Your retailers sold almost exactly as much in March as in October. Your distributors ordered forty percent more in one of those months and nothing at all in the other.
Nobody changed their behaviour. The swing was manufactured inside your own distribution chain, and you are about to build capacity for it.
Why these three models
The decision is whether to respond to an order swing at all, and what to change if you do. The features that fire are a delay between ordering and receiving, several layers each reacting to the layer next to it rather than to the end customer, and a short run of numbers that looks like a trend.
Three lenses. Delay and oscillation produce a complex answer about where the amplification comes from. Bargaining produces an equilibrium answer about why your partners inflate orders on purpose and are right to. Luck and skill produce a random answer about whether the swing you are staring at contains any information. The second is the one founders miss entirely, because it looks like a supply problem and it is a contract problem.
1. The chain manufactures the swing
Take a chain with four layers: retailer, distributor, wholesaler, you. Each one orders from the layer above. Each order takes time to arrive. Each layer holds stock and wants to protect against running out.
Now raise end demand by a small amount and hold it there. The retailer sells more, dips into stock, and orders extra to both meet the new demand and rebuild the buffer. The distributor sees an order larger than the true increase, and does the same thing on top of it. By the fourth layer the peak order is far above the peak at retail. Then the pipeline fills, everyone finds themselves overstocked, and orders collapse to zero for a period while the excess is worked off.
The amplification does not require anyone to panic or to forecast badly. It is produced by the structure: a delay between ordering and receiving, plus each layer holding a buffer, plus each layer seeing only its immediate neighbour.1 This behaviour reproduces reliably in a controlled four-layer setting where customer demand steps up once and then stays flat for the rest of the run. Teams of graduate students and business executives produce large oscillations anyway, average costs run about ten times the optimum available from the information they actually hold, and the peak order rate at the factory averages more than double the peak at retail.2 The demand curve almost every player draws afterwards is a boom and a crash, and the real one was a single step.
Two things follow. The first is that the volatility your operations team complains about may not exist at the customer end at all, and you can check that in an afternoon by comparing the variance of your own order book against the variance of retail sell-through. If yours is several times theirs, the chain is generating it.
That comparison needs one caution or it will mislead you. Compare like periods in volume rather than in value, because a price change during the year moves the value series and leaves the physical one alone. And divide each series by its own mean before comparing, so that an order book measured in truckloads and sell-through measured in units are on the same footing. The number you want is how far each series swings relative to its own average, not how big either one is.
Where you hold no sell-through data at all, which is the usual case, there is a partial version worth a week of somebody’s time. Pick five retailers who will let you count. Weekly stock on hand at their premises, recorded by the sales representative on the visit they already make, gives you a usable series inside two months. Five is not a sample and it is enough to tell a step from a boom, which is the only question this section asks.
The second is that the standard responses make it worse. Building capacity for the peak means you carry the cost through the trough. Adding safety stock at your own layer adds another buffer to the chain, which increases the amplification for everyone above you. The intuitive fixes are additions to the structure that causes the problem.
2. Your partners are inflating orders, and they are right to
The delay explanation is incomplete on its own, because it treats every order as an honest statement of need. Many of them are not, and the reason is contractual.
When supply runs short, most companies allocate what they have in proportion to what was ordered. Think about what that rule does. If you know that ordering more gets you a larger share of a shortage, then ordering more is simply correct, and everyone who understands the rule does it. The result is an order book that inflates precisely when supply is tightest, which is exactly when you most need it to be accurate.
This is a bargaining outcome rather than dishonesty.4 Each distributor is responding rationally to the rule you wrote. And it is self-reinforcing: once one partner inflates, the shortage worsens for everyone else, and their correct response is to inflate too.
An allocation rule based on orders converts your order book into a bidding war and then asks you to forecast from it.
The fix is a rule change and it costs nothing to implement. Allocate on the basis of past sell-through rather than current orders. The moment ordering more does not get you more, the incentive to inflate disappears and orders start meaning something again. Some partners will object, and the ones who object loudest are usually the ones who have been gaming the previous rule most effectively, which is useful information in itself.
Change the rule at the start of a season rather than in the middle of a shortage, because the timing decides how it is heard. A distributor told during a stock-out that allocation now depends on last year’s sales hears a punishment. The same distributor told in a planning meeting three months ahead, with their own numbers in front of them, hears a policy. Same rule, and only one version survives the first argument.
And where a partner refuses, look at what they are protecting before you concede anything. A distributor in Kampala who inflates orders because your lead time is six weeks and unreliable is solving your problem rather than gaming you. There the allocation rule is the wrong lever and the lead time is the right one, and you recover the accuracy by making the delivery date credible, not by rewriting the contract.
A second contract lever is order batching. A distributor who orders once a month because the shipping economics require it is sending you a spike every month regardless of how smooth their sales are. Smaller, more frequent orders cost more per shipment and remove a source of amplification, and whether that trade is worth it is an arithmetic question specific to your logistics rather than a matter of principle.
3. Most of the swing deserves no response
The third lens is a brake on the first two. Having understood that the chain amplifies and that partners inflate, the temptation is to react to every movement with a structural change.
Any short run of numbers mixes signal and noise, and where the chance component is large, sequences that look like clear trends occur routinely with nothing underneath them. Reacting to one produces a change that must itself be reversed later, and the reversal arrives with its own delay.3
This is the specific failure that turns a manageable oscillation into a damaging one. The structure creates a swing. Management reacts to the swing. The reaction arrives after the swing has already turned. The reaction then amplifies the next movement in the opposite direction. The people involved experience this as a volatile market and it is a feedback loop with their own hands in it.
