In short: Suppliers discount the forecasts they receive when those numbers have moved every week and settled lower, so they apply a factor of their own and plan to that instead. Forecast, plan and commitment are three different objects, and a supplier collaboration process that sends only the first gives the supplier nothing to hold capacity against. A capacity reservation is the instrument that turns a number into something a supplier will act on, because it carries a price for both sides. The measure that tells you whether the relationship can work is your own forecast stability at the supplier's commitment horizon rather than at the point of delivery.
Every Monday a file goes out to your key suppliers. Twelve months of requirement, by item, updated from the latest planning run. It has been going out for years and the portal reports that they open it.
What happens at the other end is that a planner exports it, multiplies the first three months by a factor they worked out themselves some time ago, and plans to that. The factor is somewhere between 0.6 and 0.9 depending on which of your product families it is and how the last two years went. Nobody has told you the factor exists, and if you asked directly they would probably deny it, because it sounds like bad faith rather than what it actually is, which is a reasonable response to a series of numbers that moved every week and settled somewhere lower.
What a supplier actually needs from you
The instinct is that more forecast is better: longer horizon, finer detail, more frequent updates. What a supplier's planner needs is narrower and mostly about structure.
A stated commitment horizon. Three zones is the usual arrangement and it works. A frozen window where you own the number and it does not change, a flexible window where it can move within an agreed band, and a forecast window where it is an estimate and both parties know it. The lengths should be set by the supplier's own lead time and lot sizing, and the frozen window has to be genuinely frozen. A frozen window that gets overridden twice a quarter is a flexible window that nobody has priced.
Confidence attached to the number. A single figure carries no information about how much the supplier should hedge. A figure with a range, generated from the same model rather than added by hand afterwards, lets them decide how much raw material to commit. Conformalized quantile regression gives calibrated intervals without assuming a distribution, and the value of publishing an interval is that it makes the difference between a stable item and a volatile one visible to the person deciding what to buy.
The shape, and specifically the peak. For anything capacity constrained, the supplier's binding problem is your worst week rather than your annual total. A forecast expressed monthly hides the peak inside an average, and a supplier who plans against that average will fail in exactly the week you cared about.
Their planning unit. If their constraint is a raw material, a tool, or an autoclave hour, a forecast in your finished goods codes forces them to run your bill of materials themselves, badly, from an incomplete view. Converting once, at your end, where the data is correct, removes a whole category of error.
What actively hurts is the rolling weekly update with no structure. A number that changes every week gives the supplier no point at which they are safe to commit, so they either commit early and carry the risk unpriced, or wait until the last moment and become the constraint you complain about.
Forecast, plan and commitment are three different objects
Most supplier conversations run all three words as synonyms, and the damage from that is specific.
A forecast is an estimate of what demand will be, with an error distribution around it. A plan is what you intend to do given that estimate, including the buffer you have chosen to carry. A commitment is a quantity you will pay for whether or not you need it.
Run them together and each one degrades. The forecast gets inflated, because if the number the supplier plans against is the number you sent, and being short is expensive while being long is somebody else's problem, the rational internal behaviour is to send a bigger number. The plan gets treated as a promise, so people stop changing it when conditions change, and the planning process loses the property that made it useful. And the commitment never gets priced, because it is hiding inside a forecast, so you are extracting a real option from the supplier for free and they are recovering the cost somewhere in the unit price.
Cachon and Lariviere set this out formally in Management Science in 2001. Their result is that a buyer's forecast is credible to a supplier only when the contract makes overstating it costly to the buyer. Absent that, forecast sharing is cheap talk in the technical sense, and a supplier who discounts it is behaving correctly rather than uncooperatively.
The practical version is a labelling change in the file. Each period carries a status: committed, planned, or forecast. The committed quantities carry a stated consequence for cancellation. It is a small change and it does more for the relationship than a portal, because it lets the supplier act differently on the three, which is what they were trying to do with their private multiplier anyway.
Capacity reservation as a commercial instrument
Once the three objects are separated, the commitment becomes something you can trade rather than something you avoid mentioning.
Capacity reservation is the general form: you pay for the right to a quantity of output in a period, whether or not you take it. It appears as a reservation fee against an option to call, as a take or pay minimum, or as a tiered structure where a firm base is committed and additional tranches are optional at rising prices. The variant matters less than the principle, which is that the uncertainty in your demand has a cost, that cost is currently sitting with the supplier unacknowledged, and moving it into an explicit price makes both sides behave better.
