communicoupling
You turn the shower dial. Nothing happens. You turn it further — and scald yourself. Couples run this loop with distance and reassurance, central banks with interest rates, managers with hiring. Wherever the consequence of an action arrives later than the next decision, correction curdles into overcorrection.
The beer game — a four-tier supply chain Jay Forrester built as a teaching exercise for Industrial Dynamics (1961), which John Sterman later turned into one of the most replicated experiments in management science (1989) — is the cleanest trap ever laid for this instinct. In a feedback loop with delay, decision-makers who ignore the supply line — everything ordered but not yet arrived — systematically overshoot, and their locally rational corrections amplify into a wave that grows at every step up the chain. The bullwhip effect, Sterman showed, comes from the players themselves: it appears even when customer demand changes exactly once, and it persists in subjects who know the trap is there.
Below, the full chain — retailer, wholesaler, distributor, factory — plays Sterman's ordering rule against a single step in demand: four cases a week becomes eight, at week five, forever. That one changed number is the only surprise the world supplies. Everything after it is self-inflicted.
Correction through a delay
Each tier sees two things: the orders arriving from below, and its own shelf. Sterman's rule is what his subjects actually did — anchor on a demand forecast, then adjust for the gap between the stock you want and the stock you see. The trap is the third term, the one weighted by β. At β = 0 a tier takes no account of orders already placed but not yet delivered. So it corrects the same shortfall again next week, and the week after — the delay hides every correction until they all arrive at once. One shortfall gets ordered three and four times over, and the glut that follows is exactly the panic, returned with interest.
Then the chain multiplies it. My orders are your demand: the wholesaler cannot tell the retailer's panic from real thirst, so it forecasts from the panic, adds its own stock correction, and passes a bigger wave to the distributor, who does the same. In the classic run the peaks climb 14 → 25 → 40 → 53 cases a week — ×6.6 the new demand of eight. Each decision in that chain is defensible in isolation. The disaster is a property of the loop, which is Sterman's point: subjects blamed customers, colleagues, and luck, when the only demand shock in the whole game was one step from 4 to 8. They were fighting their own reflection, phase-shifted by the delay.
What to try
Play the defaults (delay 2 · β 0). Demand steps at week 5; the retailer's orders peak at 14 cases, the factory's at 53 — then factory orders flatline at zero for fifty-nine weeks while the glut burns off. One change in the world; a year of self-inflicted chaos.
Drag β to 1. The factory's peak collapses from 53 to 19 cases and total cost falls from about $16,000 to $4,800. Nothing else changed. The cure is one act of counting: what have I already set in motion?
Keep β = 0 and set delay = 4. Amplification jumps from ×6.6 to ×26 and the oscillation period stretches from 26 to 40 weeks. Overshoot is not linear in delay — every extra week of blindness re-orders the same correction once more.
The concept in social life
Run the game on a marriage. One partner feels the distance and asks for reassurance. Warmth has a shipping delay — the other partner is turning toward them, but slowly — so the request seems to land on nothing, and gets repeated, louder. By the time the reassurance arrives it arrives all at once, feels like smothering, and triggers the counter-correction: withdrawal. Each person is responding sensibly to what they can see. Neither is counting the supply line — the repair already in motion — so each reads the delay as refusal and orders again.
Organisations play it with headcount. A team is drowning, so the manager opens requisitions; hiring pipelines run on months-long delays; the team is still drowning next quarter, so more requisitions open. Everyone arrives in the same season, and the hiring freeze follows the panic as surely as the glut follows the shortage. Central banks run it on an eighteen-month lag between a rate change and its effect on prices — which is why the brake and the accelerator so often arrive at the wrong moments. And in any steep hierarchy, the beer game explains how a small tremor at the bottom becomes a crisis at the top: each layer reads the layer below's corrections as the state of the world, adds its own margin of safety, and passes it up. The chief executive experiences a catastrophe six times the size of anything a customer did.
The deepest reading is about the channel. Each tier's order stream is a one-number-per-week message that hopelessly confounds two things: what I need and how frightened I am. The point-of-sale preset shows what happens when the chain stops inferring and starts talking — every tier sees true end demand, and amplification falls to ×1.8 with the ordering rule still fully blind. Better coupling through shared information beats coupling through orders alone. Most of what looks like panic in a system is inference doing the work that communication should have done.
Neighbouring loops
Everything here is negative feedback — the tame kind, the thermostat kind. Delay alone turns it predatory: push the correction out of phase with the error and a stabilising loop becomes an oscillator, which is how this page connects to its siblings on limit cycles and vicious circles. And the cure has a cybernetic shape. Counting your supply line means carrying a small model of everything you have set in motion — a private echo of requisite variety: a regulator that represents less of the situation than the situation contains will be beaten by the remainder. In the beer game the unrepresented remainder is your own past orders, still in the mail, coming to punish you.
The mapping
| In the model | In the world |
|---|---|
| demand step, 4 → 8 | A small, one-time change in what the world actually wants. |
| shipping & order delays | The gap between acting and feeling the consequence. |
| supply line SL | Everything already in motion that you have stopped counting. |
| β = 0 ordering | Managing by what you can see from where you sit. |
| the order stream | The only message each level ever receives from the one below. |
| amplification ratio | How panic compounds as it climbs a hierarchy. |
| shared point-of-sale data | Talking instead of inferring. |
Where it tears
The simulation is the laboratory strain of the disease. Real bullwhips are also driven by capacity limits, order batching, price promotions, and shortage gaming — ration a scarce product and customers strategically over-order, then cancel. None of that appears here, and some real amplification would survive even perfectly patient ordering. The clean model shows that misperceived feedback is sufficient for the bullwhip; it never claims to be the only cause.
"Just watch your supply line" sounds like advice. Sterman's data say it does not work as advice: subjects under-weighted the supply line (β ≈ 0.34) even with all the information in front of them, and experience barely moved it. The honest conclusion is institutional, and less flattering — shorten the delays, share the point-of-sale signal, automate the counting. Design systems that do not require people to be what people reliably are not.
Demand really does swing — seasons, fashions, epidemics — and a chain faithfully tracking a swinging world will swing with it. The diagnostic mark of a bullwhip is variance growing upstream while end demand stays nearly flat, which is what the simulation isolates by allowing demand to change exactly once. Before blaming the loop, check the input: some whips are cracked from outside.