Operations: fulfilment, packaging & compliance
Amazon demand is a curve. Wholesale demand is a staircase. A forecast that works for one will strand you on the other, and the failure is asymmetric: running out on Amazon costs you rank for a few weeks, while running out on a retail purchase order can cost you the account.
The fix is not a more sophisticated model. It is forecasting the two channels separately, holding safety stock separately, and having one explicit rule for when wholesale is allowed to eat the Amazon buffer. This article walks through the arithmetic and the conversations that produce the inputs.
Two demand shapes in one warehouse
On Amazon, you sell some number of units a day, that number moves with season and ad spend, and the past thirty days is a decent predictor of the next thirty. Demand is continuous. It is safe to talk in daily or weekly averages, because the variance around the average is small relative to the average itself.
Wholesale is nothing like that. A distributor places one order for 600 units, then nothing for eleven weeks, then 900. Three stockists reorder in the same fortnight because a season turned. A chain runs a reset and buys once for forty doors. The average weekly wholesale demand might be 140 units, and there will be no week in the year where you sell 140 units wholesale.
Averaging a staircase produces a number that describes nothing. What you need from the wholesale channel is not an average, it is the size and timing of the next few lumps, and the only source for that is the buyers themselves.
Why an Amazon-only forecast strands you
Say a brand sells a 19 dollar SKU at a steady 40 units a day on Amazon, and reorders from the factory when stock gets low. That works for years. Then wholesale starts.
The failure runs in the same order every time. A buyer commits to a first order. It is small, so it comes out of Amazon stock without much thought. It goes well, the buyer reorders larger, and that one also comes out of the same pool. Two more accounts open. Now roughly a quarter of the units leaving the warehouse are going to stores, but the reorder point was set on Amazon velocity alone. The next container is sized for Amazon and arrives on an Amazon schedule.
Then the chain places its first real order and there is nothing to ship it with. You either cut the Amazon channel to the bone, which costs rank and triggers low-inventory charges, or you short the retail order, which is the one outcome a new account remembers. This is also the moment when low-inventory fees on FBA start showing up on a settlement that was previously clean.
Neither choice is a forecasting sophistication problem. It is a bookkeeping problem: two kinds of demand were drawing from one pool with one trigger.
Lead time, reorder point and safety stock
Three numbers run the whole system, and only one of them is about demand.
Lead time is the full clock from placing a purchase order with the factory to units being available to ship, not the factory's stated production time. It includes production, inspection, booking freight, ocean or air transit, customs clearance, drayage, receiving at your warehouse, and if the units are going to FBA, the check-in queue at the fulfilment centre. Measure it from your own last three runs rather than trusting the factory number.
Lead time demand is what you will sell while that clock runs. Safety stock is the buffer on top, sized for how wrong the lead time or the demand might be. Reorder point is the sum: place the order when stock on hand hits lead time demand plus safety stock.
Written as arithmetic, reorder point equals daily demand multiplied by lead time in days, plus safety stock. The only difficult input is daily demand when half your demand is not daily, which is why the two channels get computed separately and then added.
A worked reorder point
Take a plausible SKU: a stainless steel pour-over coffee kettle, 19 dollars wholesale, 44 dollars on Amazon, 12 to a case.
Amazon side. Sales run at 40 units a day. Total lead time from the last three runs came in at 78, 84 and 92 days, so use 90 to be honest rather than optimistic. Lead time demand is 40 times 90, which is 3,600 units. Amazon safety stock is set at three weeks of cover, so 21 times 40, which is 840 units. The Amazon reorder point is 4,440 units.
Wholesale side. Four active accounts. Two distributors order about 600 units every eight weeks each, and two independent chains order about 250 units every twelve weeks each. Over the 90 day lead time that is roughly 1,350 units from the distributors and 500 from the chains, call it 1,850 units of lead time demand. Wholesale safety stock is not set in weeks of cover, because the demand is lumpy. It is set as one full order from the largest account, 600 units, so that a single unexpected reorder does not break anything.
Combined reorder point: 3,600 plus 840 plus 1,850 plus 600, which is 6,890 units. Round up to a clean case and container quantity.
Now the comparison that matters. The Amazon-only reorder point was 4,440 units. The real one is 6,890. A brand triggering its reorder at 4,440 while shipping wholesale out of the same pool is placing every order roughly 2,450 units too late, which at a combined burn of around 60 units a day is about six weeks short every single cycle. That is not a small miss, and it repeats.
Safety stock per channel, not per SKU
The reason to compute two buffers rather than one is that they exist for different risks.
