SBA and TSB: The Corrections to Croston's Method
By MLAIA Data Science Ltd. · Published 10 November 2026 · Updated 10 November 2026
SBA (the Syntetos–Boylan Approximation) and TSB (the Teunter–Syntetos–Babai method) are the two standard corrections to Croston's method for intermittent demand. SBA multiplies Croston's forecast by (1 − α/2) to remove its upward bias; with α = 0.2 that turns a Croston forecast of 2.09 units per period into 1.88. TSB replaces the smoothed demand interval with a smoothed probability of demand that is updated every period, so in the worked example below its forecast falls from 2.22 to 1.14 across three consecutive zero-demand periods, while Croston and SBA stay frozen at their last values.
Croston's method in one paragraph
Croston (1972) splits an intermittent series into the sizes of the non-zero demands, z, and the intervals between them, p, and smooths each with a constant α, updating both only in periods where demand occurs:
ẑt = α·zt + (1 − α)·ẑt−1 p̂t = α·pt + (1 − α)·p̂t−1 forecast = ẑt / p̂t
"Demand averages ẑ units and comes every p̂ periods, so the rate is ẑ/p̂ per period." The decomposition avoids the spike-and-decay behaviour of exponential smoothing on the raw series, but it has two weaknesses that the corrections address.
SBA: removing the bias
Croston assumed that the expected value of ẑ/p̂ equals the ratio of the two expected values. It does not. Because p̂ is in the denominator, E[1/p̂] is larger than 1/E[p̂] (Jensen's inequality), so the estimator systematically over-forecasts the demand rate. Syntetos and Boylan (2001, 2005) quantified the bias and proposed the correction now called the Syntetos–Boylan Approximation:
forecastSBA = (1 − α/2) · ẑt / p̂t
With α = 0.2 the factor is 0.9; with α = 0.1 it is 0.95. The correction is approximately unbiased, and in the empirical comparison of Syntetos and Boylan (2005) it beat the original on a large sample of real intermittent series. The ADI / CV² classification cutoffs were themselves derived from the regions where SBA's expected error is lower than Croston's, which is everywhere except smooth demand.
TSB: letting the forecast decay
Croston and SBA update only when demand occurs. If an item stops selling, the forecast freezes at its last value forever and any replenishment rule driven by it keeps ordering stock for a dead item. Teunter, Syntetos and Babai (2011) replaced the interval with the probability of demand, d̂, which is updated in every period from the indicator dt (1 if demand occurred, 0 otherwise), with its own constant β:
d̂t = β·dt + (1 − β)·d̂t−1 ẑt updated on demand as in Croston forecastTSB = d̂t · ẑt
Through a run of zeros, d̂ shrinks by a factor (1 − β) each period, so the forecast declines geometrically toward zero; a new demand lifts it back up. TSB is slightly biased, but it is the standard choice for portfolios with real obsolescence risk.
Worked example: all three methods on one series
Ten periods of demand: 0, 4, 0, 0, 6, 0, 5, 0, 0, 0 (15 units in total, so the plain average is 1.5 per period). Smoothing constants α = 0.2 for sizes and intervals and β = 0.2 for the TSB probability. All three methods are initialised at the first demand in period 2: size 4, interval 2 (the two periods it took for demand to appear), and probability 1 ÷ 2 = 0.5.
| Period | Demand | ẑ (size) | p̂ (interval) | Croston ẑ/p̂ | SBA 0.9·ẑ/p̂ | d̂ (probability) | TSB d̂·ẑ |
|---|---|---|---|---|---|---|---|
| 2 | 4 | 4.00 | 2.00 | 2.00 | 1.80 | 0.500 | 2.00 |
| 3 | 0 | 4.00 | 2.00 | 2.00 | 1.80 | 0.8·0.500 = 0.400 | 1.60 |
| 4 | 0 | 4.00 | 2.00 | 2.00 | 1.80 | 0.320 | 1.28 |
| 5 | 6 | 0.2·6 + 0.8·4 = 4.40 | 0.2·3 + 0.8·2 = 2.20 | 2.00 | 1.80 | 0.2 + 0.8·0.320 = 0.456 | 2.01 |
| 6 | 0 | 4.40 | 2.20 | 2.00 | 1.80 | 0.365 | 1.61 |
| 7 | 5 | 0.2·5 + 0.8·4.4 = 4.52 | 0.2·2 + 0.8·2.2 = 2.16 | 2.09 | 1.88 | 0.2 + 0.8·0.365 = 0.492 | 2.22 |
| 8 | 0 | 4.52 | 2.16 | 2.09 | 1.88 | 0.393 | 1.78 |
| 9 | 0 | 4.52 | 2.16 | 2.09 | 1.88 | 0.315 | 1.42 |
| 10 | 0 | 4.52 | 2.16 | 2.09 | 1.88 | 0.252 | 1.14 |
Three things to read off the table. First, SBA is always 10% below Croston here, because the factor is fixed by α; the gap is the bias that Croston carries. Second, after the last demand in period 7, Croston and SBA do not move: three zero periods later they still forecast 2.09 and 1.88 units a period. Third, TSB's probability has fallen from 0.49 to 0.25 over the same three periods and its forecast from 2.22 to 1.14. If the item is genuinely dying, TSB is already saying so; if the three zeros are ordinary for an item with ADI around 2.5, SBA's patience is the better behaviour. That trade-off is the whole choice between them.
