+T_(t−1)); T_t = β·(L_t−L_(t−1)) + (1−β)·T_(t−1); S_t = γ·y_t/L_t + (1−γ)·S_(t−m)m = 12 for months, 52 for weeks; parameters between 0 and 1.
Forecast h steps aheadF_(t+h) = (L_t + h·T_t)·S_(t+h−m·⌈h/m⌉)A linear trend overstates growth on long horizons: limit the horizon or damp the trend.
Seasonal index from decompositionĨ_i = mean_k( y_(k,i) / MA_(k,i) ); I_i = Ĩ_i / mean(Ĩ)MA is the centerd 2×12; it is undefined at the series edges.
Remainder and spreade_t = y_t / (Trend_t · I_t) − 1; s_e = √[ Σ(e_t − ē)² / (n − 1) ]Sample SD, divisor n−1.
Crostonz_t = α·y_t + (1−α)·z_(t−1), p_t = α·q + (1−α)·p_(t−1) (only in weeks with a sale); F = z/pq is the number of weeks since the previous sale; α about 0.1.
SBA correction and lead-time demandF_SBA = (1 − α/2)·z/p; D_L = L·F_SBASafety stock for rare items is computed from the distribution of D_L, not from a normal approximation of weekly demand.

Formulas in notation

Holt-Winters: level
L_t = α·y_t / S_{t−m} + (1 − α)·(L_{t−1} + T_{t−1})
The deseasonalized level.
Holt-Winters: trend
T_t = β·(L_t − L_{t−1}) + (1 − β)·T_{t−1}
Level gain per period.
Holt-Winters: season
S_t = γ·y_t / L_t + (1 − γ)·S_{t−m}
The seasonal coefficient of a position within the year.
Forecast
F_{t+h} = (L_t + h·T_t)·S_{t+h−m}
For h ≤ m; beyond that the season repeats.
Seasonal index
I_i = mean_k( y_{k,i} / MA_{k,i} ) / mean(Ĩ)
MA is the centerd 2×12; normalized to an average of 1.
SBA correction
F_SBA = (1 − α/2)·z / p
Removes Croston’s overstatement.

Worked example

Four methods on a 12-month hold-out (multi-step forecast from the end of the training window)

MethodWAPE, %Bias, %
Seasonal naive (a year ago)6.2−4.8
Holt (trend without season)38.338.3
Holt-Winters (multiplicative)4.4−1.2
Decomposition: trend × index3.6−1.4

The 48-month series is generated in code (fixed seed): a linear trend, an annual pattern and 5% noise. Training 36 months, hold-out 12. The best methods with season and trend (3.6 and 4.4%), without season 38.3%. The data is illustrative; the method order on real series may differ.

Figures

Actual and forecasts on the 12-month hold-out
Line chart: Actual: 939 → 1,566; Holt-Winters: 860 → 1,645; Decomposition: 936 → 1,605; Holt without season: 1,496 → 1,7788001,0001,2001,4001,6001,800123456789101112Actual, 1: 939Actual, 2: 910Actual, 3: 1,084Actual, 4: 1,218Actual, 5: 1,137Actual, 6: 1,061Actual, 7: 919Actual, 8: 1,097Actual, 9: 1,324Actual, 10: 1,465Actual, 11: 1,482Actual, 12: 1,566Holt-Winters, 1: 860Holt-Winters, 2: 877Holt-Winters, 3: 1,065Holt-Winters, 4: 1,121Holt-Winters, 5: 1,246Holt-Winters, 6: 981Holt-Winters, 7: 897Holt-Winters, 8: 1,075Holt-Winters, 9: 1,272Holt-Winters, 10: 1,469Holt-Winters, 11: 1,518Holt-Winters, 12: 1,645Decomposition, 1: 936Decomposition, 2: 904Decomposition, 3: 1,089Decomposition, 4: 1,113Decomposition, 5: 1,248Decomposition, 6: 976Decomposition, 7: 893Decomposition, 8: 1,069Decomposition, 9: 1,249Decomposition, 10: 1,440Decomposition, 11: 1,486Decomposition, 12: 1,605Holt without season, 1: 1,496Holt without season, 2: 1,522Holt without season, 3: 1,547Holt without season, 4: 1,573Holt without season, 5: 1,599Holt without season, 6: 1,624Holt without season, 7: 1,650Holt without season, 8: 1,676Holt without season, 9: 1,701Holt without season, 10: 1,727Holt without season, 11: 1,753Holt without season, 12: 1,778RUB thousand
  • Actual
  • Holt-Winters
  • Decomposition
  • Holt without season
Data as a table
MeasureActualHolt-WintersDecompositionHolt without season
19398609361,496
29108779041,522
31,0841,0651,0891,547
41,2181,1211,1131,573
51,1371,2461,2481,599
61,0619819761,624
79198978931,650
81,0971,0751,0691,676
91,3241,2721,2491,701
101,4651,4691,4401,727
111,4821,5181,4861,753
121,5661,6451,6051,778