Amazon safety stock covers forecast error and lead time variance. Weeks of cover is the right unit because the demand is continuous, and three to four weeks is a sane starting point for a stable SKU with a long lead time. The consequence of getting it wrong is recoverable: rank slips, ads get expensive for a while, and it comes back.
Wholesale safety stock covers a different risk entirely: one buyer doing something reasonable that you did not predict. A store adds four doors. A distributor decides to hold more depth. Someone moves a seasonal order forward by a month. The right unit is not weeks, it is orders: hold enough to absorb one full order from your largest account without touching the Amazon buffer.
That framing also tells you when to raise it. Add a chain that can order 2,000 units at a reset and your wholesale safety stock has to grow to match, even though nothing about the average changed. The size of your buffer is set by your biggest customer, not by your typical week.
Getting a number out of a retail buyer
The single highest-leverage forecasting activity is not modelling. It is asking.
Buyers plan against an open-to-buy, which is the budget they have available to spend on inventory in a category for a period. They know roughly what they intend to spend with you and when, and most will tell you if the question is specific. The unhelpful version is asking whether they think they will reorder. The useful version asks how many doors are carrying it, how many units per door they want on hand, what their own reorder cadence is, and when their next planning window falls.
From four answers you get a forecast. Forty doors, six units on the shelf per door, a reorder when a door is down to two, and a buying window that opens in February: that is a shape you can plan against, and it is far better than a weekly average.
Push further where the relationship supports it. A soft commitment on quantity for the next two quarters, even without a signed order, is worth building a container around. Some accounts will agree to a firm forecast in exchange for a guarantee that stock is held for them, which is a genuinely good trade for both sides. When a buyer will not commit to anything, that is also information: treat the account as opportunistic demand and do not let it drive production. Working with distributors covers how the cadence differs from a direct retail account.
Seasonal resets and when chains actually order
Retail buys ahead of the shelf, and the gap is larger than Amazon sellers expect.
Chains plan around resets, the scheduled points when a category's shelf layout is redone. The order for a reset is placed well before the shelf changes, and the buying conversation happens before that again. In practical terms, a decision made in early spring can be about shelves that change in autumn, and the goods have to be at the distribution centre weeks before the reset date.
For a brand with a 90 day supply lead time, this arithmetic is unforgiving. If a buyer wants delivery six weeks before a reset, and the reset is nineteen weeks away, the factory order has to be placed now or the units will not exist. The habit worth building is to put every account's known reset and planning windows on one calendar, work backwards by your true lead time, and let that calendar drive when you talk to the factory.
The corollary: a missed reset is not a delayed order. The shelf gets set without you and the next chance is the following cycle.
The reserve pool and the rule for releasing it
Once the two forecasts exist, physically or logically separate the inventory. A wholesale reserve pool is simply a quantity of units that Amazon replenishment is not allowed to draw from, held in your own warehouse rather than sent into fulfilment centres.
The pool needs one written rule, decided in advance and not in the moment. A version that works: the reserve is released to Amazon only when it exceeds the next 60 days of committed and forecast wholesale demand, and only down to that level, and never within four weeks of a known reset window. Anything else and the pool quietly drains, because the Amazon side always has an urgent-sounding reason today while the retail order is still hypothetical.
Holding a reserve has a real cost and it should be counted. Units sitting in your own warehouse are cash that is not working, though they are at least not accumulating FBA storage fees, which is one reason to hold the buffer outside the fulfilment network rather than inside it. The other reason is that units already in Amazon's building cannot easily be turned into a pallet for a retail dock.
The cost of being wrong in each direction
Both errors cost money and they are not the same size, which is the whole argument for the reserve.
Over-forecast Amazon and you hold too much in fulfilment centres: monthly storage charges, aged-inventory surcharges on units that sit too long, and capital tied up in stock at the wrong end of the country. It is a measurable, survivable cost and it shows up cleanly on a settlement report, so it gets attention.
Over-forecast wholesale and you hold units in your own warehouse waiting for an order that shifts a quarter. That is cheaper per unit per month than the FBA version but it ties up the same cash, and the cash pressure is worse because wholesale pays on terms rather than on a two-week cycle. Sizing that squeeze is what Amazon seller cash flow is about.
Under-forecast Amazon and you lose rank, pay more for ads to recover it, and possibly take a low-inventory charge on the way down. Painful, temporary.
Under-forecast wholesale and you short a purchase order. The buyer fills the shelf space with a competitor, the reorder does not come, and the account that took nine months to open closes in an afternoon. There is no advertising spend that recovers it. That asymmetry is the reason the wholesale buffer gets protected first and released last.
If the accounts that would generate this demand do not exist yet, the forecast is not the problem. Paste a product link into WholesalePilot and the preview shows which distributors and retailers would plausibly stock it, which is the input every number above depends on.