Flaw of averages: the answer at the average is not the average answer
Three , 22, 30 and 38 days, average out to 30. Run 30 through the cost curve and it charges $7,200. Run all three and the average of what they charge is $9,000, and the whole of the difference came from averaging first.
- Average lead time30 days
- Cost at the average lead time$7,200
Try a case
The curve with both answers on it
- Horizontal
- Lead time, days
- Vertical
- What that lead time costs, $
- Units
- Days on the horizontal, whole dollars on the vertical.
Fixed panels
Bend at 20 days, before all three. Gap $0.
Bend at 26 days. Gap $1,800.
Bend at 30 days, on the average. Gap $3,600.
Bend at 40 days, after all three. Gap $0.
| Lead times you typed | 22 · 30 · 38 days |
|---|---|
| Weights | 25% · 50% · 25% |
| Average lead time | 30 days |
| Cost at the average lead time | $7,200 |
| Average of the three costs | $9,000 |
| The gap | $1,800, or 25.0% |
Where the weight sits
Shown in any single currency; the arithmetic is identical.
The gap against where the bend sits
- Horizontal
- Where the cost starts, days
- Vertical
- The gap, $
- Basis
- The gap is the average of the three costs less the cost at the average lead time.
- Units
- Days on the horizontal, whole dollars on the vertical.
- Assumptions
- The three lead times and weights as you set them; one fixed daily rate; only the bend moves on this axis.
What the average of the answers is not
It is not the answer either. It is the average of three answers, and a fourth lead time you did not type would move it, which is the honest position: two numbers on this page, both partial, and the useful thing is the distance between them rather than either one on its own.
What the distance measures is the bend. A curve with no bend in the range you actually face returns the same figure both ways, and a bend sitting exactly on your average returns the widest difference this page can produce. So the question the arithmetic is really asking is where your bend is, and most of the time nobody has written that number down at all.
It also assumes three lead times stand in for every lead time, that the weights are known, that the cost per day never changes once it starts, and that a day early is worth nothing. None of those hold. All of them are why the gap below is the start of a conversation about the bend rather than the end of one.
Here is the maths
- Scale
- The four fixed panels share one scale, 16 to 44 days and $0 to $45,000, and it does not move. The live chart above rescales with what you type.
- Basis
- Both answers are read against the same three lead times and the same curve.
- Assumptions
- Three lead times, no more; weights that keep the average where you set it; one bend and one rate; nothing else on the curve.
- Model version
- 1.0
The formula, in words
- Averaging is one step, and the curve is another Do them in one order and you get the cost at the average lead time. Do them in the other order and you get the average of the costs. Only a straight line lets you swap the two without changing the answer, and the moment a curve has a bend in it the order starts to matter.
- The bend is the whole of it Below the bend the curve charges nothing, above it the curve charges a fixed amount a day. That single corner is enough: it means a lead time inside the bend has slack that the arithmetic can spend but the world cannot, and the average of the answers is the version that refuses to spend it.
- The gap has a closed form and it is a triangle Multiply the daily rate by the weight on each outside lead time, then by whichever is smaller, the distance from the bend down to the low lead time or the distance from the bend up to the high one. Outside those two lead times that is zero. It peaks when the bend sits on the average, and the weight in the middle does not enter it at all.
- The average lead time never moves The two outside lead times always carry the same weight as each other, so widening the spread or moving the weight to the edges changes what the three cost without changing what they average. That is the row in the table that stays still while everything above it moves.
A worked example
Shipped defaults: average lead time 30 days, the other two 8 days either side, weights even, cost starts after 26 days, $1,800 a day after that. Monetary values are in dollars and durations in days.
The three lead times. · · , weighted 0.25, 0.50, 0.25. Their average is days, which is the number we started from, and it stays 30 for every spread and every shape.
The average of the answers.
| Lead time | Days past the bend | What it costs | Weight | Contribution |
|---|---|---|---|---|
| 22 | $0 | 0.25 | $0 | |
| 30 | $7,200 | 0.50 | $3,600 | |
| 38 | $21,600 | 0.25 | $5,400 | |
| $9,000 | ||||
The triangle, checked at its three corners. The apex is at a bend of 30, and there the plan built on the average reads while the average of the three costs is . The base is [22, 38]: at a bend of 22 the gap is , and at 38 it is . Between and beyond those points the gap is a straight line, so the whole function is an isoceles triangle sitting exactly on the three lead times, with its apex exactly on the average of them.
The shape control, at the shipped bend. never appears in the closed form, so only the tail weight moves the gap. Even gives ; Edges gives ; Middle gives . The average lead time is 30 days in all three.
Check it yourself
A curve that says a plan costs nothing is the clearest version of the argument this page can draw, and the small panels beneath the chart show that it is the widest the gap ever gets.
What this does not model
Any curve other than one bend and one rate, a cost that grows with how long the delay lasts, a cost that stops at a cap, more than three lead times, weights that are not symmetric, dependence between the lead time and anything else, the possibility that the bend itself is uncertain, the value of arriving early, and the fact that most real consequences bend more than once.
The argument on this page comes from Sam L. Savage, The Flaw of Averages, and is described here without reproducing its text. Hopp and Spearman name the same shape in waiting, Cachon and Terwiesch in inventory, and Meredith and Mantel in completion dates. Everything else is arithmetic on the numbers you type. The cost curve is a shape chosen to be checkable, not a model of a particular contract, and the three lead times are yours rather than supplied data.
Related
- Safety stock: variance beats the average lead timeTwo suppliers both quote 30 days. One of them quietly costs you triple the safety stock.Interactive lesson · 2 min
- Correlation: every estimate honest and the sum still lateEight serial tasks estimated at 8, 10 and 16 days sum to 80. Independent, four runs in five finish inside the 96 day allowance. Fully correlated they do not, and no estimate changed.Interactive lesson · 3 min