Actual and forecasts on the 12-month hold-out. Holt-Winters (4.4%) and decomposition (3.6%) follow both the growth and the seasonal pattern; Holt without season (38.3%) gives an almost straight line. Axis: month of the hold-out.

Holt-Winters forecast with an 80% interval
Line chart: Upper 80%: 939 → 1,918; Forecast: 860 → 1,645; Actual: 939 → 1,566; Lower 80%: 781 → 1,3715001,0001,5002,000123456789101112Upper 80%, 1: 939Upper 80%, 2: 988Upper 80%, 3: 1,202Upper 80%, 4: 1,280Upper 80%, 5: 1,422Upper 80%, 6: 1,175Upper 80%, 7: 1,107Upper 80%, 8: 1,298Upper 80%, 9: 1,509Upper 80%, 10: 1,719Upper 80%, 11: 1,780Upper 80%, 12: 1,918Forecast, 1: 860Forecast, 2: 877Forecast, 3: 1,065Forecast, 4: 1,121Forecast, 5: 1,246Forecast, 6: 981Forecast, 7: 897Forecast, 8: 1,075Forecast, 9: 1,272Forecast, 10: 1,469Forecast, 11: 1,518Forecast, 12: 1,645Actual, 1: 939Actual, 2: 910Actual, 3: 1,084Actual, 4: 1,218Actual, 5: 1,137Actual, 6: 1,061Actual, 7: 919Actual, 8: 1,097Actual, 9: 1,324Actual, 10: 1,465Actual, 11: 1,482Actual, 12: 1,566Lower 80%, 1: 781Lower 80%, 2: 765Lower 80%, 3: 928Lower 80%, 4: 963Lower 80%, 5: 1,069Lower 80%, 6: 788Lower 80%, 7: 688Lower 80%, 8: 851Lower 80%, 9: 1,035Lower 80%, 10: 1,219Lower 80%, 11: 1,256Lower 80%, 12: 1,371RUB thousand
  • Upper 80%
  • Forecast
  • Actual
  • Lower 80%
Data as a table
MeasureUpper 80%ForecastActualLower 80%
1939860939781
2988877910765
31,2021,0651,084928
41,2801,1211,218963
51,4221,2461,1371,069
61,1759811,061788
71,107897919688
81,2981,0751,097851
91,5091,2721,3241,035
101,7191,4691,4651,219
111,7801,5181,4821,256
121,9181,6451,5661,371

Holt-Winters forecast with an 80% interval. The interval width grows with the horizon as √h; 11 of 12 actual months fell inside, so the interval is conservative.

Seasonal indices: set and recovered by decomposition
Bar chart: 1 0.8, 2 0.78, 3 0.92, 4 1, 5 1.02, 6 0.900.511.511Built-in index: 0.80.8Recovered index: 0.830.8322Built-in index: 0.780.78Recovered index: 0.790.7933Built-in index: 0.920.92Recovered index: 0.950.9544Built-in index: 11Recovered index: 0.970.9755Built-in index: 1.021.02Recovered index: 1.081.0866Built-in index: 0.90.9Recovered index: 0.840.8477Built-in index: 0.760.76Recovered index: 0.760.7688Built-in index: 0.90.9Recovered index: 0.910.9199Built-in index: 1.121.12Recovered index: 1.061.061010Built-in index: 1.21.2Recovered index: 1.221.221111Built-in index: 1.251.25Recovered index: 1.251.251212Built-in index: 1.351.35Recovered index: 1.341.34to the yearly mean
  • Built-in index
  • Recovered index
Data as a table
MeasureBuilt-in indexRecovered index
10.80.83
20.780.79
30.920.95
410.97
51.021.08
60.90.84
70.760.76
80.90.91
91.121.06
101.21.22
111.251.25
121.351.34

Seasonal indices: set and recovered by decomposition. Decomposition over 36 months recovers the built-in annual pattern with small deviations; a peak in November-December and a trough in January-February and July. Axis: month of the year.