Choosing between SBA and TSB
| Method | Use when | Watch out for |
|---|---|---|
| Croston (1972) | Teaching the decomposition; historical baseline | Over-forecasts; SBA dominates it in practice |
| SBA (2001/2005) | Stationary intermittent or lumpy items; the default | Never decays if demand stops |
| TSB (2011) | Items with declining demand, end-of-life risk, or long runs of zeros that should lower the forecast | Slightly biased; a second constant to set; decays even when zeros are normal for the item |
What the corrections do not fix
- They give a rate, not a distribution. A forecast of 1.88 units a period says nothing about whether lead-time demand will be 0 or 11. Setting a reorder point needs the distribution of demand over the lead time, which must be assumed (Poisson, negative binomial) or built empirically, as shown in Safety Stock and Reorder Point for Intermittent Demand.
- They do not forecast timing. None of the methods predicts which period the next demand lands in.
- They assume a stationary pattern. Trend and seasonality, where present, must be handled separately.
- They are sensitive to initialisation and α on short histories. With four demands in ten periods, as above, the first demand still carries most of the weight.
Frequently asked questions
Is SBA always better than Croston's method?
For practical purposes, yes. The bias correction is small (10% of the forecast at α = 0.2, 5% at α = 0.1) but it is systematic, and Syntetos and Boylan (2005) found SBA more accurate than Croston on a large sample of real intermittent series. The ADI / CV² classification was derived from the regions where SBA beats Croston, which is every quadrant except smooth demand.
When should I use TSB instead of SBA?
Use TSB when a run of zero-demand periods should lower the forecast: items late in their lifecycle, parts for a shrinking installed base, or any portfolio where obsolescence is a real cost. In the worked example TSB's forecast falls from 2.22 to 1.14 across three zero periods while SBA stays at 1.88. Use SBA when zeros are normal for the item and you want a stable rate that does not react to every quiet month.
What smoothing constant should I use for SBA or TSB?
Small values. Croston-family methods are usually run with α between 0.05 and 0.2 for sizes and intervals; the worked example uses 0.2 so the arithmetic is visible. For TSB the probability constant β is typically also small, and a lower β makes the decay through zeros slower. Larger constants make the forecast chase the most recent order, which on lumpy items means chasing noise.
References
- Croston, J. D. (1972). Forecasting and stock control for intermittent demands. Operational Research Quarterly, 23(3), 289–303.
- Syntetos, A. A., & Boylan, J. E. (2001). On the bias of intermittent demand estimates. International Journal of Production Economics, 71(1–3), 457–466.
- Syntetos, A. A., & Boylan, J. E. (2005). The accuracy of intermittent demand estimates. International Journal of Forecasting, 21(2), 303–314.
- Teunter, R. H., Syntetos, A. A., & Babai, M. Z. (2011). Intermittent demand: linking forecasting to inventory obsolescence. European Journal of Operational Research, 214(3), 606–615.
- Syntetos, A. A., Boylan, J. E., & Croston, J. D. (2005). On the categorization of demand patterns. Journal of the Operational Research Society, 56(5), 495–503.
- Boylan, J. E., & Syntetos, A. A. (2021). Intermittent Demand Forecasting: Context, Methods and Applications. Wiley.
See also
- Demand Forecasting Wiki — all articles
- Croston's Method, SBA and TSB
- Intermittent and Lumpy Demand
- Lumpy Demand: Meaning, Detection and Forecasting
- Syntetos–Boylan Classification (ADI and CV²)
- Safety Stock and Reorder Point for Intermittent Demand
Try it on your own items
- Check which method fits your item with the free demand classifier — runs in your browser, no signup.
- See how NextDemand forecasts spare parts and other slow movers — item by item, with order recommendations.
- Open the live NextDemand demo →