Bullwhip effect

Why does a small wobble in shop sales turn into wild swings at the factory that supplies it?

Consumers are creatures of habit. The number of nappies or beer crates a country buys next week is close to the number it bought last week — retail demand for staples is remarkably steady. Yet the factories making those staples swing between overtime and idle lines, and their suppliers swing harder still.

The obvious explanation is that something upstream is badly run. But the pattern appears in well-managed chains, across industries, and even in laboratory conditions with no incompetence available. So the assumption worth questioning is that a steady input should produce a steady output. What if the amplification is generated by the ordering process itself, by people each behaving reasonably?

Put yourself behind the counter of the shop. Sales this week are five units above normal. What do you order?

The naive answer is five extra, but that is not what a sensible retailer does, and the reason is worth stating precisely. You keep a buffer — safety stock — sized against how uncertain demand looks and how long resupply takes. A rise in observed sales moves your forecast up, and a higher forecast means a higher target buffer too. So your order must cover the extra sales and the increase in the stock you want on the shelf. You order more than the rise. Not through panic: through arithmetic.

Now stand one tier up. The distributor never sees consumer sales. It sees your order — already amplified — and treats that as its demand signal, applying the same reasoning to the inflated number and amplifying again. Each tier does something locally correct with the only information it has, and the distortion compounds at every hand-off.

Then add the delay, which turns amplification into oscillation. Orders take weeks to arrive, the shortage persists meanwhile, and the natural response is to order again — forgetting the stock already in transit. John Sterman's work on the beer distribution game showed exactly this: players systematically fail to account for the supply line, over-order while waiting, then find themselves buried and cancel everything. The game gives them a single small step-up in customer demand and nothing else, and it reliably produces enormous swings from participants who are not stupid and often cannot explain afterwards what happened.

Three further engines were catalogued by Lee, Padmanabhan and Whang. Orders are batched: a shop reordering monthly to save on shipping converts smooth sales into one lumpy signal. Prices move: a promotion makes buyers stock up, so the factory sees a spike that is a shift in timing rather than consumption, followed by a trough. And when supply is short, buyers game the allocation — if a supplier rations in proportion to orders, the rational move is to order far more than you need, and those phantom orders evaporate the moment supply loosens.

Four mechanisms, none requiring anyone to behave badly, all pushing the same way. What follows for the fix? If the cause is that each tier sees only the tier below, change what is visible: share point-of-sale data upstream so the factory forecasts from consumer demand rather than a twice-distorted order stream. Let the supplier manage the customer's inventory directly, so one party sees both shelf and plant and there is one forecast rather than four. Shorten lead times, so less can happen while an order is in flight. Cut batch sizes. Hold prices steady to remove the artificial spikes. Allocate scarce supply on past sales rather than current orders, which makes over-ordering pointless.

It is tempting to read this as a psychology finding, because the beer game is a psychology experiment. That is half right. Misperception of the supply line makes the swings worse, but amplification also appears in purely mathematical models where every tier applies an optimal order-up-to policy with no error at all. Forecast updating plus lead times is sufficient on its own; the behavioural component adds overshoot on top of a distortion already there.

A second refinement concerns the word "effect". Amplification is measured, not merely modelled — order variability exceeding sales variability has been documented in real industry data. But it is not universal: studies across many firms have found some retailers actually smoothing demand rather than amplifying it, with the sharpest amplification further up, at wholesale and manufacturing tiers. That fits the theory, since a retailer ordering frequently in small batches from a nearby supplier has short lead times and little reason to amplify, but it corrects the idea that every chain whips.

Finally, note what the remedies share. None asks anyone to forecast better. They change the information structure — who can see what, and how long they wait before their action has an effect. That is the signature of a systems problem rather than a skills problem, and it is why exhorting purchasing managers to be less reactive has never worked.

• How the beer distribution game is run, and why experienced managers do no better at it than students.
• Why sharing point-of-sale data requires trust that competing firms in a chain often do not have.

Every term the collection defines is gathered in the glossary.