One-step error for different α and γ
Heat map: 7×7; min 61.7, max 78.5γ 0.05γ 0.05γ 0.1γ 0.1γ 0.2γ 0.2γ 0.3γ 0.3γ 0.5γ 0.5γ 0.7γ 0.7γ 0.9γ 0.9α 0.1α 0.162α 0.1 / γ 0.05: 6262.5α 0.1 / γ 0.1: 62.563.8α 0.1 / γ 0.2: 63.865.3α 0.1 / γ 0.3: 65.368.7α 0.1 / γ 0.5: 68.773α 0.1 / γ 0.7: 7378.5α 0.1 / γ 0.9: 78.5α 0.2α 0.261.7α 0.2 / γ 0.05: 61.762.2α 0.2 / γ 0.1: 62.263.2α 0.2 / γ 0.2: 63.264.4α 0.2 / γ 0.3: 64.467.2α 0.2 / γ 0.5: 67.270.8α 0.2 / γ 0.7: 70.875.4α 0.2 / γ 0.9: 75.4α 0.3α 0.362.7α 0.3 / γ 0.05: 62.763.1α 0.3 / γ 0.1: 63.164α 0.3 / γ 0.2: 6464.9α 0.3 / γ 0.3: 64.967.1α 0.3 / γ 0.5: 67.169.8α 0.3 / γ 0.7: 69.873.2α 0.3 / γ 0.9: 73.2α 0.4α 0.464.1α 0.4 / γ 0.05: 64.164.5α 0.4 / γ 0.1: 64.565.2α 0.4 / γ 0.2: 65.265.9α 0.4 / γ 0.3: 65.967.5α 0.4 / γ 0.5: 67.569.4α 0.4 / γ 0.7: 69.471.8α 0.4 / γ 0.9: 71.8α 0.5α 0.565.8α 0.5 / γ 0.05: 65.866.1α 0.5 / γ 0.1: 66.166.7α 0.5 / γ 0.2: 66.767.2α 0.5 / γ 0.3: 67.268.4α 0.5 / γ 0.5: 68.469.8α 0.5 / γ 0.7: 69.871.3α 0.5 / γ 0.9: 71.3α 0.6α 0.667.5α 0.6 / γ 0.05: 67.567.8α 0.6 / γ 0.1: 67.868.3α 0.6 / γ 0.2: 68.368.7α 0.6 / γ 0.3: 68.769.6α 0.6 / γ 0.5: 69.670.6α 0.6 / γ 0.7: 70.671.7α 0.6 / γ 0.9: 71.7α 0.8α 0.871.3α 0.8 / γ 0.05: 71.371.4α 0.8 / γ 0.1: 71.471.7α 0.8 / γ 0.2: 71.772α 0.8 / γ 0.3: 7272.6α 0.8 / γ 0.5: 72.673.1α 0.8 / γ 0.7: 73.173.6α 0.8 / γ 0.9: 73.6
  • 61.7 … 78.5 RUB thousand
Data as a table
γ 0.05γ 0.1γ 0.2γ 0.3γ 0.5γ 0.7γ 0.9
α 0.16262.563.865.368.77378.5
α 0.261.762.263.264.467.270.875.4
α 0.362.763.16464.967.169.873.2
α 0.464.164.565.265.967.569.471.8
α 0.565.866.166.767.268.469.871.3
α 0.667.567.868.368.769.670.671.7
α 0.871.371.471.77272.673.173.6

One-step error for different α and γ. One-step root-mean-square error on the training window at β = 0.01: the minimum is at small α and γ; fast parameters react to noise and err more.