It changes the supplier's problem from guessing to allocating. A supplier with three customers all sending optimistic rolling forecasts has to guess whose is real. A supplier with one reserved commitment and two forecasts knows exactly what to build first, and the customer who reserved gets served during the shortage, which is when they actually needed the relationship.
It also produces a number that is hard to get any other way. When you have paid a specific sum for the option to take an extra 20 percent in the peak quarter, you have a market price for your own forecast uncertainty, and the internal conversation about whether forecast improvement is worth funding becomes arithmetic instead of opinion.
Where it fits is narrower than it sounds. Constrained assets, long lead time inputs where the supplier must commit before you must, seasonal peaks that everyone hits at once, and contract manufacturing where the alternative is to be third in the queue. On commodity inputs with deep supply and short lead times there is nothing to reserve and the instrument is unnecessary overhead.
The information that should come back
Collaboration that only runs one way is a reporting requirement. The supplier holds facts that should change your plan, and most buyers never ask for them.
The two most useful are almost never in the standard supplier onboarding pack. Minimum run quantity, because it determines whether your order pattern is cheap or expensive to serve, and a supplier absorbing the cost of uneconomic runs will recover it in price or in flexibility. And changeover time and cost between your variants, because that is what decides whether ordering three items every week is worse for both of you than ordering one item every three weeks at the same total volume.
Beyond those, ask for planned maintenance shutdowns twelve months out, capacity by product family rather than in total, the lead time of their own critical inputs, and the point in their cycle at which your order becomes unchangeable. That last one is frequently different from the lead time on the purchase order, and the gap is where most expedite arguments live.
The right artefact is a constraint sheet rather than a capacity figure, refreshed quarterly, held next to the supply plan. Where a supplier's constraint changes your plan, the plan should change; that sounds obvious and it fails in practice because the constraint information arrives in an email to a buyer and never reaches the planning model. Mapping suppliers for concentration and sub-tier exposure is a related exercise with its own method, covered elsewhere on this site.
The measure that matters from the supplier's side
Forecast accuracy is measured from your point of view: how close the final number was to what happened. A supplier experiences something different. What matters to them is how much the number moved between the moment they had to commit and the moment you ordered.
A forecast that was 15 percent wrong but never moved after their commitment point is easy to supply. A forecast that landed exactly right after swinging 40 percent in the eight weeks before their cut off is expensive, because they carried the swing in raw material, overtime and idle capacity. From their side the second one is the worse relationship, and your accuracy metric shows it as the better one.
The measure to add is straightforward. For each target period and each supplier, take the quantity you forecast at their commitment lead time, take the quantity you forecast four weeks later, and record the absolute percentage change. Average those across periods. That is forecast churn at the horizon that matters to them, and it is the number to put on the supplier review agenda next to on time delivery.
Add bias at the same horizon, signed rather than absolute. Persistent over forecasting is precisely what teaches a supplier to apply a discount, and the size of your average overstatement is a good estimate of the factor they are quietly applying.
Terwiesch, Ren, Ho and Cohen published an empirical study of this in Management Science in 2005, looking at forecast sharing in the semiconductor equipment supply chain. They found buyers systematically inflating forecasts, suppliers discounting them in response, and the effects persisting as reputation on both sides, with buyers who had inflated receiving worse treatment later. The discounting behaviour has been observed and quantified in a working supply chain, which is worth knowing when a supplier's planner tells you they take your numbers at face value.
Publish your own stability numbers to the supplier before they ask. It converts a metric that looks like self criticism into the only evidence that would let them reduce the discount.
The limit
A supplier will discount your forecast by however unreliable it has historically been, and there is no implementation that shortens that.
The rebuild takes as many commitment cycles as it takes. If their lead time is sixteen weeks, a year of honoured frozen windows is three or four observations, which is barely enough for anyone to update a belief they formed over five years. Expect two years of consistent behaviour before the discount visibly moves, expect a single broken quarter to reset a good part of it, and plan the relationship on that timescale rather than on the project plan.
Some of the discount is not about you at all. A supplier allocating capacity across a customer base where most forecasts are inflated will apply a default haircut to everyone, including the customer who has never inflated anything. Your good behaviour only pays if they measure it per customer, and asking them whether they do, and offering to be measured, is a more productive conversation than asserting that your numbers are reliable.
There is also a case where none of this has any force. If you are a small share of a supplier's volume, the collaborative apparatus is overhead and the forecast will be read by nobody. The only instrument with weight in that position is a paid commitment, and the honest version of the strategy is to reserve what you genuinely need and accept market treatment for the rest.
Pull your last twelve monthly forecast files, pick the three suppliers with the longest lead times, and compute how much your number for a given month moved between the week they had to commit and the week you actually ordered.