The discipline is to set the response rule before you see the numbers. Decide in advance how many consecutive periods of movement in one direction, at what size, justify a change in production or in ordering. Then hold to it. A rule chosen in advance is worse than perfect judgement and far better than judgement applied in the moment, because judgement in the moment is what the oscillation is made of.
Write that rule with numbers in it, because a rule without numbers is a preference. Three consecutive periods moving the same way, each by more than a stated percentage of your own trailing average, before production changes. Anything smaller is absorbed by the buffer you already carry. Whatever thresholds you pick will be roughly wrong, and having them on paper is what stops the argument being reopened every month by whoever feels strongest that week.
The rule also needs an exception clause, or the first real shock will destroy it. Name in advance what overrides the waiting period: a port closure, a currency move beyond a stated size, a large customer announcing a change. Those are observable events rather than readings of your own series, which is why they are safe to act on immediately, and naming them is what stops an ordinary swing being argued into the same category.
What the three say together
- Measure the amplification. Variance of your order book against variance of retail sell-through. A ratio well above one means the chain is generating the swing.
- Change the allocation rule from current orders to past sales. This is the highest-return change in the article and it is a paragraph in a contract.
- Get end-demand data. Any visibility of actual sell-through, even weekly and partial, shortens the chain of guesses.
- Write the response rule before the next swing. How many periods, what size, before you change anything.
Where they disagree
The structural view and the noise view conflict on how fast to act.
The structural view says the amplification is real and predictable, so you should intervene in the structure now: change the rules, shorten the delays, share the data. The noise view says most movements contain nothing and every intervention has a lag, so acting quickly is how oscillation is produced in the first place.
The resolution is that they are talking about different objects. Structural changes to rules and information flow should be made immediately and permanently, because they alter the system rather than responding to a reading. Volume responses, meaning production and ordering decisions, should be slow and rule-bound. Founders routinely do the reverse: they leave the allocation rule untouched for years while adjusting production every month.
What none of them contain
None of the three prices the relationship cost of changing an allocation rule. Distributors who have optimised against the old rule will experience the new one as a demotion, and in markets where distribution relationships are personal and hard to replace, that cost can exceed the inventory saving. The models treat partners as responsive to rules and they are also people with alternatives.
None of them handles genuine supply shocks. All three assume the underlying demand and supply are stable while the chain misreads them. When a port closes or a currency moves, the swing is real, and the discipline of not reacting is exactly wrong. Telling the two situations apart requires the end-demand data, which is the recommendation this article keeps returning to.
And none of them addresses the case where you are not the top of the chain. If you are the distributor rather than the manufacturer, you cannot change the allocation rule, and your available moves are limited to your own buffer and your own ordering discipline. That is a weaker position and the honest advice is to secure sell-through visibility from your retailers, which is the one piece of information nobody upstream has.
The one action that survives the ignorance: pull twelve months of your own order book and twelve months of retail sell-through, and compare how much each one moved. If your orders swung several times as much as end sales, you have measured the amplification and you now know the volatility is internal, which changes what you should be building next quarter.
Who has to move
This belongs to whoever writes the distribution agreements, which in most companies is commercial rather than operations, while the pain is felt entirely in operations. That split is why the allocation rule survives for years. The cheapest first test is the variance comparison, and it requires no new systems and no partner cooperation, only two columns of numbers you already hold. If the ratio comes back near one, the volatility is real and this article does not apply to you. In most distribution businesses it does not come back near one.
Sources and notes
- John D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World, McGraw-Hill. Supply chain amplification, its structural causes in delays and inventory buffers, and the analysis of the beer distribution game are developed in the chapters on supply chains and on delays. Used in section 1 for the mechanism and for the point that amplification is a property of the structure rather than of the participants.
- John D. Sterman, The Beer Distribution Game, MIT. https://web.mit.edu/jsterman/www/SDG/beergame.html. The page states that customer demand begins at four cases per week, rises to eight in week five and remains completely constant thereafter; that average team costs are about $2,000 against an optimum of about $200 computed from the information players actually hold; that orders and inventories are dominated by large oscillations; that the amplitude and variance of orders increases from customer to retailer to factory, with the peak factory order rate averaging more than double the peak at retail; and that the vast majority of players afterwards draw a rising and plunging customer demand pattern that did not occur. The full analysis is John D. Sterman, Modeling Managerial Behavior: Misperceptions of Feedback in a Dynamic Decision Making Experiment, Management Science 35(3), 1989, pages 321 to 339, which is behind a publisher paywall and is therefore cited here through the openly available page above rather than linked directly.
- Michael J. Mauboussin, The Success Equation: Untangling Skill and Luck in Business, Sports, and Investing, Harvard Business Review Press. The luck-skill continuum and the consequence that short sequences in high-variance settings frequently resemble trends. Used in section 3 for the direction of the error. No specific number of periods is claimed, since the required run length depends on the variance of your own series.
- Jeffrey Carpenter and Andrea Robbett, Game Theory and Behavior, MIT Press. Bargaining and the strategic response to allocation rules, including the general result that a rule which rewards a stated quantity will be met with inflation of that quantity, sit in the chapters on bargaining and mechanism design. Used in section 2 for the claim that order inflation under proportional allocation is a rational response rather than a behavioural failure.
A note on a number this article does not give. Published amplification ratios exist for specific industries and none of them transfers to yours. The ratio depends on the number of layers, the length of your delays and the size of the buffers, all of which you can measure and none of which I can assume. Compute your own from the two columns.
Joshua Agonya Pi’Rwot, Founder.