Intermittent demand: bias and error of the demand rate estimate
Bar chart: Moving avg 8 5.3, Smoothing 0.2 3.1, Croston 6, SBA 0.7020406080Moving avg 8Moving avg 8Bias: 5.35.3Root-mean-square error: 68.768.7Smoothing 0.2Smoothing 0.2Bias: 3.13.1Root-mean-square error: 6464CrostonCrostonBias: 66Root-mean-square error: 28.128.1SBASBABias: 0.70.7Root-mean-square error: 26.126.1%
  • Bias
  • Root-mean-square error
Data as a table
MeasureBiasRoot-mean-square error
Moving avg 85.368.7
Smoothing 0.23.164
Croston628.1
SBA0.726.1

Intermittent demand: bias and error of the demand rate estimate. Croston is biased by 6%, SBA by 0.7%; both have a root-mean-square error less than half that of the moving average and plain smoothing. 300 simulated items, 78 training weeks, evaluated against the true demand rate.

One rare item: sales and three forecasts
Line chart: Sales, units: 0 → 0; Moving avg 8: 1.13 → 0.13; Croston: 1.42 → 0.64; SBA: 1.35 → 0.610123491317212529333741454952Sales, units, 9: 0Sales, units, 52: 0Moving avg 8, 9: 1.13Moving avg 8, 52: 0.13Croston, 9: 1.42Croston, 52: 0.64SBA, 9: 1.35SBA, 52: 0.61units per week
  • Sales, units
  • Moving avg 8
  • Croston
  • SBA
Data as a table
MeasureSales, unitsMoving avg 8CrostonSBA
901.131.421.35
1001.131.421.35
1120.751.421.35
1200.751.241.18
1300.751.241.18
1400.381.241.18
1500.381.241.18
1600.251.241.18
1700.251.241.18
1800.251.241.18
1900.251.241.18
20001.241.18
21301.241.18
2220.380.920.87
2300.630.960.91
2400.630.960.91
2500.630.960.91
2600.630.960.91
2730.630.960.91
28010.90.85
29010.90.85
3010.630.90.85
3100.50.840.8
3200.50.840.8
3320.50.840.8
3430.750.820.78
3501.130.90.86
3610.750.90.86
3740.880.870.83
3801.3810.95
3901.2510.95
4001.2510.95
4101.2510.95
420110.95
4300.6310.95
4400.6310.95
4500.510.95
460010.95
470010.95
480010.95
490010.95
500010.95
511010.95
5200.130.640.61

One rare item: sales and three forecasts. In 73% of weeks there are no sales; the moving average jumps, Croston and SBA give a steady level and a single forecast for stock planning. Axis: week.

Specifics by category

Apparel
A multiplicative season with two peaks; lines with a trend (network expansion, online) are forecast by Holt-Winters, basic knitwear more simply by a seasonal index.
Footwear
Seasonal switches (mid-season, winter) are sharp: γ is higher than for apparel; sizes are forecast by Croston within a model, especially the edge ones.
Accessories
Many rare items and gift peaks: Croston for the tail, weekly decomposition around holiday dates for the core.
Kids
The school and age calendar creates sharp seasonal jumps; the trend depends on demography; a series by age group, not for the whole category.
Men
Weak seasonality and a steady trend: double smoothing or decomposition with few indices is enough; Croston for rare sizes.
Women
A high share of new items and fashion waves: series trends are unstable, β is kept small, the interval is shown wider; new items by analogs, Holt-Winters for mature lines.

Common mistakes

  • α, β, γ are fitted on the whole series and tested on it → false accuracy → a hold-out window excluded from tuning.
  • A linear trend is extended 18-24 months → the forecast flies up or down → limit the horizon or damp the trend.
  • Sales rather than demand are forecast with frequent stock-outs → trend and season are understated → reconstruct demand on out-of-stock days.
  • Plain smoothing is applied to series that are mostly zeros → the forecast jumps and triggers false reorders → Croston or SBA.
  • Rare items are evaluated by weekly WAPE → a zero forecast looks “best” → evaluation by lead-time cumulative demand and stock-outs.

Russian specifics

  • Retailer: the Russian calendar with long holiday weeks in January and May shifts the season by week; for weekly series use an event calendar or weeks around dates, not the week number.
  • Forecasts built on personal customer data require a legal basis and anonymization; demand by store and SKU needs no person identifiers. Rule: Law 152-FZ of 27.07.2006: Federal Law of 27.07.2006 No. 152-FZ on personal data (rule check not completed).

Open interactively →