Intermittent · Free
Teunter–Syntetos–Babai
Demand probability updates every period, so a dying item can decay to zero.
What it assumes
Demand probability updates every period, so a dying item can decay to zero.
When to use
Intermittent items with obsolescence risk.
When to avoid
Stable intermittent demand where Croston/SBA already work.
Knobs
ŷ = z × p_prob; p_prob shrinks on zeros.
How it works
TSB tracks size and occurrence probability. Probability is updated every period — including zeros — so a long silence shrinks the forecast. That is the obsolescence story Croston cannot tell. If demand is a stable intermittent drip, Croston or SBA may already be enough and TSB can under-forecast.
Sample forecast question
A replacement filter sold, then went quiet. What is TSB’s mean weekly demand for the next four weeks after those trailing zeros?
Step-by-step on these numbers
| Period | Actual | Fitted |
|---|---|---|
| W1 | 5 | 5 |
| W2 | 0 | 5 |
| W3 | 6 | 3.25 |
| W4 | 0 | 4.13 |
| W5 | 0 | 2.69 |
| W6 | 4 | 1.75 |
| W7 | 0 | 2.74 |
| W8 | 0 | 1.78 |
| W9 | 0 | 1.16 |
| W10 | 3 | 0.75 |
| W11 | 0 | 1.90 |
| W12 | 0 | 1.24 |
| W13 | 0 | 0.80 |
| W14 | 0 | 0.52 |
| Period | Forecast | 95% interval |
|---|---|---|
| W15 | 0.34 | -4.20 – 4.87 |
| W16 | 0.34 | -6.08 – 6.75 |
| W17 | 0.34 | -7.52 – 8.20 |
| W18 | 0.34 | -8.73 – 9.41 |
Parameters the engine found
alpha= 0.350level= 0.339z= 4.220p= 0
The question
A replacement filter sold, then went quiet. What is TSB’s mean weekly demand for the next four weeks after those trailing zeros? Sample weeks: W1=5, W2=0, W3=6, W4=0, W5=0, W6=4, W7=0, W8=0, W9=0, W10=3, W11=0, W12=0, W13=0, W14=0 (units).
Probability every period
TSB tracks size z and occurrence probability p. Unlike Croston, p is updated on zeros too: p ← (1−α) p, so a dying item can decay toward 0.
This sample
α = 0.350. Last 4 weeks are zero, so p has been shrinking. End states: z = 4.220, p = 0.080, ŷ = z × p = 0.34.
Forecast
Hold that mean rate: W15: 0.34; W16: 0.34; W17: 0.34; W18: 0.34.
Interval
Residual σ = 2.314. For h = 1 the 95% band is [-4.20, 4.87] around 0.34. The engine labels this a residual-Gaussian heuristic (σ√h), not a simulation interval.
Graph of this sample
The chart plots the canned table on this page (actuals, in-sample fitted, forecast, 95% interval). Not your Excel series. Free analysis never uploads raw data.
Educational only. Not investment, weather, or operational advice. In the add-in, rank is rolling-origin MASE — “best supported among candidates on this series.”