Thought Toys · Emergence · Exhibit 81
Customers walk into a shop and buy an almost-steady number of things. Four links up the chain, the factory is being asked for wildly swinging quantities. No one hoarded, no one lied, no one guessed badly — every link followed the same sensible rule.
Orders placed, day by day — same vertical scale throughout orders placed what customers actually bought
—
—
Each strip is one link, and each link can see only the strip directly above it — the shop sees customers, the wholesaler sees the shop's orders, and so on. Nobody has a view of the actual shoppers except the shop. Every link runs the same rule: keep a forecast of demand, hold enough stock to cover the delivery wait, and order enough to replace what sold plus whatever your stock target moved by. That last clause is the whole story. When sales tick up, a link doesn't just order the extra units it sold — it also orders more to raise its target, because a higher forecast means it now wants more days of cover on the shelf. So the order it places is bigger than the change it saw. And the link above has no way to tell the two apart. It sees a big order and does the same thing again.
Press Send one 10% rise. Customer demand goes perfectly flat and then steps up once, by a tenth, and stays there — no noise, no drama, one honest change. Watch it travel. By the factory it arrives as a spike far bigger than the step that caused it, followed by an overshoot in the other direction as everyone realises they over-ordered. Nothing in that sequence was a mistake, and nobody could have seen it from where they stood.
Two sliders can switch the effect off entirely, and they're the interesting ones. Drag the lead time to zero — if goods arrived instantly, nobody would need a stock target, so nobody revises one, and the chain passes demand through untouched at every link. Or drag forecast follows sales to zero — now each link ignores what it just saw and never revises its target, and again the amplification is exactly one. Note what you can't switch it off with: the customer-variation slider. Crank the noise all the way up with lead time at zero and the strips stay identical. The randomness was never the cause. The cause is a reasonable rule, applied by people who can only see one step downstream.
improve/verify/81-bullwhip.js, 19 checks): that closed form is
derived from the policy and then cross-examined against a 400,000-day Monte Carlo simulation of the policy
itself at five different settings, agreeing to within 0.3%; at α=1, L=2 it
lands on exactly 13, which you can check by hand. A four-link chain at the default settings measures
4.4× → 20.9× → 101.4× → 497.7× the customers' own variance. Five negative controls isolate the cause: zeroing
the lead time, or the forecast responsiveness, drops amplification to 1.000000000 with the demand noise
completely unchanged; a single 10% step in provably noiseless demand still spikes to 277 units four links
up; and mean orders equal mean sales to three decimal places at every link, so nobody is over-ordering on
average — only the swing grew.
Four identical links, one product, no capacity limits, no price promotions and no one
gaming an allocation — all of which make it worse in the real world, not better.
Also in Emergence: Phantom Traffic Jams →
← the cabinet · Thought Toys — a cabinet of explorable explanations. Exhibit 81.